2022 AI Index Report Master

Artificial Intelligence
Index Report 2022

Index Report 2022
INTRODUCTION TO THE
AI INDEX REPORT 2022
Welcome to the fifth edition of the AI Index Report! The latest edition includes data from a broad set of academic,
private, and nonprofit organizations as well as more self-collected data and original analysis than any previous
editions, including an expanded technical performance chapter, a new survey of robotics researchers around the
world, data on global AI legislation records in 25 countries, and a new chapter with an in-depth analysis of technical
AI ethics metrics.
The AI Index Report tracks, collates, distills, and visualizes data related to artificial intelligence. Its mission is to
provide unbiased, rigorously vetted, and globally sourced data for policymakers, researchers, executives, journalists,
and the general public to develop a more thorough and nuanced understanding of the complex field of AI. The report
aims to be the world’s most credible and authoritative source for data and insights about AI.
F R O M T H E C O - D I R E CTO R S
This year’s report shows that AI systems are starting to be deployed widely into the economy, but at the same time
they are being deployed, the ethical issues associated with AI are becoming magnified. Some of this is natural—after
all, we tend to care more about the ethical aspects of a given technology when it is being rolled out into the world.
But some of it is bound up in the peculiar traits of contemporary AI—larger and more complex and capable AI systems
can generally do better on a broad range of tasks while also displaying a greater potential for ethical concerns.
This is bound up with the broad globalization and industrialization of AI—a larger range of countries are developing,
deploying, and regulating AI systems than ever before, and the combined outcome of these activities is the creation
of a broader set of AI systems available for people to use, and reductions in their prices. Some parts of AI are not very
globalized, though, and our ethics analysis reveals that many AI ethics publications tend to concentrate on English-
language systems and datasets, despite AI being deployed globally.
If anything, we expect the above trends to continue: 103% more money was invested in the private investment of AI
and AI-related startups in 2021 than in 2020 ($96.5 billion versus $46 billion).
Jack Clark and Ray Perrault
2
Index Report 2022
TOP TAKEAWAYS
Private investment in AI soared while investment concentration intensified:
•T
he private investment in AI in 2021 totaled around $93.5 billion—more than double the total private
investment in 2020, while the number of newly funded AI companies continues to drop, from 1051 companies in
2019 and 762 companies in 2020 to 746 companies in 2021. In 2020, there were 4 funding rounds worth $500
million or more; in 2021, there were 15.
U.S. and China dominated cross-country collaborations on AI:

• Despite rising geopolitical tensions, the United States and China had the greatest number of cross-country
collaborations in AI publications from 2010 to 2021, increasing five times since 2010. The collaboration between
the two countries produced 2.7 times more publications than between the United Kingdom and China—the second
highest on the list.
Language models are more capable than ever, but also more biased:
• L arge language models are setting new records on technical benchmarks, but new data shows that larger models are
also more capable of reflecting biases from their training data. A 280 billion parameter model developed in 2021
shows a 29% increase in elicited toxicity over a 117 million parameter model considered the state of the art as
of 2018. The systems are growing significantly more capable over time, though as they increase in capabilities, so
does the potential severity of their biases.
The rise of AI ethics everywhere:

• Research on fairness and transparency in AI has exploded since 2014, with a fivefold increase in related publications
at ethics-related conferences. Algorithmic fairness and bias has shifted from being primarily an academic pursuit to
becoming firmly embedded as a mainstream research topic with wide-ranging implications. Researchers with industry
affiliations contributed 71% more publications year over year at ethics-focused conferences in recent years.
AI becomes more affordable and higher performing:

•S
ince 2018, the cost to train an image classification system has decreased by 63.6%, while training times
have improved by 94.4%. The trend of lower training cost but faster training time appears across other MLPerf task
categories such as recommendation, object detection and language processing, and favors the more widespread
commercial adoption of AI technologies.
Data, data, data:

•T
op results across technical benchmarks have increasingly relied on the use of extra training data to set new state-of-
the-art results. As of 2021, 9 state-of-the-art AI systems out of the 10 benchmarks in this report are trained with
extra data. This trend implicitly favors private sector actors with access to vast datasets.
More global legislation on AI than ever:

• An AI Index analysis of legislative records on AI in 25 countries shows that the number of bills containing “artificial
intelligence” that were passed into law grew from just 1 in 2016 to 18 in 2021. Spain, the United Kingdom, and the
United States passed the highest number of AI-related bills in 2021 with each adopting three.
Robotic arms are becoming cheaper:

• An AI Index survey shows that the median price of robotic arms has decreased by 46.2% in the past five years—
from $42,000 per arm in 2017 to $22,600 in 2021. Robotics research has become more accessible and affordable.
3
Index Report 2022
Steering Committee
Co-Directors
Jack Clark Raymond Perrault
Anthropic, OECD SRI International
Members
Erik Brynjolfsson James Manyika Michael Sellitto
Stanford University Google, University Stanford University
of Oxford
John Etchemendy Yoav Shoham
Stanford University Juan Carlos Niebles (Founding Director)
Stanford University, Stanford University,
Terah Lyons Salesforce AI21 Labs
Staff and Researchers

Research Manager and Editor in Chief Research Associate
Daniel Zhang Nestor Maslej
Stanford University Stanford University
Affiliated Researchers
Andre Barbe Helen Ngo
The World Bank Cohere
Latisha Harry Ellie Sakhaee

Independent Consultant Microsoft
Graduate Researcher
Benjamin Bronkema-Bekker
Stanford University
4
Index Report 2022
How to Cite This Report

Daniel Zhang, Nestor Maslej, Erik Brynjolfsson, John Etchemendy, Terah Lyons, James
Manyika, Helen Ngo, Juan Carlos Niebles, Michael Sellitto, Ellie Sakhaee, Yoav Shoham,
Jack Clark, and Raymond Perrault, “The AI Index 2022 Annual Report,” AI Index Steering
Committee, Stanford Institute for Human-Centered AI, Stanford University, March 2022.
The AI Index 2022 Annual Report by Stanford University is licensed under

Attribution-NoDerivatives 4.0 International. To view a copy of this license,
visit http://creativecommons.org/licenses/by-nd/4.0/.
Public Data and Tools

The AI Index 2022 Report is supplemented by raw data and an interactive tool. We invite each
reader to use the data and the tool in a way most relevant to their work and interests.
•R
aw data and charts: The public data and high-resolution images of all the charts in the
report are available on Google Drive.
•G
lobal AI Vibrancy Tool: We redesigned the Global AI Vibrancy Tool this year with a
better visualization to compare up to 29 countries across 23 indicators.
AI Index and Stanford HAI

The AI Index is an independent initiative at the
Stanford Institute for Human-Centered Artificial Intelligence (HAI).
The AI Index was conceived within the One Hundred Year Study on AI (AI100).
We welcome feedback and new ideas for next year.

Contact us at AI-Index-Report@stanford.edu.
5
Index Report 2022
Supporting Partners
Analytics and Research Partners
6
Index Report 2022
Contributors
We want to acknowledge the following individuals by chapter and section for their contributions
of data, analysis, advice, and expert commentary included in the AI Index 2022 Report:
Research and Development

Sara Abdulla, Catherine Aiken, Jack Clark, James Dunham, Nezihe Merve Gürel, Nestor Maslej, Ray Perrault,
Sarah Tan, Daniel Zhang
Technical Performance
Jack Clark, David Kanter, Nestor Maslej, Deepak Narayanan, Juan Carlos Niebles, Konstantin Savenkov,
Yoav Shoham, Daniel Zhang
Technical AI Ethics
Jack Clark, Nestor Maslej, Helen Ngo, Ray Perrault, Ellie Sakhaee, Daniel Zhang
The Economy and Education

Betsy Bizot, Erik Brynjolfsson, Jack Clark, John Etchemendy, Murat Erer, Akash Kaura, Julie Kim, Nestor Maslej,
James Manyika, Brenden McKinney, Julia Nitschke, Ray Perrault, Brittany Presten, Tejas Sirohi, Bledi Taska,
Rucha Vankudre, Daniel Zhang
AI Policy and Governance

Amanda Allen, Benjamin Bronkema-Bekker, Jack Clark, Latisha Harry, Taehwa Hong, Cameron Leuthy, Terah
Lyons, Nestor Maslej, Ray Perrault, Michael Sellitto, Teruki Takiguchi, Daniel Zhang
Conference Attendance
Terri Auricchio (ICML), Christian Bessiere (IJCAI), Meghyn Bienvenu (KR), Andrea Brown (ICLR),
Alexandra Chouldechova (FAccT), Nicole Finn (ICCV, CVPR), Enrico Gerding (AAMAS), Carol Hamilton (AAAI),
Seth Lazar (FAccT), Max Qing Hu Meng (ICRA), Jonas Martin Peters (UAI), Libor Preucil (IROS),
Marc’Aurelio Ranzato (NeurIPS), Priscilla Rasmussen (EMNLP, ACL), Hankz Hankui Zhuo (ICAPS)
Global AI Vibrancy Tool

Andre Barbe, Latisha Harry, Daniel Zhang
Robotics Survey
Pieter Abbeel, David Abbink, Farshid Alambeigi, Farshad Arvin, Nikolay Atanasov, Ruzena Bajcsy, Philip Beesley,
Tapomayukh Bhattacharjee, Jeannette Bohg, David J. Cappelleri, Qifeng Chen, I-Ming Chen, Jack Cheng,
Cynthia Chestek, Kyujin Cho, Dimitris Chrysostomou, Steve Collins, David Correa, Brandon DeHart,
Katie Driggs-Campbell, Nima Fazeli, Animesh Garg, Maged Ghoneima, Tobias Haschke, Kris Hauser, David Held,
Yue Hu, Josie Hughes, Soo Jeon, Dimitrios Kanoulas, Jonathan Kelly, Oliver Kroemer, Changliu Liu, Ole Madsen,
Anirudha Majumdar, Genaro J. Martinez, Saburo Matunaga, Satoshi Miura, Norrima Mokhtar, Elena De Momi,
Chrystopher Nehaniv, Christopher Nielsen, Ryuma Niiyama, Allison Okamura, Necmiye Ozay, Jamie Paik,
Frank Park, Karthik Ramani, Carolyn Ren, Jan Rosell, Jee-Hwan Ryu, Tim Salcudean, Oliver Schneider,
Angela Schoellig, Reid Simmons, Alvaro Soto, Peter Stone, Michael Tolley, Tsu-Chin Tsao, Michiel van de Panne,
Andy Weightman, Alexander Wong, Helge Wurdemann, Rong Xiong, Chao Xu, Geng Yang, Junzhi Yu,
Wenzhen Yuan, Fu Zhang, Yuke Zhu
7
Index Report 2022
We thank the following organizations and individuals who provided data for
inclusion in the AI Index 2022 Report:
Organizations
Bloomberg Government Intento
Amanda Allen, Cameron Leuthy Grigory Sapunov, Konstantin Savenkov
Center for Security and Emerging LinkedIn

Technology, Georgetown University Murat Erer, Akash Kaura
Sara Abdulla, Catherine Aiken,
James Dunham McKinsey Global Institute
Brenden McKinney, Brittany Presten
Computing Research Association
Betsy Bizot MLCommons
David Kanter
Emsi Burning Glass
Julia Nitschke, Bledi Taska, NetBase Quid
Rucha Vankudre Julie Kim, Tejas Sirohi
Women in Machine Learning

Nezihe Merve Gürel, Sarah Tan
We also would like to thank Jeanina Casusi, Nancy King, Shana Lynch, Jonathan Mindes,
Stacy Peña, Michi Turner, and Justin Sherman for their help in preparing this report, and
Joe Hinman, Travis Taylor, and the team at Digital Avenues for their efforts in designing and
developing the AI Index and HAI websites.
8
Index Report 2022
Table of Contents
REPORT HIGHLIGHTS
10
CHAPTER 1 Research and Development 13
CHAPTER 2 Technical Performance 47
CHAPTER 3 Technical AI Ethics 100
CHAPTER 4 The Economy and Education 139
CHAPTER 5 AI Policy and Governance 172
APPENDIX 196
ACCESS THE PUBLIC DATA
9
Index Report 2022
REPORT HIGHLIGHTS
C H A P T E R 1 : R E S E A R C H A N D D E V E LO P M E N T
• Despite rising geopolitical tensions, the United States and China had the greatest number of cross-country
collaborations in AI publications from 2010 to 2021, increasing five times since 2010. The collaboration between
the two countries produced 2.7 times more publications than between the United Kingdom and China—the
second highest on the list.
• In 2021, China continued to lead the world in the number of AI journal, conference, and repository
publications—63.2% higher than the United States with all three publication types combined. In the meantime,
the United States held a dominant lead among major AI powers in the number of AI conference and repository
citations.
• From 2010 to 2021, the collaboration between educational and nonprofit organizations produced the
highest number of AI publications, followed by the collaboration between private companies and educational
institutions and between educational and government institutions.
• The number of AI patents filed in 2021 is more than 30 times higher than in 2015, showing a compound
annual growth rate of 76.9%.
C H A P T E R 2: T E C H N I CA L P E R F O R M A N C E
• Data, data, data: Top results across technical benchmarks have increasingly relied on the use of extra training
data to set new state-of-the-art results. As of 2021, 9 state-of-the-art AI systems out of the 10 benchmarks
in this report are trained with extra data. This trend implicitly favors private sector actors with access to vast
datasets.
• Rising interest in particular computer vision subtasks: In 2021, the research community saw a greater level
of interest in more specific computer vision subtasks, such as medical image segmentation and masked-face
identification. For example, only 3 research papers tested systems against the Kvasir-SEG medical imaging
benchmark prior to 2020. In 2021, 25 research papers did. Such an increase suggests that AI research is
moving toward research that can have more direct, real-world applications.
• AI has not mastered complex language tasks, yet: AI already exceeds human performance levels on basic
reading comprehension benchmarks like SuperGLUE and SQuAD by 1%–5%. Although AI systems are still
unable to achieve human performance on more complex linguistic tasks such as abductive natural
language inference (aNLI), the difference is narrowing. Humans performed 9 percentage points better on
aNLI in 2019. As of 2021, that gap has shrunk to 1.
10
Index Report 2022
• Turn toward more general reinforcement learning: For the last decade, AI systems have been able to master
narrow reinforcement learning tasks in which they are asked to maximize performance in a specific skill, such
as chess. The top chess software engine now exceeds Magnus Carlsen’s top ELO score by 24%. However, in
the last two years AI systems have also improved by 129% on more general reinforcement learning tasks
(Procgen) in which they must operate in novel environments. This trend speaks to the future development of
AI systems that can learn to think more broadly.
• AI becomes more affordable and higher performing: Since 2018, the cost to train an image classification
system has decreased by 63.6%, while training times have improved by 94.4%. The trend of lower training
cost but faster training time appears across other MLPerf task categories such as recommendation, object
detection and language processing, and favors the more widespread commercial adoption of AI technologies.
• Robotic arms are becoming cheaper: An AI Index survey shows that the median price of robotic arms has
decreased by 46.2% in the past five years—from $42,000 per arm in 2017 to $22,600 in 2021. Robotics
research has become more accessible and affordable.
C H A P T E R 3: T E C H N I CA L A I E T H I C S
• Language models are more capable than ever, but also more biased: Large language models are setting new
records on technical benchmarks, but new data shows that larger models are also more capable of reflecting
biases from their training data. A 280 billion parameter model developed in 2021 shows a 29% increase in
elicited toxicity over a 117 million parameter model considered the state of the art as of 2018. The systems
are growing significantly more capable over time, though as they increase in capabilities, so does the potential
severity of their biases.
• The rise of AI ethics everywhere: Research on fairness and transparency in AI has exploded since 2014, with a
fivefold increase in related publications at ethics-related conferences. Algorithmic fairness and bias has shifted
from being primarily an academic pursuit to becoming firmly embedded as a mainstream research topic with
wide-ranging implications. Researchers with industry affiliations contributed 71% more publications year
over year at ethics-focused conferences in recent years.
• Multimodal models learn multimodal biases: Rapid progress has been made on training multimodal language-
vision models which exhibit new levels of capability on joint language-vision tasks. These models have set new
records on tasks like image classification and the creation of images from text descriptions, but they also reflect
societal stereotypes and biases in their outputs—experiments on CLIP showed that images of Black people
were misclassified as nonhuman at over twice the rate of any other race. While there has been significant
work to develop metrics for measuring bias within both computer vision and natural language processing, this
highlights the need for metrics that provide insight into biases in models with multiple modalities.
11
Index Report 2022
C H A P T E R 4 : T H E E C O N O M Y A N D E D U CAT I O N
• New Zealand, Hong Kong, Ireland, Luxembourg, and Sweden are the countries or regions with the highest growth
in AI hiring from 2016 to 2021.
• In 2021, California, Texas, New York, and Virginia were states with the highest number of AI job postings in the
United States, with California having over 2.35 times the number of postings as Texas, the second greatest.
Washington, D.C., had the greatest rate of AI job postings compared to its overall number of job postings.
• The private investment in AI in 2021 totaled around $93.5 billion—more than double the total private
investment in 2020, while the number of newly funded AI companies continues to drop, from 1051 companies in
2019 and 762 companies in 2020 to 746 companies in 2021. In 2020, there were 4 funding rounds worth $500
million or more; in 2021, there were 15.
• “Data management, processing, and cloud” received the greatest amount of private AI investment in 2021—
2.6 times the investment in 2020, followed by “medical and healthcare” and “fintech.”
• In 2021, the United States led the world in both total private investment in AI and the number of newly funded AI
companies, three and two times higher, respectively, than China, the next country on the ranking.
• Efforts to address ethical concerns associated with using AI in industry remain limited, according to a McKinsey
survey. While 29% and 41% of respondents recognize “equity and fairness” and “explainability” as risks
while adopting AI, only 19% and 27% are taking steps to mitigate those risks.
• In 2020, 1 in every 5 CS students who graduated with PhD degrees specialized in artificial intelligence/
machine learning, the most popular specialty in the past decade. From 2010 to 2020, the majority of AI PhDs in
the United States headed to industry while a small fraction took government jobs.
C H A P T E R 5: A I P O L I CY A N D G OV E R N A N C E
• An AI Index analysis of legislative records on AI in 25 countries shows that the number of bills containing
“artificial intelligence” that were passed into law grew from just 1 in 2016 to 18 in 2021. Spain, the United
Kingdom, and the United States passed the highest number of AI-related bills in 2021, with each adopting three.
• The federal legislative record in the United States shows a sharp increase in the total number of proposed bills
that relate to AI from 2015 to 2021, while the number of bills passed remains low, with only 2% ultimately
becoming law.
• State legislators in the United States passed 1 out of every 50 proposed bills that contain AI provisions in 2021,
while the number of such bills proposed grew from 2 in 2012 to 131 in 2021.
• In the United States, the current congressional session (the 117th) is on track to record the greatest number of
AI-related mentions since 2001, with 295 mentions by the end of 2021, half way through the session,
compared to 506 in the previous (116th) session.
12
Index Report 2022
CHAPTER 1:
Research &
Development
Index Report 2022 CHAPTER 1: RESEARCH & DEVELOPMENT
CHAPTER 1:
Chapter Preview
Overview 15 AI Repositories 32
Chapter Highlights 16 Overview 32
By Region 33
1.1 PUBLICATIONS 17 By Geographic Area 34
Overview 17 Citations 35
Total Number of AI Publications 17 AI Patents 36
By Type of Publication 18 Overview 36
By Field of Study 19 By Region and Application Status 37
By Sector 20 By Geographic Area and
Cross-Country Collaboration 22 Application Status 39
Cross-Sector Collaboration 23
AI Journal Publications 24 1.2 CONFERENCES 41
Overview 24 Conference Attendance 41
By Region 25 Women in Machine Learning (WiML)

NeurIPS Workshop 43
By Geographic Area 26
Workshop Participants 43
Citations 27
Demographics Breakdown 44
AI Conference Publications 28
Overview 28
1.3 AI OPEN-SOURCE SOFTWARE
By Region 29 LIBRARIES 45
By Geographic Area 30 GitHub Stars 45
Citations 31
Table of Contents 14
Overview
Research and development is an integral force that drives the rapid progress
of artificial intelligence (AI). Every year, a wide range of academic, industry,
government, and civil society experts and organizations contribute to AI
R&D via a slew of papers, journal articles, and other AI-related publications,
conferences on AI or on particular subtopics like image recognition or
natural language processing, international collaboration across borders, and
the development of open-source software libraries. These R&D efforts are
diverse in focus and geographically dispersed.
Another key feature of AI R&D, making it somewhat distinct from other areas
of STEM research, is its openness. Each year, thousands and thousands of
AI publications are released in the open source, whether at conferences
or on file-sharing websites. Researchers will openly share their findings at
conferences; government agencies will fund AI research that ends up in the
open source; and developers use open software libraries, freely available to
the public, to produce state-of-the-art AI applications. This openness also
contributes to the globally interdependent and interconnected nature of
modern AI R&D.
This first chapter draws on multiple datasets to analyze key trends in the
AI research and development space in 2021. It first looks at AI publications,
including conference papers, journal articles, patents, and repositories.
It then analyzes AI conference attendance. And finally, it examines
AI open-source software libraries used in the R&D process.
Table of Contents Chapter 1 Preview 15

CHAPTER HIGHLIGHTS
• Despite rising geopolitical tensions, the United States and China had the greatest number of
cross-country collaborations in AI publications from 2010 to 2021, increasing five times since
2010. The collaboration between the two countries produced 2.7 times more publications
than between the United Kingdom and China—the second highest on the list.
• In 2021, China continued to lead the world in the number of AI journal, conference, and
repository publications—63.2% higher than the United States with all three publication types
combined. In the meantime, the United States held a dominant lead among major AI powers
in the number of AI conference and repository citations.
• From 2010 to 2021, the collaboration between educational and nonprofit organizations
produced the highest number of AI publications, followed by the collaboration between
private companies and educational institutions and between educational and government
institutions.
• The number of AI patents filed in 2021 is more than 30 times higher than in 2015, showing a
compound annual growth rate of 76.9%.

Artificial Intelligence CHAPTER 1: RESEARCH & DEVELOPMENT
Index Report 2022 1.1 Publications
This section draws on data from the Center for Security and Emerging Technology (CSET) at Georgetown University. CSET maintains
a merged corpus of scholarly literature that includes Digital Science’s Dimensions, Clarivate’s Web of Science, Microsoft Academic
Graph, China National Knowledge Infrastructure, arXiv, and Papers with Code. In that corpus, CSET applied a classifier to identify
English-language publications related to the development or application of AI and ML since 2010.1
1.1 PUBLICATIONS2
OV E R V I E W Total Number of AI Publications
The figures below capture the total number of English- Figure 1.1.1 shows the number of AI publications in
language AI publications globally from 2010 to 2021—by the world. From 2010 to 2021, the total number of AI
type, affiliation, cross-country collaboration, and cross- publications doubled, growing from 162,444 in 2010 to
industry collaboration. The section also breaks down 334,497 in 2021.
publication and citation data by region for AI journal
articles, conference papers, repositories, and patents.
NUMBER of AI PUBLICATIONS in the WORLD, 2010–21

Source: Center for Security and Emerging Technology, 2021 | Chart: 2022 AI Index Report
350
334.50
300
Number of AI Publications (in thousands)
250
200
150
100
50
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.1
1 See the Appendix for more information on CSET’s methodology. Due to the change in data provider and classification method, the publication trend/data may be different from past reports. For more
on the challenge of defining AI and correctly capturing relevant bibliometric data, see the AI Index team’s discussion in the paper “Measurement in AI Policy: Opportunities and Challenges.”
2 The number of AI publications in 2021 in this section may be lower than the actual count due to the lag in the collection of publication metadata by aforementioned databases.

By Type of Publication Books, book chapters, theses, and unknown document

Figure 1.1.2 shows the variation overtime in the types of types comprised the remaining 10.1% of publications.
AI publications released globally. In 2021, 51.5% of all AI While journal and repository publications have grown
documents published were journal articles, 21.5% were 2.5 and 30 times, respectively, in the past 12 years, the
conference papers, and 17.0% were from repositories. number of conference papers has declined since 2018.
NUMBER of AI PUBLICATIONS by TYPE, 2010–21

180 172.11, Journal
160
140
120
100
80
71.92, Conference
60 56.73, Repository
40
20.60, Book Chapter

20 10.84, Unknown
1.16, Book
0 1.14, Thesis
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.2

By Field of Study Other areas strongly influenced by deep learning, such

Figure 1.1.3 shows that publications in pattern recognition as computer vision, data mining and natural language
and machine learning have more than doubled since 2015. processing, have shown smaller increases.
NUMBER of AI PUBLICATIONS by FIELD of STUDY (EXCLUDING OTHER AI), 2010–21

51.69, Pattern Recognition

50
39.93, Machine Learning

40
30
24.80, Computer Vision
20
18.30, Algorithm
15.27, Data Mining
13.43, Natural Language Processing
10 9.70, Human–Computer Interaction

8.45, Control Theory
5.78, Linguistics
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.3

By Sector United Kingdom (Figure 1.1.4d).3 The education sector

This section shows the number of AI publications affiliated dominates in each of the regions. The level of company
with industry, education, government, and nonprofit in participation is highest in the United States, then in the
the world (Figure 1.1.4a), China (Figure 1.1.4b), the United European Union. China is the only area in which the share
States (Figure 1.1.4c), and the European Union plus the of education has been rising.
AI PUBLICATIONS (% of TOTAL) by SECTOR, 2010–21

60% 59.58%, Education
50%
AI Publications (% of Total)
40%
30%
20% 20.77%, Unknown
11.27%, Nonpro t
10%
5.21%, Company
3.17%, Government
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.4a
AI PUBLICATIONS in UNITED STATES (% of TOTAL) by SECTOR, 2010–21

60%
57.63%, Education
50%
40%
30%
20%
17.63%, Unknown
12.36%, Nonpro t
10% 9.76%, Company
2.62%, Government
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.4b
3 The categorization is adapted based on the Global Research Identifier Database (GRID). See definitions of each category here. Healthcare, including hospitals and facilities,
are included under nonprofit. Publications affiliated with state-sponsored universities are included in the education sector.

AI PUBLICATIONS in CHINA (% of TOTAL) by SECTOR, 2010–21

60% 60.24%, Education
50%
40%
30%
23.74%, Unknown
20%
10%
8.47%, Nonpro t
3.93%, Company
0% 3.62%, Government
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.4c
AI PUBLICATIONS in EUROPEAN UNION and UNITED KINGDOM (% of TOTAL) by SECTOR, 2010–21

60%
54.81%, Education
50%
40%
30%
20% 19.76%, Unknown

16.14%, Nonpro t
10%
5.68%, Company
3.61%, Government
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.4d

Cross-Country Collaboration
Cross-border collaborations between academics, By far, the greatest
researchers, industry experts, and others are a key
component of modern STEM development that accelerate
number of collaborations
the dissemination of new ideas and the growth of research in the past 12 years
teams. Figures 1.1.5a and 1.1.5b depict the top cross-
country AI collaborations from 2010 to 2021. CSET counted
took place between the
cross-country collaborations as distinct pairs of countries United States and China,
across authors for each publication (e.g., four U.S. and
four Chinese-affiliated authors on a single publication are increasing five times
counted as one U.S.-China collaboration; two publications since 2010.
between the same authors counts as two collaborations).
By far, the greatest number of collaborations in the past

12 years took place between the United States and China,
increasing five times since 2010. The next largest set of
collaborations is between the United Kingdom and both the
United States and China, which have increased more than
three times since 2010. In 2021, the number of collaborations
between the United States and China was 2.7 times greater
than between the United Kingdom and China.
UNITED STATES and CHINA COLLABORATIONS in AI PUBLICATIONS, 2010–21

10
9.66
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.5a

CROSS-COUNTRY COLLABORATIONS in AI PUBLICATIONS (EXCLUDING U.S. and CHINA), 2010–21

3.56, United Kingdom and China

3.34, United States and United Kingdom
3
2.63, China and Australia

2.44, United States and Germany
2.37, United States and Canada
2 2.16, United States and Australia
1.42, United States and France
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.5b
Cross-Sector Collaboration (29,839), followed by companies and educational institutions

The increase in AI research outside of universities has brought (11,576), and governments and educational institutions
collaborations between universities and other industries. (8,087). There were 2.5 times as many collaborations
Figure 1.1.6 shows that in 2021, educational institutions between educational institutions and nonprofits in 2021 as
and nonprofits had the greatest number of collaborations between educational institutions and companies.
CROSS-SECTOR COLLABORATIONS in AI PUBLICATIONS, 2010–21

29.84, Education and Nonpro t

30
25
20
15
11.58, Company and Education

10
8.09, Education and Government
2.48, Government and Nonpro t

1.83, Company and Nonpro t
0 0.66, Company and Government
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.6

A I J O U R N A L P U B L I CAT I O N S since 2015 (Figure 1.1.7). As a percentage of all journal

publications, as captured in Figure 1.1.8, AI journal
Overview publications in 2021 were about 2.5% of all publications,
After growing only slightly from 2010 to 2015, the compared to 1.5% in 2010.
number of AI journal publications grew almost 2.5 times
NUMBER of AI JOURNAL PUBLICATIONS, 2010–21

172.11
Number of AI Journal Publications (in thousands)
150
100
50
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.7
AI JOURNAL PUBLICATIONS (% of TOTAL JOURNAL PUBLICATIONS), 2010–21

2.53%
2.50%
AI Journal Publications (% of Total Journal Publications)
2.00%
1.50%
1.00%
0.50%
0.00%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.8

By Region4 In addition, South Asia

Figure 1.1.9 shows the share of AI journal publications
by region between 2010 and 2021. In 2021, East Asia and the Middle East and
and Pacific leads with 42.9%, followed by Europe and
Central Asia (22.7%) and North America (15.6%). In
North Africa saw the most
addition, South Asia and the Middle East and North significant growth as
Africa saw the most significant growth as their number
of AI journal publications grew around 12 and 7 times,
their number of AI journal
respectively, in the last 12 years. publications grew around
12 and 7 times, respectively,
in the last 12 years.
AI JOURNAL PUBLICATIONS (% of WORLD TOTAL) by REGION, 2010–21

50%
AI Journal Publications (% of World Total)
42.87%, East Asia and Paci c

40%
30%
22.70%, Europe and Central Asia

20%
15.63%, North America
10%
7.99%, South Asia
6.18%, Middle East and North Africa
3.57%, Latin America and the Caribbean
0% 1.06%, Sub-Saharan Africa
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.9
4 Regions in this chapter are classified according to the World Bank analytical grouping.

By Geographic Area5 powers. China has remained the leader throughout, with
Figure 1.1.10 breaks down the share of AI journal 31.0% in 2021, followed by the European Union plus the
publications over the past 12 years by three major AI United Kingdom at 19.1% and the United States at 13.7%
AI JOURNAL PUBLICATIONS (% of WORLD TOTAL) by GEOGRAPHIC AREA, 2010–21

40%
AI Journal Publications (% of World Total)
31.04%, China
30%
20% 19.05%, European Union and United Kingdom
13.67%, United States
10%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.10
5 Geographic areas in this chapter combine the number of publications between the European Union and the United Kingdom to reflect the historically strong association between them with regards to
research collaboration.

Citations
On the number of citations of AI journal The three geographic areas
publications, China’s share has gradually increased
while those of the European Union plus the United
combined accounted for
Kingdom and the United States have decreased. more than 66% of the total
The three geographic areas combined accounted for
more than 66% of the total citations in the world.
citations in the world.
AI JOURNAL CITATIONS (% of WORLD TOTAL) by GEOGRAPHIC AREA, 2010–21

30%
27.84%, China
25%
AI Journal Citations (% of World Total)
21.13%, European Union and United Kingdom

20%
15%
10%
5%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.11

A I C O N F E R E N C E P U B L I CAT I O N S (Figure 1.1.12). Despite the decline in the total numbers,

however, the share of AI conference publications among
Overview total conference publications in the world has increased
The number of AI conference publications peaked in by more than five percentage points since 2010 (Figure
2019, and fell about 19.4% below the peak in 2021 1.1.13).
NUMBER of AI CONFERENCE PUBLICATIONS, 2010–21

Number of AI Conference Publications (in thousands)
80
71.92
60
40
20
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.12
AI CONFERENCE PUBLICATIONS (% of TOTAL CONFERENCE PUBLICATIONS), 2010–21

AI Conference Publications (% of Total Conference Publications)
17.83%
16%
12%
8%
4%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.13

By Region share represented by East Asia and Pacific continues to

Figure 1.1.14 shows the number of AI conference rise since taking the lead in 2014, accounting for 40.4%
publications by region. Similar to the trend in AI journal in 2021, followed by Europe and Central Asia (23.0%) and
publication, East Asia and Pacific, Europe and Central North America (19.0%). The percentage of AI conference
Asia, and North America account for the world’s highest publications in South Asia saw a noticeable rise in the
numbers of AI conference publications. Specifically, the past 12 years, from 4.0% in 2010 to 10.4% in 2021.
AI CONFERENCE PUBLICATIONS (% of WORLD TOTAL) by REGION, 2010–21

40% 40.39%, East Asia and Paci c

AI Conference Publications (% of World Total)
30%
20%
10% 10.44%, South Asia

2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.14

Index Report 2022
By Geographic Area greater lead than in 2020, while the European Union plus
In 2021, China produced the greatest share of the world’s the United Kingdom followed at 19.0% and the United
AI conference publications at 27.6%, opening an even States came in third at 16.9% (Figure 1.1.15).
AI CONFERENCE PUBLICATIONS (% of WORLD TOTAL) by GEOGRAPHIC AREA, 2010–21

27.64%, China
25%
AI Conference Publications (% of World Total)
20%
15%
10%
5%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.15

Citations the number of AI conference citations, with 29.5% in

Despite China producing the most AI conference 2021, followed by the European Union plus the United
publications in 2021, Figure 1.1.16 shows that the United Kingdom (23.3%) and China (15.3%).
States led among the major powers with respect to
AI CONFERENCE CITATIONS (% of WORLD TOTAL) by GEOGRAPHIC AREA, 2010–21

35%
30% 29.52%, United States

AI Conference Citations (% of World Total)
25%
20%
15%
15.32%, China
10%
5%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.16

A I R E P O S I TO R I E S work outside traditional avenues for publications. Those

repositories allow researchers to share their findings before
Overview submitting them to journals and conferences, which greatly
Publishing pre-peer-reviewed papers on repositories of accelerates the cycle of information discovery. The number
electronic preprints (such as arXiv and SSRN) has become of AI repository publications grew almost 30 times in the
a popular way among AI researchers to disseminate their past 12 years (Figure 1.1.17), now account for 15.3% of all
repository publications (Figure 1.1.18).
NUMBER of AI REPOSITORY PUBLICATIONS, 2010–21
56.73
50
Number of AI Repository Publications (in thousands)
20
10
1
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.17
AI REPOSITORY PUBLICATIONS (% of TOTAL REPOSITORY PUBLICATIONS), 2010–21
15.30%
AI Repository Publications (% of Total Repository Publications)
15%
10%
5%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.18

By Region repository publications in the world since 2014 while that

The analysis by region in Figure 1.1.19 shows that North of Europe and Central Asia has declined. Since 2013, the
America has maintained a steady lead in the share of AI share of East Asia and Pacific has grown significantly.
AI REPOSITORY PUBLICATIONS (% of WORLD TOTAL) by REGION, 2010–21

40%

AI Repository Publications (% of World Total)
30%
20%
10%
4.23%, South Asia

2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.19

By Geographic Area the United States accounted for 32.5% of the world’s AI
While the United States has held the lead in the percentage repository publications—a higher percentage compared
of AI repository publications in the world since 2011, China to journal and conference publications, followed by the
is catching up while the European Union plus the United European Union plus the United Kingdom (23.9%) and
Kingdom’s share continues to drop (Figure 1.1.20). In 2021, China (16.6%).
AI REPOSITORY PUBLICATIONS (% of WORLD TOTAL) by GEOGRAPHIC AREA, 2010–21

AI Repository Publications (% of World Total)
30%
20%
16.60%, China
10%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.20

Citations of overall citations in 2021, establishing a dominant

On the citations of AI repository publications, Figure 1.1.21 lead over the European Union plus the United Kingdom
shows that the United States tops the list with 38.6% (20.1%) and China (16.4%).
AI REPOSITORY CITATIONS (% of WORLD TOTAL) by GEOGRAPHIC AREA, 2010–21

40%
AI Repository Citations (% of World Total)
30%
20% 20.13%, European Union and United Kingdom
16.44%, China
10%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.21

A I PAT E N T S
This section draws on data from CSET and 1790 Analytics The number of patents
on patents relevant to AI development and applications—
indicated by Cooperative Patent Classification (CPC)/
filed in 2021 is more
International Patent Classification (IPC) codes and than 30 times higher
keywords. Patents were grouped by country and year and
then counted at the “patent family” level, before CSET
than in 2015, showing
extracted year values from the most recent publication a compound annual
date within a family.
growth rate of 76.9%.
Overview
Figure 1.1.22 captures the number of AI patents filed from
2010 to 2021. The number of patents filed in 2021 is more
than 30 times higher than in 2015, showing a compound
annual growth rate of 76.9%.
NUMBER of AI PATENT FILINGS, 2010–21

141.24
140
120
Number of AI Patents (in thousands)
100
80
60
40
20
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.22

By Region and Application Status Central Asia. In terms of granted patents in those regions,
Figure 1.1.23a breaks down AI patent filings by region. North America leads with 57.0%, followed by East Asia
The share of East Asia and Pacific took off in 2014 and and Pacific (31.0%), and Europe and Central Asia (11.3%)
led the rest of the world in 2021 with 62.1% of all patent (Figure 1.1.23b). The other regions combine to make up
applications, followed by North America and Europe and roughly 1% of world patents (Figure 1.1.23c).
AI PATENT FILINGS (% of WORLD TOTAL) by REGION, 2010–21

70%
60% 62.14%, East Asia and Paci c

AI Patent Filings (% of World Total)
50%
40%
30%
20%
10%

0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.23a

GRANTED AI PATENTS (% of WORLD TOTAL) by REGION, 2010–21 (Part 1)

80%
70%
Granted AI Patents (% of World Total)
60%
50%
40%

30%
20%

10%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.23b
GRANTED AI PATENTS (% of WORLD TOTAL) by REGION, 2010–21 (Part 2)

1.50%
1.25%
1.00%
0.75%
0.50%

0.25%
0.25%, South Asia

0.00%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.23c

By Geographic Area and Application Status United Kingdom. The United States, which files almost
Trends revealed by the regional analysis can also be all the patents in North America, does so at one-third the
observed in AI patent data broken down by geographic rate of China. Figure 1.1.24c shows that compared to the
area (Figure 1.1.24a and Figure 1.1.24b). China is now increasing numbers of AI patents applied and granted,
filing over half of the world’s AI patents and being granted China has far greater numbers of patent applications
about 6%, about the same as the European Union plus the (87,343 in 2021) than those granted (1,407 in 2021).
AI PATENT FILINGS (% of WORLD TOTAL) by GEOGRAPHIC AREA, 2010–21

60%
50% 51.69%, China

AI Patent Filings (% of World Total)
40%
30%
20%
10%

0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.24a
GRANTED AI PATENTS (% of WORLD TOTAL) by GEOGRAPHIC AREA, 2010–21
70%
60%
50%
40%
30%
20%
10%
5.90%, China
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.1.24b

AI PATENTS by APPLICATION STATUS by GEOGRAPHIC AREA, 2010–21

China European Union and United Kingdom United States

90
87.34
Application
80 Grant
Number of AI Patent Filings (in thousands)
70
60
50
40
30
19.61
20
9.45
10
4.88
1.41 1.81
0
2011 2016 2021 2011 2016 2021 2011 2016 2021

Figure 1.1.24c

Index Report 2022 1.2 Conferences
AI conferences are key venues for researchers to publish and communicate their work, as well as to connect with peers and
collaborators. Conference attendance is an indication of broader industrial and academic interest in a scientific field. In the past
20 years, AI conferences have grown not only in size but also in number and prestige. This section presents data on the trends in
attendance at major AI conferences, covering more conferences (16) than previous Index reports.
1.2 CONFERENCES
C O N F E R E N C E AT T E N DA N C E Figure 1.2.1 shows that attendance at top AI conferences
Similar to 2020, most AI conferences were offered in 2021 was relatively consistent with 2020, with more
virtually in 2021. Only the International Conference on than 88,000 participants worldwide. Figure 1.2.2 and
Robotics and Automation (ICRA) and the Conference Figure 1.2.3 show the attendance data for individual
on Empirical Methods in Natural Language Processing conferences, with 16 major AI conferences split into
(EMNLP) were held using a hybrid format. Conference two categories: large AI conferences with over 2,500
organizers reported measuring the exact attendance attendees and small AI conferences with fewer than
numbers at a virtual conference is difficult, as virtual 2,500 attendees.6
conferences allow for higher attendance of researchers
from all around the world.
NUMBER of ATTENDEES at SELECT AI CONFERENCES, 2010–21

Source: Conference Data, 2021 | Chart: 2022 AI Index Report
90 88.76
80
Number of Attendees (in thousands)
70
60
50
40
30
20
10
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.2.1
6 The International Conference on Machine Learning (ICML) used the number of session visitors as a proxy for the number of conference attendees, which explains the high attendance count in 2021.
The International Conference on Intelligent Robots and Systems (IROS) extended the virtual conference to allow users to watch events for up to three months, which explains the high attendance count
in 2020. For the AAMAS conference, the attendance in 2020 is based on the number of users on site reported by the platform that recorded the talks and managed the online conference, while the 2021
number is for total registrants.

ATTENDANCE at LARGE AI CONFERENCES, 2010–21

30
29.54, ICML
25
20
17.09, NeurIPS
15
10
8.24, CVPR
6.31, ICLR
5 5.01, ICCV
4.09, AAAI
3.76, EMNLP
2.59, IROS
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.2.2
ATTENDANCE at SMALL AI CONFERENCES, 2010–21

3
2.10, UAI
2
1.90, IJCAI
1.35, FaccT
1.08, AAMAS
1
1.00, ICRA
0.67, KR
0.51, ICAPS
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.2.3

WO M E N I N M AC H I N E L E A R N I N G Workshop Participants
( W I M L ) N E U R I P S WO R KS H O P The number of participants attending the WiML workshop
Founded in 2006, Women in Machine Learning is an has steadily increased since it was first introduced in
organization dedicated to supporting and increasing 2006. For the 2021 edition, Figure 1.2.4 shows an estimate
the impact of women in machine learning. This section of 1,486 attendees over all workshop sessions, counted
presents data from its annual technical workshop as the number of unique individuals who accessed the
colocated with NeurIPS. Starting in 2020, WiML has also virtual workshop platform at neurips.cc. The 2021 WiML
been hosting the Un-Workshop, which aims to advance Workshop at NeurIPS happened as multiple sessions
research via collaboration and interaction among over three days, which was a change in format from 2020.
participants from diverse backgrounds at ICML. As in 2020, the workshop was held virtually due to the
pandemic.
ATTENDANCE at NEURIPS WOMEN in MACHINE LEARNING WORKSHOP, 2010–21

Source: Women in Machine Learning, 2021 | Chart: 2022 AI Index Report
1,486
1400
1200
Number of Attendees
1000
800
600
400
200
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 1.2.4

Demographics Breakdown by Europe (19.9%), Asia (16.2%), and Africa (7.3%) (Figure
This section shows the continent of residence and 1.2.5). Figure 1.2.6 shows that Ph.D. students made
professional position breakdowns of the 2021 workshop up almost half of the survey participants, while the
participants based on a survey filled by participants share of university faculty is around 1.2%. Researcher
who consented to have such information aggregated. scientists/engineers, data scientists/engineers, and
Among the survey respondents, more than half of the software engineers were among the most commonly held
survey respondents were from North America, followed professional positions.
CONTINENT of RESIDENCE of PARTICIPANTS at NEURIPS WOMEN in MACHINE LEARNING WORKSHOP, 2021

North
53.40%
America
Europe 19.90%
Asia 16.20%
Africa 7.30%
South
1.00%
America
Australia 1.00%
0% 10% 20% 30% 40% 50%

% of Respondents Figure 1.2.5
PROFESSIONAL POSITIONS of PARTICIPANTS at NEURIPS WOMEN in MACHINE LEARNING WORKSHOP, 2021

PhD Student 49.40%
MSc Student 16.50%
Research Scientist / Engineer 14.20%
Undergraduate Student 10.20%
Data Scientist / Engineer 7.40%
Postdoctoral Researcher 5.70%
Software Engineer 2.30%
High School Student 0.60%
Professor (Post-Tenure) 0.60%
Professor (Pre-Tenure) 0.60%
0% 10% 20% 30% 40% 50%

% of Respondents Figure 1.2.6

Index Report 2022 1.3 AI Open-Source Software Libraries
A software library is a collection of computer code that is used to create applications and products. Popular AI-specific software
libraries—such as TensorFlow and PyTorch—help developers create their AI solutions quickly and efficiently. This section analyzes the
popularity of software libraries through GitHub data.
1.3 AI OPEN-SOURCE
SOFTWARE LIBRARIES
G I T H U B S TA R S GitHub open-source AI software library, OpenCV, which
Figures 1.3.1 and 1.3.2 reflect the number of users of was followed by Keras, PyTorch, and Scikit-learn.
GitHub open-source AI software libraries from 2015 to Figure 1.3.2 shows library popularity for libraries with
2021. TensorFlow remained by far the most popular in fewer than 40,000 GitHub stars—led by FaceSwap with
2021, with around 161,000 cumulative GitHub stars—a around 40,000 stars, followed by 100-Days-Of-ML-Code,
slight increase over 2020. TensorFlow was about three AiLearning, and BVLC/caffe.
times as popular in 2021 as the next-most-starred
NUMBER of GITHUB STARS by AI LIBRARY (OVER 40K STARS), 2014–21

Source: GitHub, 2021 | Chart: 2022 AI Index Report
160 160.7, TensorFlow

Number of Cumulative GitHub Stars (in thousands)
140
120
100
80
58.6, OpenCV
60 53.2, Keras
52.7, PyTorch
48.0, Scikit-learn
40 46.7, DeepLearning-500-questions
41.5, TenserFlow-Examples
20
2014 2015 2016 2017 2018 2019 2020 2021

Figure 1.3.1

Index Report 2022 1.3 AI Open-Source Software Libraries
NUMBER of GITHUB STARS by AI LIBRARY (UNDER 40K STARS), 2014–21

Source: GitHub, 2021 | Chart: 2022 AI Index Report
39.88, faceswap
40
33.58, 100-Days-Of-ML-Code
Number of Cumulative GitHub Stars (in thousands)
35 32.27, AiLearning
32.14, BVLC/ca e
32.02, Real-Time-Voice-Cloning
30 32.00, deeplearningbook-chinese
31.28, Deep Learning Papers Reading Roadmap
30.26, DeepFaceLab
25
20
15
10
2014 2015 2016 2017 2018 2019 2020 2021

Figure 1.3.2

Index Report 2022
CHAPTER 2:
Technical
Performance
Index Report 2022 CHAPTER 2: TECHNICAL PERFORMANCE
CHAPTER 2:
Chapter Preview
Overview 50 Medical Image Segmentation 61
Chapter Highlights 51 CVC-ClinicDB and Kvasir-SEG 61
Face Detection and Recognition 62
2.1 COMPUTER VISION—IMAGE 52 National Institute of Standards
Image Classification 52 and Technology (NIST) Face
Recognition Vendor Test (FRVT) 62
ImageNet 52
Face Detection: Effects of Mask-Wearing 63
ImageNet: Top-1 Accuracy 52
Face Recognition Vendor Test (FRVT):
ImageNet: Top-5 Accuracy 52
Face-Mask Effects 63
Image Generation 54
Highlight: Masked Labeled Faces
STL-10: Fréchet Inception in the Wild (MLFW) 64
Distance (FID) Score 54
Visual Reasoning 65
CIFAR-10: Fréchet Inception
Visual Question Answering
Distance (FID) Score 55
(VQA) Challenge 65
Deepfake Detection 56
FaceForensics++ 56 2.2 COMPUTER VISION—VIDEO 67
Celeb-DF 57 Activity Recognition 67
Human Pose Estimation 57 Kinetics-400, Kinetics-600, Kinetics-700 67
Leeds Sports Poses: Percentage ActivityNet: Temporal Action
of Correct Keypoints (PCK) 58 Localization Task 69
Human3.6M: Average Mean Object Detection 70
Per Joint Position Error (MPJPE) 59
Common Object in Context (COCO) 71
Semantic Segmentation 60
You Only Look Once (YOLO) 72
Cityscapes 60
Visual Commonsense Reasoning (VCR) 73
CHAPTER 2: CHAPTER PREVIEW (CONT’D)
2.3 LANGUAGE 74 2.4 SPEECH 86

English Language Understanding 74 Speech Recognition 86
SuperGLUE 74 Transcribe Speech: LibriSpeech
Stanford Question Answering (Test-Clean and Other Datasets) 86
Dataset (SQuAD) 75 VoxCeleb 87
Reading Comprehension Dataset
Requiring Logical Reasoning (ReClor) 76 2.5 RECOMMENDATION 88
Text Summarization 78 Commercial Recommendation:
arXiv 78 MovieLens 20M 88
PubMed 79 Click-Through Rate Prediction: Criteo 89
Natural Language Inference 80

2.6 REINFORCEMENT LEARNING 90
Stanford Natural Language
Inference (SNLI) 80 Reinforcement Learning Environments 90
Abductive Natural Language Arcade Learning Environment: Atari-57 90

Inference (aNLI) 81 Procgen 91
Sentiment Analysis 82 Human Games: Chess 93
SemEval 2014 Task 4 Sub Task 2 82
Machine Translation (MT) 83 2.7 HARDWARE 94
WMT 2014, English-German MLPerf: Training Time 94

and English-French 84 MLPerf: Number of Accelerators 96
Number of Commercially IMAGENET: Training Cost 97
Available MT Systems 85
2.8 ROBOTICS 98
Price Trends in Robotic Arms 98
AI Skills Employed by Robotics Professors 99
Overview
This year, the technical performance chapter includes more analysis
than ever before of the technical progress in various subfields of
artificial intelligence, including trends in computer vision, language,
speech, recommendation, reinforcement learning, hardware, and
robotics. It uses a number of quantitative measurements, from
common AI benchmarks and prize challenges to a field-wide survey,
to highlight the development of top-performing AI systems.

CHAPTER HIGHLIGHTS
• Data, data, data: Top results across technical benchmarks have increasingly relied on the use of
extra training data to set new state-of-the-art results. As of 2021, 9 state-of-the-art AI systems
out of the 10 benchmarks in this report are trained with extra data. This trend implicitly favors
private sector actors with access to vast datasets.
• Rising interest in particular computer vision subtasks: In 2021, the research community saw
a greater level of interest in more specific computer vision subtasks, such as medical image
segmentation and masked-face identification. For example, only 3 research papers tested
systems against the Kvasir-SEG medical imaging benchmark prior to 2020. In 2021, 25
research papers did. Such an increase suggests that AI research is moving toward research that
can have more direct, real-world applications.
• AI has not mastered complex language tasks, yet: AI already exceeds human performance levels
on basic reading comprehension benchmarks like SuperGLUE and SQuAD by 1%–5%. Although
AI systems are still unable to achieve human performance on more complex linguistic tasks
such as abductive natural language inference (aNLI), the difference is narrowing. Humans
performed 9 percentage points better on aNLI in 2019. As of 2021, that gap has shrunk to 1.
• Turn toward more general reinforcement learning: For the last decade, AI systems have
been able to master narrow reinforcement learning tasks in which they are asked to maximize
performance in a specific skill, such as chess. The top chess software engine now exceeds
Magnus Carlsen’s top ELO score by 24%. However, in the last two years AI systems have also
improved by 129% on more general reinforcement learning tasks (Procgen) in which they
must operate in novel environments. This trend speaks to the future development of AI systems
that can learn to think more broadly.
• AI becomes more affordable and higher performing: Since 2018, the cost to train an image
classification system has decreased by 63.6%, while training times have improved by 94.4%.
The trend of lower training cost but faster training time appears across other MLPerf task
categories such as recommendation, object detection and language processing, and favors the
more widespread commercial adoption of AI technologies.
• Robotic arms are becoming cheaper: An AI Index survey shows that the median price of
robotic arms has decreased by 46.2% in the past five years—from $42,000 per arm in 2017
to $22,600 in 2021. Robotics research has become more accessible and affordable.

Artificial Intelligence CHAPTER 2: TECHNICAL PERFORMANCE
Index Report 2022 2.1 Computer Vision–Image
Computer vision is the subfield of AI that teaches machines to understand images and videos. There is a wide range of computer
vision tasks, such as image classification, object recognition, semantic segmentation, and face detection. As of 2021, computers can
outperform humans on a plethora of computer vision tasks. Computer vision technologies have a variety of important real-world
applications, such as autonomous driving, crowd surveillance, sports analytics, and video-game creation.
2.1 COMPUTER VISION—IMAGE

I M AG E C L A S S I F I CAT I O N A DEMONSTRATION OF IMAGE CLASSIFICATION

Image classification refers to the ability of machines to Source: Krizhevsky, 2020
categorize what they see in images (Figure 2.1.1). In a

practical sense, image recognition systems can help cars
identify objects in their surroundings, doctors detect
tumors, and factory managers spot production defects.
The past decade has seen tremendous advances in the
technical capacity of image recognition systems, especially
as researchers have embraced more machine learning
techniques. Moreover, progress in algorithmic, hardware,
and data technologies has meant that image recognition
has become more affordable, widely applicable, and
accessible than ever before.
ImageNet
ImageNet is a database that includes over 14 million images
across 20,000 categories publicly available to researchers
Figure 2.1.1
working on image classification problems. Created in 2009,
ImageNet is now one of the most common ways scientists
benchmark improvement on image classification.
the top pretrained system was CoAtNets, produced by
ImageNet: Top-1 Accuracy researchers on the Google Brain Team.
Benchmarking on ImageNet is measured through accuracy
metrics, which quantify how frequently AI systems assign ImageNet: Top-5 Accuracy
the right labels to the given images. Top-1 accuracy Top-5 accuracy considers whether any of the model’s 5
measures the rate at which the top prediction made by highest probability answers align with the image label. As
a classification model for a given image matches the highlighted in Figure 2.1.3, AI systems presently achieve
image’s actual target label. In recent years, it has become near perfect Top-5 estimation. Currently, the state-of-the-
increasingly common to improve the performance of art performance on Top-5 accuracy with pretraining is
systems on ImageNet by pretraining them with additional 99.0%, achieved in November 2021 by Microsoft Cloud and
data from other image datasets. Microsoft AI’s Florence-CoSwim-H model.
As of late 2021, the top image classification system makes Improvements in Top-5 accuracy on ImageNet seem to be
on average 1 error for every 10 classification attempts plateauing, which is perhaps unsurprising. If your system
on Top-1 accuracy compared to an average of 4 errors is classifying correctly 98 or 99 out of 100 times, there is
for every 10 attempts in late 2012 (Figure 2.1.2). In 2021, only so much higher you can go.

IMAGENET CHALLENGE: TOP-1 ACCURACY

Source: Papers with Code, 2021; arXiv, 2021 | Chart: 2022 AI Index Report
90.88%, With Extra Training Data

90%
87.80%, Without Extra Training Data

Top-1 Accuracy (%)
80%
70%
60%
2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 2.1.2
IMAGENET CHALLENGE: TOP-5 ACCURACY

100% 99.02%, With Extra Training Data

94.90 %, Human Baseline
95%
Top-5 Accuracy (%)
90%
85%
80%
2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 2.1.3

I M AG E G E N E R AT I O N
Image generation is the task of generating images that are systems generate different design patterns), and healthcare
indistinguishable from real ones. Image generation can be (image generators can synthetically create novel drug
widely useful in generative domains where visual content compounds). Figure 2.1.4 illustrates progress made in
has to be created, for example entertainment (companies image generation by presenting several human faces that
like NVIDIA have already used image generators to create were synthetically generated by AI systems in the last year.
virtual worlds for gaming), fashion (designers can let AI
GAN PROGRESS ON FACE GENERATION

Source: Goodfellow et al., 2014; Radford et al., 2016; Liu & Tuzel, 2016;
Karras et al., 2018; Karras et al., 2019; Goodfellow, 2019; Karras et al.,
2020; AI Index, 2021; Vahdat et al., 2021
2021 Figure 2.1.4
STL-10: Fréchet Inception Distance (FID) Score Figure 2.1.5 documents the gains generative models have
The Fréchet Inception Distance score tracks the similarity made in FID on the STL-10 dataset, one of the most widely
between an artificially generated set of images and the cited datasets in computer vision. The state-of-the-art
real images from which it was generated. A low score model on STL-10 developed by researchers at the Korea
means that the generated images are more similar to Advanced Institute of Science and Technology as well as the
the real ones, and a score of zero indicates that the fake University of Seoul posted a FID score of 7.7, significantly
images are identical to the real ones. better than the state-of-the-art result from 2020.
STL-10: FRÉCHET INCEPTION DISTANCE (FID) SCORE

Fréchet Inception Distance (FID) Score
30
20
10
7.71
0
2018 2019 2020 2021
Figure 2.1.5

CIFAR-10: Fréchet Inception Distance (FID) Score The FID scores achieved by the top image generation
Progress on image generation can also be benchmarked models are much lower on CIFAR-10 than STL-10. This
on CIFAR-10, a dataset of 60,000 color images across 10 difference is likely attributable to the fact that CIFAR-10
different object classes. The state-of-the-art results on contains images of much lower resolution (32 x 32 pixels)
CIFAR-10 posted in 2021 were achieved by researchers than those on STL-10 (96 x 96 pixels).
from NVIDIA.
CIFAR-10: FRÉCHET INCEPTION DISTANCE (FID) SCORE

30
Fréchet Inception Distance (FID) Score
20
10
2.10
0
2017 2018 2019 2020 2021
Figure 2.1.6

D E E P FA K E D E T E CT I O N dataset that contains approximately 1,000 original video

Many AI systems can now generate fake images that are sequences sourced from YouTube videos. Progress on
indistinguishable from real ones. A related technology FaceForensics++ is measured in terms of accuracy: the
involves superimposing one person’s face onto another, percentage of altered images an algorithm can correctly
creating a so-called “deepfake.” Deepfakes are used identify.
for purposes ranging from advertising to generating Although FaceForensics++ was introduced in 2019,
misogynistic pornography and disinformation (in 2018, researchers have tested previously existing deepfake
for example, a deepfake video of Barack Obama uttering detection methods on the dataset in order to track
profanities about Donald Trump was circulated online over progress over time in deepfake detection (Figure 2.1.7).
2 million times). In the last few years, AI researchers have In the last decade, AI systems have become better and
sought to keep up with improving deepfake technologies by better at detecting deepfakes. In 2012, the top-performing
crafting stronger deepfake detection algorithms. systems could correctly identify 69.9% of deepfakes
FaceForensics++ across all four FaceForenics++ datasets. In 2021, that
FaceForensics++ is a deepfake detection benchmarking number increased to 97.7%.1
FACEFORENSICS++: ACCURACY
Source: arXiv, 2021 | Chart: 2022 AI Index Report
100% 99.98%, Face2Face

99.47%, DeepFake
98.27%, FaceSwap
93.25%, NeuralTextures
90%
Accuracy (%)
80%
70%
60%
2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 2.1.7
1 These numbers were taken by averaging performance across all four FaceForensics++ datasets.

Celeb-DF Top detection models perform significantly worse (by 20

The Celeb-DF deepfake detection dataset is composed of percentage points) on Celeb-DF than FaceForensics++,
590 original celebrity YouTube videos manipulated into suggesting that Celeb-DF is a more challenging dataset to
5,639 deepfakes. Celeb-DF was introduced in 2019. In test out techniques on. As deepfake technologies continue
2021, the top score on Celeb-DF was 76.9 and came from to improve in the upcoming years, it will be important
researchers at the University of Science and Technology of to continue monitoring progress on Celeb-DF and other
China and Alibaba group (Figure 2.1.8). similarly challenging deepfake detection datasets.
CELEB-DF: AREA UNDER CURVE SCORE (AUC)

80
76.88
75
Area Under Curve Score (AUC)
70
65
60
2018 2019 2020 2021
Figure 2.1.8
H U M A N P O S E E S T I M AT I O N Human pose estimation can be used to facilitate activity

Human pose estimation is the task of estimating different recognition for purposes such as sports analytics, crowd
positions of human body joints (arms, head, torso, etc.) surveillance, CGI development, virtual environment design,
from a single image (Figure 2.1.9), and then combining these and transportation (for example, identifying the body
estimates to correctly label the pose the human is taking. language signs of an airport runway controller).
A DEMONSTRATION OF
HUMAN POSE ESTIMATION
Source: Cao et al., 2019
Figure 2.1.9

Leeds Sports Poses: Percentage of Correct In 2021, the top-performing human pose estimation
Keypoints (PCK) model correctly identified 99.5% of keypoints on Leeds
The Leeds Sports Poses dataset contains 2,000 images Sports Poses (Figure 2.1.10). Given that maximum
collected from Flickr of athletes playing a sport. Each performance on Leeds Sports is 100.0%, more challenging
image includes information on 14 different body joint benchmarks for human pose estimation will have to be
locations. Performance on the Leeds Sports Poses developed, as we are close to saturating the benchmark.
benchmark is assessed by the percentage of correctly
estimated keypoints.
LEEDS SPORTS POSES: PERCENTAGE of CORRECT KEYPOINTS (PCK)

100% 99.50%
Percentage of Correct Keypoints (PCK)
90%
80%
70%
2014 2015 2016 2017 2018 2019 2020 2021
Figure 2.1.10

Human3.6M: Average Mean Per Joint Position on Human3.6M is measured in average mean per joint
Error (MPJPE) position error in millimeters, which is the average difference
3D human pose estimation is a more challenging type of between an AI model’s position estimations and the actual
pose estimation, where AI systems are asked to estimate position annotation.
poses in a three- rather than two-dimensional space. The
In 2014, the top-performing model was making an average
Human3.6M dataset tracks progress in 3D human pose
per joint error of 16 centimeters, half the size of a standard
estimation. Human3.6M is a collection of over 3.6 million
school ruler. In 2021, this number fell to 1.9 centimeters,
images of 17 different types of human poses (talking on
less than the size of an average paper clip.
the phone, discussing, and smoking, etc.). Performance
HUMAN3.6M: AVERAGE MEAN PER JOINT POSITION ERROR (MPJPE)

150
120
Average MPJPE (mm)
90
60
22.70, Without Extra Training Data

30
18.70, With Extra Training Data

0
2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 2.1.11

S E M A N T I C S E G M E N TAT I O N A DEMONSTRATION OF SEMANTIC SEGMENTATION

Source: Visual Object Classes Challenge, 2012
Semantic segmentation is the task of assigning
individual image pixels a category (such as person,
bicycle, or background) (Figure 2.1.12). A plethora
of real world domains require pixel-level image
segmentation such as autonomous driving (identifying
which parts of the image a car sees are pedestrians and
which parts are roads), image analysis (distinguishing Figure 2.1.12
the foreground and background in photos), and

medical diagnosis (segmenting tumors in lungs).
Cityscapes metric, with a higher IoU score corresponding to better

The Cityscapes dataset contains images of urban street segmentation accuracy. In practical terms, a higher score
environments from 50 cities, taken during the daytime in means that a greater proportion of the image segments
different seasons, and it allows for evaluation on a wide predicted by the model overlap with the image’s actual
range of semantic segmentation tasks (instance-level, segments.
panoptic, and 3D vehicle).
The top-performing AI systems on Cityscapes in 2021
The task to which most researchers submit is pixel-level reported scores that are 14.6 percentage points higher than
semantic labeling, a challenge in which AI systems must those in 2015. As with other computer vision tasks, the top
semantically label an image on a per-pixel level. Challengers performing models on Cityscapes have been pretrained on
are evaluated on the intersection-over-union (IoU) additional training data in the last few years.
CITYSCAPES CHALLENGE, PIXEL-LEVEL SEMANTIC LABELING TASK: MEAN INTERSECTION-OVER-UNION (IOU)

Source: Cityscapes Challenge, 2021 | Chart: 2022 AI Index Report
90%

Mean Intersection-Over-Union (mIoU)
80%
70%
60%
2014 2015 2016 2017 2018 2019 2020 2021
Figure 2.1.13

M E D I CA L I M AG E S E G M E N TAT I O N A DEMONSTRATION OF KIDNEY SEGMENTATION

Source: Kidney and Kidney Tumor Segmentation, 2021
Medical image segmentation refers to the ability of AI
systems to segment objects of interest, such as organs,
lesions, or tumors, in medical images (Figure 2.1.14).
Technical progress in this task is vital to streamlining
medical diagnoses. Advances in medical image
segmentation mean doctors can spend less time on
diagnosis and more time treating patients.
CVC-ClinicDB and Kvasir-SEG

CVC-ClinicDB is a dataset that includes over 600 high-
resolution images from 31 colonoscopies. Kvasir-SEG is a
public dataset of 1,000 high-resolution gastrointestinal polyp
images that were manually segmented by doctors and cross-
verified by professional gastroenterologists. Both datasets
Figure 2.1.14
are used to track progress in medical image segmentation.
Performance is measured in mean DICE, which represents the
2020 (Figure 2.1.15). Similar progress has been made on
average overlap between the polyp segments identified by an
Kvasir-SEG, where presently the top-performing AI model
AI system and the actual polyp segments.
can accurately segment gastrointestinal polyps at a rate
AI systems are now capable of correctly segmenting of 92.2%. The top results on both the CVC-ClinicDB and
colonoscopy polyps at a rate of 94.2% on CVC-ClinicDB, Kvasir-SEG benchmarks were achieved by the MSRF-Net
representing an 11.9 percentage point improvement model, one of the first convolutional neural networks
since 2015, and a 1.8 percentage improvement since designed specifically for medical image segmentation.
CVC-CLINICDB: MEAN DICE KVASIR-SEG: MEAN DICE

Source: Papers with Code, 2021; arXiv, 2021 |Chart: 2022 AI Index Report Source: Papers with Code, 2021; arXiv, 2021 | Chart: 2022 AI Index Report
95% 95%
94.20%
92.17%
90% 90%
Mean DICE
Mean DICE
85% 85%
80% 80%
2015 2017 2019 2021 2015 2017 2019 2021
Figure 2.1.15a Figure 2.1.15b

The Kvasir-SEG benchmark also points to the explosion of which makes the technology extremely appealing to
interest in medical image segmentation. Prior to 2020, the militaries and governments all over the world (e.g., 18 out
dataset was referenced in only three academic papers. In of 24 U.S. government agencies are already using some
2020 that number rose to six, and in 2021 it shot up to 25. kind of facial recognition technology).
Last year also saw the hosting of KiTS21 (the Kidney and
Kidney Tumor Segmentation Challenge), which challenged National Institute of Standards and Technology
medical researchers from academia and industry to create (NIST) Face Recognition Vendor Test (FRVT)
the best systems for automatically segmenting renal The National Institute of Standards and Technology’s
tumors and the surrounding anatomy of kidneys. Face Recognition Vendor Test measures how well facial
recognition algorithms perform on a variety of homeland
security and law enforcement tasks, such as face
FACE DETECTION AND RECOGNITION recognition across photojournalism images, identification
In facial detection, AI systems are tasked with identifying of child trafficking victims, deduplication of passports,
individuals in images or videos. Although facial recognition and cross-verification of visa images. Progress on facial
technology has existed for several decades, the technical recognition algorithms is measured according to the false
progress in the last few years has been significant. Some of non-match rate (FNMR) or the error rate (the frequency
today’s top-performing facial recognition algorithms have a with which a model fails to match an image to a person).
near 100% success rate on challenging datasets.
In 2017, some of the top-performing facial recognition
Facial recognition can be used in transportation to facilitate algorithms had error rates of over 50.0% on certain FRVT
cross-border travel, in fraud prevention to protect sensitive tests. As of 2021, none has posted an error rate greater than
documents, and in online proctoring to identify illicit 3.0%. The top-performing model across all datasets in 2021
examination behavior. The greatest practical promise of (visa photos) registered an error rate of 0.1%, meaning that
facial recognition, however, is in its potential to aid security, for every 1,000 faces, the model correctly identified 999.
NATIONAL INSTITUTE OF STANDARDS AND TECHNOLOGY (NIST) FACE RECOGNITION VENDOR TEST (FRVT):
VERIFICATION ACCURACY by DATASET
Source: National Institute of Standards and Technology, 2021 | Chart: 2022 AI Index Report
0.5
False Non-Match Rate: FMNR (Log-Scale)
0.2
0.1
0.05
0.0297, WILD Photos FNMR @ FMR 0.00001

0.02
0.01
0.005
0.0044, BORDER Photos FNMR @ FMR = 0.000001
0.0023, VISABORDER Photos FNMR@FMR 0.000001
0.0022, MUGSHOT Photos FNMR @ FMR 0.00001
0.002
0.0021, MUGSHOT Photos FNMR @ FMR 0.00001 DT>=12 YRS
0.001 0.0013, VISA Photos FNMR @ FMR 0.000001
2017 2018 2019 2020 2021

Figure 2.1.16

FACE DETECTION:
EFFECTS OF MASK-WEARING Although facial recognition
Face Recognition Vendor Test (FRVT):
technology has existed
Face-Mask Effects for several decades, the
Facial recognition has become a more challenging
task with the onset of the COVID-19 pandemic and
technical progress in the
accompanying mask mandates. The face-mask effects test last few years has been
asks AI models to identify faces on two datasets of visa
border photos, one of which includes masked faces, the significant. Some of today’s
other which does not. top-performing facial
Three important trends can be gleaned from the FRVT
recognition algorithms have
face-mask test: (1) Facial recognition systems still perform
relatively well on masked faces; (2) the performance on a near 100% success rate on
masked faces is worse than on non-masked faces; and (3)
the gap in performance has narrowed since 2019.
challenging datasets.
NIST FRVT FACE MASK EFFECTS: FALSE-NON MATCH RATE

Source: National Institute of Standards and Technology, 2021 | Chart: 2022 AI Index Report
0.025
False Non-Match Rate: FMNR (Log-Scale)
0.020
0.015
0.014, Masked
0.010
0.005
0.002, Non-masked
0.000
2019 2020 2021
Figure 2.1.17

Masked Labeled Faces in the Wild (MLFW)

In 2021, researchers from the
Beijing University of Posts and
Telecommunications released a
facial recognition dataset of 6,000
masked faces in response to the new
recognition challenges posed by
large-scale mask-wearing.
EXAMPLES OF MASKED FACES IN

THE MASKED LABELED FACES IN
THE WILD (MLFW) DATABASE
Source: Wang et al., 2021
Figure 2.1.18
As part of the dataset release, the researchers ran a series of existing state-of-the-art detection
algorithms on a variety of facial recognition datasets, including theirs, to determine how much detection
performance decreased when faces were masked. Their estimates suggest that top methods perform 5
to 16 percentage points worse on masked faces compared to unmasked ones. These findings somewhat
confirm the insights from the FRVT face-mask tests: Performance deteriorates when masks are included,
but not by an overly significant degree.
STATE-OF-THE-ART FACE DETECTION METHODS on MASKED LABELED FACES IN THE WILD (MLFW): ACCURACY
Source: Wang et. al, 2021 | Chart: 2022 AI Index Report
Face Detection Method / Dataset

ArcFace1 ArcFace3 Arcface4 CosFace2 Curricularface5 SFace6
100%
90% 91% 91%

90%
85%
83%
Accuracy (%)
80%
75%
70%
60%
50%
CALFW11
CPLFW10
LFW7
MLFW
CALFW11
CPLFW10
LFW7
MLFW
CALFW11
CPLFW10
LFW7
MLFW
CALFW11
CPLFW10
LFW7
MLFW
CALFW11
CPLFW10
LFW7
MLFW
CALFW11
CPLFW10
LFW7
MLFW
SLLFW8
SLLFW8
SLLFW8
SLLFW8
SLLFW8
SLLFW8
Figure 2.1.19

VISUAL REASONING
Visual reasoning assesses how well AN EXAMPLE OF A VISUAL REASONING TASK
Source: Goyal et al., 2021
AI systems can reason across a
combination of visual and textual
data. Visual reasoning skills are
essential in developing AI that can
reason more broadly. Existing AI can
already execute certain narrow visual
tasks better than humans, such as
classifying images, detecting faces,
and segmenting objects. But many AI
systems struggle when challenged to
reason more abstractly—for example, Figure 2.1.20
generating valid inferences about the

actions or motivations of agents in an
image (Figure 2.1.20).
SAMPLE QUESTIONS IN THE VISUAL QUESTION

Visual Question Answering
ANSWERING (VQA) CHALLENGE
(VQA) Challenge Source: Goyal et al., 2017
In the Visual Question Answering
challenge, AI systems are tasked with
answering open-ended questions
about images (Figure 2.1.21). To
answer the questions at a high level,
AI systems must have a combined
understanding of language, vision,
and commonsense reasoning.
Figure 2.1.21

In the six years since the VQA challenge began, there has systems could correctly answer only 55.4% of questions
been a 24.4 absolute percentage point improvement in (Figure 2.1.22). As of 2021, top performance stood at
state-of-the-art performance. In 2015, the top-performing 79.8%—close to the human baseline of 80.8%.
VISUAL QUESTION ANSWERING (VQA) CHALLENGE: ACCURACY

Source: VQA Challenge, 2021 | Chart: 2022 AI Index Report
80.80%, Human Baseline

80%
79.78%
70%
Accuracy (%)
60%
50%
2015 2016 2017 2018 2019 2020 2021
Figure 2.1.22

Index Report 2022 2.2 Computer Vision—Video
Video analysis concerns reasoning or task operation across sequential frames (videos), rather than single frames (images). Video
computer vision has a wide range of use cases, which include assisting criminal surveillance efforts, sports analytics, autonomous
driving, navigation of robots, and crowd monitoring.
2.2 COMPUTER VISION—VIDEO

AC T I V I T Y R E C O G N I T I O N Kinetics-400, Kinetics-600, Kinetics-700
fundamental subtask in video computer vision is
A Kinetics-400, Kinetics-600, and Kinetics-700 are a series of
activity recognition: identifying the activities that occur datasets for benchmarking video activity recognition. Each
in videos. AI systems have been challenged to classify dataset includes 650,000 large-scale, high-quality video
activities that range from simple actions, like walking, clips from YouTube that display a wide range of human
waving, or standing, to ones that are more complex and activities, and asks AI systems to classify an action from a
contain multiple steps, like preparing a salad (which possible set of 400, 600, and 700 categories, respectively.
requires an AI system to recognize and chain together Some of the new and more challenging activity classes
discrete actions like cutting tomatoes, washing the added to the Kinetics-700 series include pouring wine,
greens, applying dressing, etc.) playing the oboe, and making latte art.
EXAMPLE CLASSES FROM THE KINETICS DATASET

Source: Kay et al., 2017
Figure 2.2.1

As of 2022, one model tops all three Kinetics datasets. has narrowed between performance on the datasets. In
MTV, a collaboration between Google Research, Michigan 2020, the gap between performance on Kinetics-400 and
State University, and Brown University, released in January Kinetics-700 was 27.14 percentage points. In one short
2022, achieved a 89.6% Top-1 accuracy on the 600 series, year, that gap has narrowed to 7.4 points, which means
89.1% accuracy on the 400 series, and 82.20% accuracy that performance on the newer, harder dataset is occurring
on the 700 series (Figure 2.2.2). The most striking aspect more rapidly than performance on the easier dataset and
about technical progress on Kinetics is how rapidly the gap suggests that the easier ones are starting to asymptote.
KINETICS-400, KINETICS-600, KINETICS-700: TOP-1 ACCURACY

Source: Papers with Code, 2021; arXIv, 2021 | Chart: 2022 AI Index Report
90% 89.60%, Kinetics-600

89.10%, Kinetics-400
82.20%, Kinetics-700
80%
Top-1 Accuracy (%)
70%
60%
50%
2016 2017 2018 2019 2020 2021 2022
Figure 2.2.2

ActivityNet: Temporal Action Localization Task the most complex and difficult tasks in computer vision.
ActivityNet is a video dataset for human activity Performance on TALT is measured in terms of mean average
understanding that contains 700 hours of videos of precision, with a higher score indicating greater accuracy.
humans doing 200 different activities (long jump, dog
As of 2021 the top-performing model on TALT, developed
walking, vacuuming, etc.). For an AI system to successfully
by HUST-Alibaba, scores 44.7%, a 26.9 percentage point
complete the ActivityNet Temporal Action Localization
improvement over the top scores posted in 2016 when the
Task (TALT) task, it has to execute two separate steps: (1)
challenge began (Figure 2.2.3). Although state-of-the-art
localization (identify the precise interval during which the
results on the task have been posted for each subsequent
activity occurs); and (2) recognition (assign the correct
year, the gains have become increasingly small.
category label). Temporal action localization is one of
ACTIVITYNET, TEMPORAL ACTION LOCALIZATION TASK: MEAN AVERAGE PRECISION (mAP)

Source: ActivityNet, 2021 | Chart: 2022 AI Index Report
44.67%
40%
Mean Average Precision (mAP)
30%
20%
2016 2017 2018 2019 2020 2021

Figure 2.2.3

O B J E C T D E T E CT I O N take longer to process a video. This tradeoff between

Object detection is the task of identifying objects within speed and accuracy is also reflected in the types of object
an image (Figure 2.2.4). There are different philosophies detection methods pioneered in the last decade. There
bearing on priority, speed, and accuracy that guide the are one-stage methods which prioritize speed, such as
design of object detection systems. Systems that train SSD, RetinaNet, and YOLO, and two-stage methods which
quickly might be more efficient but are less accurate. prioritize accuracy, such as Mask R-CNN, Faster R-CNN,
Those that are more accurate might perform better but and Cascade R-CNN.
A DEMONSTRATION OF HOW OBJECT DETECTION APPEARS TO AI SYSTEMS

Source: COCO, 2020
Figure 2.2.4

Common Object in Context (COCO) Since 2016, there has been a 23.8 percentage point
Microsoft’s Common Object in Context (COCO) object improvement on COCO object detection, with this year’s
detection dataset contains over 328,000 images across more top model, GLIP, registering a mean average precision of
than 80 object categories. There are many accuracy metrics 79.5%.2 Figure 2.2.5 illustrates how the use of extra training
used to track performance on object detection, but for the data has taken over object detection, much as it has with
sake of consistency, this section and the majority of this other domains of computer vision.
report considers mean average precision (mAP50).
COCO-TEST-DEV: MEAN AVERAGE PRECISION (mAP50)

80%
75%

Mean Average Precision (mAP50)
70%
65%
60%
55%
50%
2015 2016 2017 2018 2019 2020 2021
Figure 2.2.5
2 GLIP (Grounded Language-Image Pretraining), a model designed to master the learning of language contextual visual representations, was a collaboration of researchers from UCLA, Microsoft Re-
search, University of Washington, University of Wisconsin–Madison, Microsoft Cloud, Microsoft AI, and International Digital Economy Academy.

You Only Look Once (YOLO) on the COCO dataset. YOLO detectors have become
You Only Look Once is an open-source object detection much better in terms of performance since 2017 (by 28.4
model that emphasizes speed (inference latency) over percentage points). Second, the gap in performance
absolute accuracy. between YOLO and the best-performing object detectors
has narrowed. In 2017 the gap stood at 11.7%, and it
Over the years, there have been different iterations of YOLO,
decreased to 7.1% in 2021. In the last five years, object
and Figure 2.2.6 plots the performance of YOLO object
detectors have been built that are both faster and better.
detectors versus the absolute top performing detectors
STATE OF THE ART (SOTA) vs. YOU ONLY LOOK ONCE (YOLO): MEAN AVERAGE PRECISION (mAP50)
Source: arXiv, 2021; GitHub, 2021 | Chart: 2022 AI Index Report
79.50%, SOTA
80%
Mean Average Precision (mAP50)
70%
72.40%, YOLO
60%
50%
40%
2016 2017 2018 2019 2020 2021
Figure 2.2.6

Visual Commonsense Reasoning (VCR) which only requires an answer). The dataset contains
The Visual Commonsense Reasoning challenge is a 290,000 pairs of multiple-choice questions, answers,
relatively new benchmark for visual understanding. VCR and rationales from 110,000 image scenarios taken from
asks AI systems to answer challenging questions about movies. Figure 2.2.7 illustrates the kinds of questions
scenarios presented from images, and also to provide the posed in the VCR.
reasoning behind their answers (unlike the VQA challenge,
A SAMPLE QUESTION OF THE VISUAL COMMONSENSE REASONING (VCR) CHALLENGE

Source: Zellers et al., 2018
Figure 2.2.7
Performance on VCR is measured in the Q->AR score, which lag far behind human levels of performance (Figure 2.2.8).
aggregates how well machines can choose the right answer At the end of 2021, the best mark on VCR stood at 72.0, a
for a given multiple-choice question (Q->A) and then select score that represents a 63.6% increase in performance since
the correct rationale for the answer (Q->R). 2018. Although progress has been made since the challenge
was launched, improvements have become increasingly
Since the challenge debuted, AI systems have become
marginal, suggesting that new techniques may need to be
better at visual commonsense reasoning, although they still
invented to significantly improve performance.
VISUAL COMMONSENSE REASONING (VCR) TASK: Q->AR SCORE
Source: VCR Leaderboard, 2021 | Chart: 2022 AI Index Report
90
85.00, Human Baseline
80
72.00
70
Q->AR Score
60
50
40
2018 2019 2020 2021
Figure 2.2.8

Index Report 2022 2.3 Language
Natural language processing (NLP) is a subfield of AI, with roots that stretch back as far as the 1950s. NLP involves research into systems that
can read, generate, and reason about natural language. NLP evolved from a set of systems that in its early years used handwritten rules and
statistical methodologies to one that now combines computational linguistics, rule-based modeling, statistical learning, and deep learning.
This section looks at progress in NLP across several language task domains, including: (1) English language understanding; (2) text
summarization; (3) natural language inference; (4) sentiment analysis; and (5) machine translation. In the last decade, technical progress in
NLP has been significant: The adoption of deep neural network–style machine learning methods has meant that many AI systems can now
execute complex language tasks better than many human baselines.
2.3 LANGUAGE
E N G L I S H L A N G UAG E As part of the benchmark, AI systems are tested on eight
different tasks (such as answering yes/no questions,
U N D E R S TA N D I N G
identifying causality in events, and doing commonsense
English language understanding challenges AI systems
reading comprehension), and their performance on these
to understand the English language in various contexts,
tasks is then averaged into a single score. SuperGLUE is
such as sentence understanding, yes/no reading
the successor to GLUE, an earlier benchmark that also
comprehension, reading comprehension with logical
tests on multiple tasks. SuperGLUE was released in May
reasoning, etc.
2019 after AI systems began to saturate the GLUE metric,
SuperGLUE creating demand for a harder benchmark.
SuperGLUE is a single-number metric that tracks technical
progress on a diverse set of linguistic tasks (Figure 2.3.1).
A SET OF SUPERGLUE TASKS3

Source: Wang et al., 2019
Figure 2.3.1
3 For the sake of brevity, this figure only displays 3 of the 8 tasks.

At the top of the SuperGLUE leaderboard sits the SS-MoE that progress on SuperGLUE was achieved so rapidly
model with a state-of-the-art score of 91.0 (Figure 2.3.2), suggests that researchers will need to develop more
which exceeds the human performance score of 89.8 complex suites of natural language tasks to challenge
given by the SuperGLUE benchmark developers. The fact the next generation of AI systems.
SUPERGLUE: SCORE
Source: SuperGLUE Leaderboard, 2021 | Chart: 2022 AI Index Report
92
91.00
91
Score
90
89.80, Human Performance
89
88
2019 2020 2021
Figure 2.3.2
Stanford Question Answering Dataset (SQuAD) HARDER QUESTIONS ADDED TO STANFORD

The Stanford Question Answering Dataset (SQuAD) QUESTION ANSWERING DATASET (SQUAD) 2.0
Source: Rajpurkar et al., 2018
benchmarks performance on reading comprehension.
The dataset includes 107,785 question-and-answer pairs
taken from 536 Wikipedia articles. Performance on SQuAD
is measured by the F1 score, which is the average overlap
between the AI system’s answers and the actual correct
answers: The higher the score, the better the performance.
As was the case with GLUE, AI systems improved so rapidly

on SQuAD that only two years after launching SQuAD in
2016, researchers released SQuAD 2.0. This second version
included more challenging reading comprehension tasks,
namely a set of 50,000 unanswerable questions that were
written in a way to appear answerable (Figure 2.3.3). Figure 2.3.3

At the end of 2021, the leading scores on SQuAD 1.1 and scores (0.4% and 0.2%). Both SQuAD datasets have seen
SQuAD 2.0 stood at 95.7 and 93.2, respectively (Figure a trend whereby immediately after the initial launches,
2.3.4). Although these scores are state of the art, they human-performance-exceeding scores were realized and
are marginal improvements over the previous year’s top then followed by small, plateau-like increases.
SQUAD 1.1 and SQUAD 2.0: F1 SCORE

Source: SQuAD 1.1 and SQuAD 2.0, 2021 | Chart: 2022 AI Index Report
100
95.72, Squad 1.1

95
93.21, Squad 2.0
91.20, Human Baseline (v1)

F1 Score
90 89.50, Human Baseline (v2)
85
80
2016 2017 2018 2019 2020 2021

Figure 2.3.4
Reading Comprehension A SAMPLE QUESTION IN READING COMPREHENSION DATASET

Dataset Requiring Logical REQUIRING LOGICAL REASONING (RECLOR)
Reasoning (ReClor) Source: Yu et al., 2020
The plateauing progress on

benchmarks like SQuAD suggests that
NLP models need to be tested on more
complex linguistic challenges like those
offered by ReClor. Created in 2020 by
computer scientists from the National
University of Singapore, ReClor requires
that AI systems engage in reading
comprehension that also necessitates
logical reasoning. The ReClor dataset is
Figure 2.3.5
made up of logical reasoning questions
from the LSAT, the entrance exam for
law schools in the United States and
Canada (Figure 2.3.5).

There are two sets of questions on ReClor, easy and

hard, with AI systems being judged on accuracy based Although AI systems
on the percentage of questions they answer correctly
(Figure 2.3.6). Although AI systems are presently capable
are presently capable of
of achieving a relatively high level of performance on achieving a relatively high
the easy set of questions, they struggle on the hard set.
In 2021, the top-performing model on ReClor (hard set) level of performance on the
scored 69.3%, roughly 22.5 percentage points lower easy set of questions, they
than the top-performing model on the easy set. Datasets
like ReClor suggest that while NLP models can execute struggle on the hard set.
straightforward reading comprehension tasks, they face
more difficulty when those tasks are coupled with logical
reasoning requirements.
READING COMPREHENSION DATASET REQUIRING LOGICAL REASONING (RECLOR): ACCURACY

Source: ReClor Leaderboard, 2021 | Chart: 2022 AI Index Report
91.82%, Test Easy

90%
80%
Accuracy (%)
70% 69.29%, Test Hard
60%
50%
2020 2021
Figure 2.3.6

T E X T S U M M A R I Z AT I O N a summary produced by an AI system and the reference

Text summarization is the challenge of synthesizing summary produced by a human. The higher the ROUGE
a piece of text while capturing its core content. score, the greater the overlap and the more accurate the
Summarizing texts is an important component of summary.
text classification, reading comprehension, and
information dissemination; however, when done
arXiv
manually by humans, it is time- and labor-intensive.
ArXiv is a text summarization benchmark dataset that
Developing AI systems that can functionally summarize
contains over 27,770 different papers from arXiv, the
texts has a number of practical use cases, from aiding
open-access repository of scientific papers. In the
universities in classifying academic papers to helping
five years since benchmarking on arXiv began, AI text
lawyers generate case summaries.
summarization models have improved their performance
Progress in text summarization is often scored on by 47.1% (Figure 2.3.7). However, as is the case with other
ROUGE (Recall-Oriented Understudy for Gisting natural language benchmarks, progress seems to be
Evaluation). ROUGE calculates the overlap between plateauing.
ARXIV: ROUGE-1
45
ROUGE-1
40
35
30
2017 2018 2019 2020 2021
Figure 2.3.7

Index Report 2022
PubMed on text classification since 2017 (34.6%), but recently the

PubMed is a text summarization dataset consisting of pace of that progress has slowed (Figure 2.3.8). In 2021, the
19,717 scientific publications from the PubMed database of top-performing model on PubMed was HAT (hierarchical
scientific papers. Progress on PubMed validates the trend attention transformer model), created by researchers at
seen on arXiv: There has been significant improvement Birch AI and the University of Washington.
PUBMED: ROUGE-1
50
45
ROUGE-1
40
35
2017 2018 2019 2020 2021
Figure 2.3.8

N AT U R A L L A N G UAG E I N F E R E N C E Stanford Natural Language Inference (SNLI)

Natural language inference is the task of determining The Stanford Natural Language Inference (SNLI) dataset
whether, given a premise, a hypothesis is true contains around 600,000 sentence pairs (premise
(entailment), false (contradiction), or undetermined and associated hypothesis) that are labeled as either
(neutral). This skill is also known as textual entailment, contradiction, or neutral. As part of this
entailment as it requires determining whether a challenge, AI systems are asked whether premises
particular premise logically entails a hypothesis. logically entail certain hypotheses (Figure 2.3.9).
Natural language inference necessitates language- Performance on SNLI is measured in accuracy based on
processing skills, such as named entity recognition the percentage of questions answered correctly.
(understanding the words you see), as well as being
able to use commonsense knowledge to distinguish
between reasonable and unreasonable inferences.
QUESTIONS AND LABELS IN STANFORD NATURAL LANGUAGE INFERENCE (SNLI)

Source: Bowman et al., 2015
Figure 2.3.9

The top-performing model on SNLI is Facebook AI USA’s EFL, which in April 2021 posted a score of 93.1% (Figure 2.3.10).
STANFORD NATURAL LANGUAGE INFERENCE (SNLI): ACCURACY

95%
93.10%
Accuracy (%)
90%
85%
2014 2015 2016 2017 2018 2019 2020 2021
Figure 2.3.10
Abductive Natural Language Inference (aNLI) abduction is regarded as an essential element in how
Abductive natural language inference is a more humans communicate with one another, few studies have
difficult type of textual entailment. Abductive inference attempted to study the abductive ability of AI systems.
requires drawing the most plausible conclusion from a
ANLI, a new benchmark for abductive natural language
context of limited information and uncertain premises.
inference created in 2019 by the Allen Institute for AI,
For instance, if Jenny were to return from work and
comes with 170,000 premise and hypothesis pairs. Figure
find her home in a disheveled state and then recall
2.3.11 illustrates the types of statements included in the
that she left a window open, she can plausibly infer
dataset.
that a burglar broke in and caused the mess.4 Although
EXAMPLE QUESTIONS IN ABDUCTIVE NATURAL LANGUAGE INFERENCE (ANLI)

Source: Allen Institute for AI, 2021
Figure 2.3.11
4 This particular example of abductive commonsense reasoning is taken from Bhagavatula et al. ( 2019), the first paper that investigates the ability of AI systems to perform language-based
abductive reasoning.

AI performance on abductive commonsense reasoning has increased by 7.7 percentage points since 2019; however, the
top AI systems, while close, are unable to achieve human performance levels (Figure 2.3.12). Abductive reasoning is
therefore still a challenging linguistic task for AI systems.
ABDUCTIVE NATURAL LANGUAGE INFERENCE (aNLI): ACCURACY

Source: Allen Institute for AI, 2021 | Chart: 2022 AI Index Report
92.90%, Human Baseline
92% 91.87%
90%
Accuracy (%)
88%
86%
84%
2019 2020 2021

Figure 2.3.12
S E N T I M E N T A N A LYS I S Sentiment analysis has many commercial use cases, from

Sentiment analysis is the task of using NLP techniques parsing customer reviews and field survey responses to
to identify the sentiment (very negative, negative, identifying the emotional states of customers.
neutral, positive, very positive) of a given text. Sentiment
SemEval 2014 Task 4 Sub Task 2
analysis can be straightforward if sentences are worded
The SemEval 2014 Task 4 Sub Task 2 is a benchmark
clearly and unambiguously, such as “I dislike winter
for sentiment analysis that asks machines to engage
weather.” However, sentiment analysis can become
in sentiment analysis. This specific task tests whether
more challenging when AI systems encounter sentences
AI systems can identify the sentiment associated with
with flipped structures or negations, such as “to say
particular aspects of a text, rather than the sentiment of
that disliking winter weather is not really my thing is
entire sentences or paragraphs (Figure 2.3.13).
completely inaccurate.”
A SAMPLE
SEMEVAL TASK
Source: Pontiki et al., 2014
Figure 2.3.13

The SemEval dataset is composed of 7,686 restaurant and laptop reviews, whose emotional polarities have been rated
by humans. On SemEval, AI systems are tasked with assigning the right sentiment labels to particular components of
the text, with their performance measured in terms of the percentage of the labels they correctly assign.
In the past seven years, AI systems have become much better at sentiment analysis. As of last year, top-performing
systems estimate sentiment correctly 9 out of 10 times, whereas in 2016, they made correct estimates only 7 out of
10 times. As of 2021, the state-of-the-art scores on SemEval stood at 88.6%, realized by a team of Chinese researchers
from South China Normal University and Linklogis Co. Ltd. (Figure 2.3.14).
SEMEVAL 2014 TASK 4 SUB TASK 2: ACCURACY

90%
88.64%
85%
Accuracy (%)
80%
75%
70%
2015 2016 2017 2018 2019 2020 2021
Figure 2.3.14
M AC H I N E T R A N S L AT I O N ( M T ) Since 2017, neural networks have taken over machine

Machine translation examines how AI software can translation. Unlike their predecessors, neural translators
translate languages. In domains where fluency in learn from a series of prior translation tasks and predict
multiple languages is required, machine translation the likelihood of a sequence of words. Neural translation
can be extremely impactful. The European Union, for models have revolutionized the field of machine translation
example, is required to translate all its cross-national not only because they do not require human supervision,
policy documents into the 24 languages of its member but also because they produce the most accurate
states. Using machine translators can save time, improve translations. As a result, they have been widely deployed by
efficiency, and lead to more consistent outcomes. search engines and social networks.

WMT 2014, English-German and English-French Both the English-French and English-German WMT 2014
The WMT 2014 family of datasets, first introduced at the benchmarks showcase the significant progress made
Meeting of the Association for Computational Linguistics in AI machine translation over the last decade (Figure
(ACL) 2014, consist of different kinds of translation 2.3.15). Since submissions began, there has been a 23.7%
tasks, including translation between English-French and improvement in English-French and a 68.1% improvement
English-German language pairs. A machine’s translation in English-German translation ability. Relatively speaking,
capabilities are measured by the Bilingual Evaluation although performance improvements have been more
Understudy, or BLEU, score, which compares the extent significant on the English-German language pair, absolute
to which a machine-translated text matches a reference translation ability remains meaningfully higher on English-
human-generated translation. The higher the score, the French translation.
better the translation.
WMT2014, ENGLISH-FRENCH: BLEU SCORE WMT2014, ENGLISH-GERMAN: BLEU SCORE

Source: Papers with Code, 2021; arXiv, 2021 | Chart: 2022 AI Index Report Source: Papers with Code, 2021; arXiv, 2021 | Chart: 2022 AI Index Report
45 45

40 40

BLEU Score
BLEU score
35 35
30 30
25 25
20 20
2015 2017 2019 2021 2015 2017 2019 2021

Figure 2.3.15

Number of Commercially Available MT Systems to Intento (Figure 2.3.16). 2021 also saw the introduction of
The growing interest in machine translation is also three open-source machine translation services (M2M-100,
reflected in the rise of commercial machine translation mBART, and OPUS). The emergence of publicly available,
services such as Google Translate. Since 2017, there high-functioning machine translation services speaks to
has been a nearly fivefold increase in the number of the increasing accessibility of such services and bodes well
commercial machine translators on the market, according for anybody who routinely relies on translation.
NUMBER of INDEPENDENT MACHINE TRANSLATION SERVICES

Source: Intento, 2021 | Chart: 2022 AI Index Report
50
Commercial 46
Number of Independent Machine Translation Services
Open Source Pre-trained

Preview
40
34
30
26 38
23
21
20 28
16
13 23
12 21
18
10
10 9
10 15
9
3
8 9
6 5
0 3 3 3 3
05/2017 07/2017 11/2017 03/2018 07/2018 12/2018 06/2019 11/2019 07/2020 10/2021
Figure 2.3.16

Index Report 2022 2.4 Speech
Another important domain of AI research is the analysis, recognition, and synthesis of human speech. In this AI subfield, AI systems are
typically rated on their ability to recognize speech and identify words and convert them into text; and also to recognize speakers and
identify the individuals speaking. Modern home assistance tools, such as Siri, are one of the many examples of commercially applied AI
speech technology.
2.4 SPEECH
SPEECH RECOGNITION English speech taken from a collection of audiobooks. On
Speech recognition is the process of training machines LibriSpeech, AI systems are asked to transcribe speech
to recognize spoken words and convert them into text. to text and then measured on word error rate, or the
Research in this domain began at Bell Labs in the 1950s, percentage of words they fail to correctly transcribe.
when the world was introduced to the automatic digit LibriSpeech is subdivided into two datasets. First, there
recognition machine (named “Audrey”), which could is LibriSpeech Test Clean, which includes higher-quality
recognize a human saying any number from zero to nine. recordings. Performance on Test Clean suggests how
Speech recognition has come a long way since then and in well AI systems can transcribe speech in ideal conditions.
the last decade has benefited tremendously from deep- Second, there is LibriSpeech Test Other, which includes
learning techniques and the availability of rich speech lower-quality recordings. Performance on Test Other is
recognition datasets. indicative of transcription performance in environments
Transcribe Speech: LibriSpeech (Test-Clean and where sound quality is less than ideal.
Other Datasets) AI systems perform incredibly well on LibriSpeech, so
Introduced in 2015, LibriSpeech is a speech transcription much so that progress appears to be plateauing (Figure
database that contains around 1,000 hours of 16 khz 2.4.1). A state-of-the-art result on the Test Clean dataset
LIBRISPEECH, TEST CLEAN: WORD ERROR RATE (WER) LIBRISPEECH, TEST OTHER: WORD ERROR RATE (WER)
Source: Papers with Code, 2021; arXiv, 2021 | Chart: 2022 AI Index Report Source: Papers with Code, 2021; arXiv, 2021 | Chart: 2022 AI Index Report
14 14
12 12
10 10
Word Error Rate (WER)
Word Error Rate (WER)
8 8 3.30, Without Extra Training Data
6 6
4 4
2 2 2.50, With Extra Training Data

0 0
2015 2017 2019 2021 2015 2017 2019 2021
Figure 2.4.1

Index Report 2022 2.4 Speech
was not realized in 2021, which speaks to the fact that the VoxCeleb
error rate of the top system was already low at 1.4%. For VoxCeleb is a large-scale audiovisual dataset of human
every 100 words that top-performing transcription models speech for speaker recognition, which is the task of matching
heard, they correctly transcribed 99. certain speech with a particular individual. Each year, the
makers of VoxCeleb host a speaker verification challenge.
Performance on the Test Other dataset, while worse than
A low score or equal error rate on the VoxCeleb challenge
Test Clean, was still relatively poor. The state-of-the-art
is indicative of an AI system that makes few errors in its
results on Test Other were realized by the W2V-BERT
attribution of speech to particular individuals.5 Figure 2.4.2
model, an MIT and Google Brain collaboration, which
plots performance over time on VoxCeleb-1, the original
posted an error rate of 2.0%.
VoxCeleb dataset. Since 2017, performance on VoxCeleb has
improved: Systems that once reported equal error rates of
7.8% now report errors that are less than 1.0%.
VOXCELEB: EQUAL ERROR RATE (EER)

Source: VoxCeleb, 2021 | Chart: 2022 AI Index Report
8%
6%
Equal Error Rate (%)
4%
2%
0.42%
0%
2017 2018 2019 2020 2021
Figure 2.4.2
5 The Equal Error Rate (EER) is not only a measure of the false positive rate, assigning a bad label, but also the false negative rate (failure to assign the correct label).

Index Report 2022 2.5 Recommendation
Recommendation is the task of suggesting items that might be of interest to a user, such as movies to watch, articles to read, or products
to purchase. Recommendation systems are crucial to businesses, such as Amazon, Netflix, Spotify, and YouTube. For example, one of
the earliest open recommendation competitions in AI was the Netflix Prize; hosted in 2009, it challenged computer scientists to develop
algorithms that could accurately predict user ratings for films based on previously submitted ratings.
2.5 RECOMMENDATION
Commercial Recommendation: MovieLens 20M quality. A higher nDCG score means that an AI system
The MovieLens 20M dataset contains around 20 delivers more accurate recommendations.
million movie ratings for 27,000 movies from 138,000 Since 2018, top models now perform roughly 5.2% better
users. The ratings are taken from MovieLens (a movie on MovieLens 20M than they did in 2018 (Figure 2.5.1).
recommendation platform), and AI systems are challenged In 2021, the state-of-the-art system on MovieLens 20M
to see if they can predict a user’s movie preferences based posted an nDCG of 0.448 and came from researchers at
on their previously submitted ratings. The metric used to the Czech Technical University in Prague.
track performance on MovieLens is Normalized Discounted
Cumulative Gain (nDCG), which is a measure of ranking
MOVIELENS 20M: NORMALIZED DISCOUNTED CUMULATIVE GAIN@100 (nDCG@100)

0.460
Normalized Discounted Cumulative Gain@100 (nDCG@100)
0.450
0.448
0.440
0.430
0.420
2018 2019 2020 2021
Figure 2.5.1

Index Report 2022 2.5 Recommendation
Click-Through Rate Prediction: Criteo Performance on Criteo also indicates that recommender
Click-through rate prediction is the task of predicting systems have been slowly and steadily improving in the
the likelihood that something on a website, say an past decade. Last year’s top model (Sina Weibo Corp’s
advertisement, will be clicked. In 2014, the online MaskNet) performed 1.8% higher on Criteo than the top
advertising platform Criteo launched an open click- model from 2016. An improvement of 1.8% may seem
through prediction challenge. Included as part of this small in absolute terms, but it can be a valuable margin in
challenge dataset was information on a million ads the commercial world.
that were displayed during a 24-day period, whether A limit of the Criteo and MovieLens benchmarks is that
they were clicked, and additional information on their they are primarily academic measures of technical
characteristics. Since the competition launched, the progress in recommendation (Figure 2.5.2). Most of the
Criteo dataset has been widely used to test recommender research work on recommendation occurs in commercial
systems. On Criteo, systems are measured on area under settings. Given that companies have an incentive to keep
the curve (AUC). A higher AUC means a better click- their recommendation improvements proprietary, the
through prediction rate and a stronger recommender academic metrics included in this section might not be
system. complete measures of the technical progress made in
recommendation.
CRITEO: AREA UNDER CURVE SCORE (AUC)

0.813
0.810
Area Under Curve Score (AUC)
0.800
0.790
2016 2017 2018 2019 2020 2021
Figure 2.5.2

Index Report 2022 2.6 Reinforcement Learning
In reinforcement learning, AI systems are trained to maximize performance on a given task by interactively learning from their prior
actions. Researchers train systems to optimize by rewarding them if they achieve a desired goal and then punishing them if they fail.
Systems experiment with different strategy sequences to solve their stated problem (e.g., playing chess or navigating through a maze)
and select the strategies which maximize their rewards.
Reinforcement learning makes the news whenever programs like DeepMind’s AlphaZero demonstrate superhuman performance on
games like Go and Chess. However, reinforcement learning is useful in any commercial domain where computer agents need to maximize
a target goal or stand to benefit from learning from previous experiences. Reinforcement learning can help autonomous vehicles change
lanes, robots optimize manufacturing tasks, or time-series models predict future events.
2.6 REINFORCEMENT LEARNING

REINFORCEMENT LEARNING
ENVIRONMENTS
Creating reinforcement
A reinforcement learning environment is a computer learning models that are
platform where AI agents are challenged to maximize
their performance on a defined task. Unlike other AI tasks both high performing
which require systems to train on a dataset, reinforcement and highly efficient is
learning necessitates that AI systems have an environment
in which they can test various strategies and, in the process, an important step in the
identify the set of strategies that will maximize rewards.
commercial deployment of
Arcade Learning Environment: Atari-57
reinforcement learning.
Introduced in 2013, the Arcade Learning Environment
is an interface that includes various Atari 2600 game
environments (such as “Pac-Man,” “Space Invaders,”
and “Frogger”) in which AI agents are challenged to
In 2021, however, researchers from Tsinghua University
optimize performance. To enable standard comparisons,
and ByteDance launched the GDI-H3 model, which
researchers typically report average performance on the
surpassed (and nearly doubled) MuZero’s performance
ALE across a suite of 57 games. There are various metrics
on Atari-57 (Figure 2.6.1). Moreover, GDI-H3 achieved
in which performance is measured, but one of the most
this performance with less training. It only used 200
common is mean human-normalized score. A human-
million training frames, whereas MuZero used 20 billion:
normalized score of 0% represents random performance,
GDI-H3 was twice as effective and one hundred times
and a 100% score represents average human performance.
more efficient. Creating reinforcement learning models
The mean human-normalized score is then the average
that are both high performing and highly efficient is
human-normalized score achieved by an AI system.
an important step in the commercial deployment of
In late 2019, DeepMind’s MuZero algorithm achieved reinforcement learning.
state-of-the-art performance on Atari-57. MuZero
not only performed 48.3% better on Atari-57 than
the previous best-performing model, but it also set a
new world record on Go and achieved superhuman
performance on chess and shogi.

ATARI-57: MEAN HUMAN-NORMALIZED SCORE

Source: Papers with Code, 2021; arXIv, 2021 | Chart: 2022 AI Index Report
10 9.62
Mean-Human Normalized Score (in thousands)
0
2015 2016 2017 2018 2019 2020 2021
Figure 2.6.1
Procgen
Procgen is a new reinforcement learning A SCREENSHOT OF THE 16 GAME ENVIRONMENTS IN PROCGEN
Source: Cobbe et al. 2019
environment introduced by OpenAI in 2019. It
includes 16 procedurally generated video-game-
like environments specifically designed to test the
ability of reinforcement learning agents to learn
generalizable skills (Figure 2.6.2). Procgen was
developed to overcome some of the criticisms
leveled at benchmarks like Atari that incentivized
AI systems to become narrow learners that
maximized capacity in one particular skill.
Procgen encourages broad learning by introducing
a reinforcement learning environment that
emphasizes high diversity and forces AI systems
to train in generalizable ways. Performance on
Procgen is measured in terms of mean-normalized
score. Researchers typically train their systems on
200 million training runs and report an average
score across the 16 Procgen games. The higher the
system scores, the better the system.
Figure 2.6.2

In November 2021, the MuZero model from DeepMind the environment was first released. Rapid progress on
posted a state-of-the-art score of 0.6 on Procgen. such a diverse benchmark signals that AI systems are
DeepMind’s results were a 128.6% improvement over slowly improving their capacity to reason in broader
the baseline performance established in 2019 when environments.
PROCGEN: MEAN-NORMALIZED SCORE

0.64
0.60
Mean-Normalized Score
0.50
0.40
0.30
2019 2020 2021

Figure 2.6.3

Human Games: Chess range of games, such as shogi and Go, and have in
Progress in reinforcement learning can also be captured fact beaten some of the top-ranked chess engines.
by the performance of the world’s top chess software Nevertheless, looking at chess engine performance is an
engines. A chess engine is a computer program that is effective way to relativize the progress made in AI and
trained to play chess at a high level by analyzing chess compare it to a widely understandable human baseline.
positions. The performance of chess engines is ranked
Computers surpassed human performance in chess
on Elo, a method for identifying the relative skill levels
a long time ago, and since then have not stopped
of players in zero-sum games like chess: A higher score
improving (Figure 2.6.4). By the mid-1990s, the top chess
means a stronger player.
engines exceeded expert-level human performance, and
One caveat is that tracking the performance of by the mid-2000s they surpassed the peak performance
chess engines is not a complete reflection of general of Magnus Carlsen, one of the best chess players of all
reinforcement learning progress; chess engines are time. Magnus Carlsen’s 2882 Elo, recorded in 2014, is
specifically trained to maximize performance on chess. the highest level of human chess performance ever
Other popular reinforcement learning systems, like documented. As of 2021, the top chess engine has
DeepMind’s AlphaZero, are capable of playing a broader exceeded that level by 24.3%.
CHESS SOFTWARE ENGINES: ELO SCORE

Source: Swedish Computer Chess Association, 2021 | Chart: 2022 AI Index Report
3,581
3500
3000 2882, Magnus Carlsen
2500
2300, Expert
Elo Score
2000
1700, Intermediate
1500
1000
800, Novice
500
0
1987 1992 1997 2002 2007 2012 2017 2022
Figure 2.6.4

Index Report 2022 2.7 Hardware
In evaluating technical progress in AI, it is relevant not only to consider improvements in technical performance but also the speed of
operation. As this section shows, AI systems continue to improve in virtually every skill category. This performance is often realized by
increasing parameters and training systems on greater amounts of data. However, all else being equal, models that use more parameters
and source more data will take longer to train. Longer train times mean slower real-world deployment. Given that the potential of
increased training times can be offset by stronger and more robust computational infrastructures, it is important to keep track of
progress in the hardware that powers AI systems.
2.7 HARDWARE
MLPerf: Training Time increased. Top-performing hardware systems can reach
MLPerf is an AI training competition run by the ML baseline levels of performance in task categories like
Commons organization. In this challenge, participants recommendation, light-weight objection detection, image
train systems to execute various AI tasks (image classification, and language processing in under a minute.
classification, image segmentation, natural language
Figure 2.7.2 more precisely profiles the magnitude of
processing, etc.) using a common architecture. Entrants
improvement across each skill category since MLPerf first
are then ranked on their absolute wall clock time, which is
introduced the category.6 For example, training times
how long it takes for the system to train.
on image classification increased roughly twenty-seven-
Since the MLPerf competitions began in December 2018, fold, as top times dropped from 6.2 minutes in 2018 to
two key trends have emerged: (1) Training times for 0.2 minutes (or 13.8 seconds) in 2021. It might be hard to
virtually every AI skill category have massively decreased; fathom the magnitude of a 27-time decrease in training
while (2) AI hardware robustness has substantially time, but in real terms it is the difference between waiting
one hour for a bus versus a little more than two minutes.
MLPERF TRAINING TIME of TOP SYSTEMS by TASK: MINUTES

Source: MLPerf, 2021 | Chart: 2022 AI Index Report
50
20
Training Time (Minutes; Log Scale)
10 13.57, Reinforcement Learning
5
3.24, Object Detection (heavy-weight)
2.38, Speech recognition
2
1.26, Image Segmentation

1
0.63, Recommendation
0.5
0.34, Object Detection (light-weight)
0.23, Language Processing
0.2 0.23, Image Classi cation
2018 2019 2020 2021

Figure 2.7.1
6 The solitary point for reinforcement learning on Figure 2.7.1 indicates that a faster time was not registered in the MLPerf competitions in 2020 or 2021. The solitary points for speech recognition and
image segmentation are indicative of the fact that those AI subtask categories were added to the MLPerf competition in 2021.

Top-performing hardware systems can reach baseline levels

of performance in task categories like recommendation,
light-weight objection detection, image classification, and
language processing in under a minute.
MLPERF: SCALE of IMPROVEMENT across TASK

26.96
25
22.25
20
Scale of Improvement
16.47
15
10
5
2.38
1.70 1.92
Improvement Baseline 1.16
0
Reinforcement Speech Language Recommendation Segmentation Object Detection Object Detection Image
Recognition Processing (light-weight) (heavy-weight) Classi!cation
Figure 2.7.2

MLPerf: Number of Accelerators accelerators used by all entrants increased 3.5 times.
The cross-task improvements in training time are being Most notable, however, is the growing gap between
driven by stronger underlying hardware systems, as shown the average number of accelerators used by the top-
in Figure 2.7.3. Since the competition began, the highest performing systems and the average accelerators used
number of accelerators used—where an accelerator is by all systems. This gap was 9 times larger at the end of
a chip used predominantly for the machine learning 2021 than it had been in 2018. This growth clearly means
component of a training run, such as a GPU or a TPU—and that, on average, building the fastest systems requires
the mean number of accelerators used by the top system the most powerful hardware.
increased roughly 7 times while the mean number of
MLPERF HARDWARE: ACCELERATORS

4,320, Maximum Number of Accelerators Used
4,000
3,000
Number of Accelerators
2,000 1,785, Average Accelerators Used by Top System
1,000
337, Mean Number of Accelerators
0
12.12.2018 06.10.2019 07.29.2020 06.30.2021 12.01.2021
Figure 2.7.3

IMAGENET: Training Cost classification system. This cost is marginal, especially

Stronger hardware has not necessarily meant costlier when compared to the $1,112.6 it cost to train a similarly
training. Figure 2.7.4 plots the lowest training cost per performing system in 2017. In simpler terms, in four
year on MLPerf’s image classification subtask (ImageNet). short years, image classification training costs have
In 2021, it cost only $4.6 to train a high-performing image decreased by a factor of 223.
IMAGENET: TRAINING COST (to 93% ACCURACY)

Source: AI Index and Narayanan, 2021 | Chart: 2022 AI Index Report
$1,000.00
$500.00
$200.00
Cost (in U.S. Dollars; Log Scale)
$100.00
$50.00
$20.00
$10.00
$5.00 $4.59
$2.00
$1.00
2017 2018 2019 2020 2021
Figure 2.7.4

Index Report 2022 2.8 Robotics
In 2021, the AI Index developed a survey that asked professors who specialize in robotics at top-ranked universities around the world and
in emerging economies about changes in the pricing of robotic arms as well as the uses of robotic arms in research labs. The survey was
completed by 101 professors and researchers from over 40 universities and collected data on 117 robotic arm purchase events from 2016
to 2022. The survey results suggest that there has been a notable decline in the price of robotic arms since 2016.
2.8 ROBOTICS
Price Trends in Robotic Arms7 $22,600 in 2021 (Figure 2.8.1). Figure 2.8.2, which plots the
The survey results show a clear downward trend in the distribution of robotic arm prices, paints a similar picture:
pricing of robotic arms in the last seven years. In 2017, Despite some high-priced outliers, there has been a clear
the median price of a robotic arm was $42,000. Since downward trend since 2017 in the price of robotic arms.
then, the price has declined by 46.2% to reach roughly
MEDIAN PRICE of ROBOTIC ARMS, 2017–21

Source: AI Index, 2022 | Chart: 2022 AI Index Report
$40
Price (in thousands of U.S. Dollars)
$30
$22.60
$20
$10
$0
2017 2018 2019 2020 2021
Figure 2.8.1
7 We have corrected Figure 2.8.1 and Figure 2.8.2 after noticing a data filtering issue with the survey result. The correct chart has since been updated. View the appendix here for links to the data. In
addition, note that academic researchers may get a discount when purchasing robotic arms so prices are lower than retail.

Index Report 2022 2.8 Robotics
DISTRIBUTION of ROBOTIC ARM PRICES, 2017–21

$100
Price (in thousands of U.S. Dollars)
$50
$0
2017 2018 2019 2020 2021
Figure 2.8.2
AI Skills Employed by Robotics AI SKILLS EMPLOYED by ROBOTICS PROFESSORS

Professors Source: AI Index, 2022 | Chart: 2022 AI Index Report
In the survey, the AI Index also asked

robotics professors to what extent
they employ AI skills in their research.
Responses revealed that both deep Deep Learning 67.00%
learning and reinforcement learning are

popular AI skills employed by roboticists.
More specifically, 67.0% of professors
reported using deep learning and 46.0%
reported using reinforcement learning.
Reinforcement
46.00%
Learning
0% 20% 40% 60% 80%

% of Respondents
Figure 2.8.3

Index Report 2022
CHAPTER 3:
Technical AI Ethics
Text and Analysis by
Helen Ngo and Ellie Sakhaee
Index Report 2022 CHAPTER 3: TECHNICAL AI ETHICS
CHAPTER 3:
Chapter Preview
Overview 102 3.3 AI ETHICS TRENDS AT
Acknowledgment 103 FACCT AND NEURIPS 123
Chapter Highlights 105 ACM Conference on Fairness,

Accountability, and Transparency (FAccT) 123
3.1 META-ANALYSIS OF FAIRNESS NeurIPS Workshops 125

AND BIAS METRICS 106 Interpretability, Explainability,
AI Ethics Diagnostic Metrics and Causal Reasoning 126
and Benchmarks 107 Privacy and Data Collection 127
Fairness and Bias 129
3.2 NATURAL LANGUAGE
PROCESSING BIAS METRICS 109 3.4 FACTUALITY AND TRUTHFULNESS 130
Toxicity: RealToxicityPrompts Fact-Checking With AI 130
and the Perspective API 109
Measuring Fact-Checking Accuracy
Highlight: Large Language Models With FEVER Benchmark 133
and Toxicity 111
Toward Truthful Language Models 134
Detoxification of Models Can
Model Size and Truthfulness 134
Negatively Influence Performance 113
Highlight: Multimodal Biases in Contrastive
StereoSet 114
Language-Image Pretraining (CLIP) 136
CrowS-Pairs 115
Denigration Harm 136
Winogender and WinoBias 117
Gender Bias 136
WinoMT: Gender Bias in Machine
Propagating Learned Bias Downstream 138
Translation Systems 119
Underperformance on
Word and Image Embedding
Non-English Languages 138
Association Tests 120
Highlight: Multilingual Word Embeddings 122
Mitigating Bias in Word Embeddings ACCESS THE PUBLIC DATA
With Intrinsic Bias Metrics 122

Overview
In recent years, AI systems have started to be deployed into the
world, and researchers and practitioners are reckoning with their
real-world harms. Some of these harms include commercial facial
recognition systems that discriminate based on race, résumé
screening systems that discriminate on gender, and AI-powered
clinical health tools that are biased along socioeconomic and racial
lines. These models have been found to reflect and amplify human
social biases, discriminate based on protected attributes, and
generate false information about the world. These findings have
increased interest within the academic community in studying AI
ethics, fairness, and bias and prompted industry practitioners to
direct resources toward remediating these issues, and attracted
attention from the media, governments, and the people who use and
are affected by these systems.
This year, the AI Index highlights metrics which have been adopted
by the community for reporting progress in eliminating bias and
promoting fairness. Tracking performance on these metrics alongside
technical capabilities provides a more comprehensive perspective
on how fairness and bias change as systems improve, which will be
important to understand as systems are increasingly deployed.

Index Report 2022
AC K N OW L E D G M E N T
The AI Index would like to thank all those involved in research and advocacy around the development and governance
of responsible AI. This chapter builds upon the work of scholars from across the AI ethics community, including those
working on measuring technical capabilities as well those focused on shaping thoughtful societal norms. There is much
more work to be done, but we are inspired by the progress made by this community and its collaborators.
Publications cited in this Chapter include:
Sandhini Agarwal, Gretchen Krueger, Jack Clark, Alec Radford, Jong Wook Kim, and Miles Brundage. 2021. Evaluating
CLIP: Towards Characterization of Broader Capabilities and Downstream Implications. arXiv preprint arXiv:2108.02818.
Jack Bandy and Nicholas Vincent. 2021. Addressing “Documentation Debt” in Machine Learning Research: A
Retrospective Datasheet for Book Corpus. arXiv preprint arXiv:2105.05241.
Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. 2021. Multimodal Datasets: Misogyny, Pornography, and
Malignant Stereotypes. arXiv preprint arXiv:2110.01963.
Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den
Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman
Ring, Tom Hennigan, Saffron Huang, Loren Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey
Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack W. Rae, Erich Elsen, and Laurent Sifre. 2021. Improving
Language Models by Retrieving from Trillions of Tokens. arXiv preprint arXiv:2112.04426.
Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. 2017. Word Embeddings Quantify 100 Years of Gender and
Ethnic Stereotypes. arXiv preprint arXiv:1711.08412.
Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A.Smith. 2020. RealToxicityPrompts: Evaluating
Neural Toxic Degeneration in Language Models. arXiv preprint arXiv:2009.11462.
Wei Guo and Aylin Caliskan. 2020. Detecting Emergent Intersectional Biases: Contextualized Word Embeddings Contain a
Distribution of Human-like Biases. arXiv preprint arXiv:2006.03955.
Aylin Caliskan Islam, Joanna J. Bryson, and Arvind Narayanan. 2016. Semantics Derived Automatically from Language
Corpora Necessarily Contain Human Biases. arXiv preprint arXiv:1608.07187.
Opher Lieber, Or Sharir, Barak Lenz, and Yoav Shoham. 2021. Jurassic-1: Technical Details and Evaluation. (2021).
https://uploads-ssl.webflow.com/60fd4503684b466578c0d307/61138924626a6981ee09caf6_jurassic_tech_paper.pdf
Chandler May, Alex Wang, Shikha Bordia, Samuel R. Bowman, and Rachel Rudinger. 2019. On Measuring Social Biases in
Sentence Encoders. arXiv preprint arXiv:1903.10561.
Moin Nadeem, Anna Bethke, and Siva Reddy. 2020. StereoSet: Measuring Stereotypical Bias in Pretrained Language
Models. arXiv preprint arXiv:2004.09456.
Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain,
Vineet Kosaraju, William Saunders, Xu Jiang, Karl Cobbe, Tyna Eloundou, Gretchen Krueger, Kevin Button, Matthew
Knight, Benjamin Chess, John Schulman. WebGPT: Browser-Assisted Question-Answering with Human Feedback. 2021.
arXiv preprint arXiv:2112.09332.

Index Report 2022
Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R. Bowman. 2020. CrowS-Pairs: A Challenge Dataset for
Measuring Social Biases in Masked Language Models. arXiv preprint arXiv:2010.00133.
Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal,
Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter
Welinder, Paul Christiano, Jan Leike, Ryan Lowe. Training Language Models to Follow Instructions with Human Feedback.
2022. arXiv preprint arXiv:2203.02155.
Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, H. Francis Song, John Aslanides, Sarah
Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan, Jacob Menick, Albin Cassirer, Richard
Powell, George van den Driessche, Lisa Anne Hendricks, Maribeth Rauh, Po-Sen Huang, Amelia Glaese, Johannes Welbl,
Sumanth Dathathri, Saffron Huang, Jonathan Uesato, John Mellor, Irina Higgins, Antonia Creswell, Nat McAleese, Amy
Wu, Erich Elsen, Siddhant M. Jayakumar, Elena Buchatskaya, David Budden, Esme Sutherland, Karen Simonyan, Michela
Paganini, Laurent Sifre, Lena Martens, Xiang Lorraine Li, Adhiguna Kuncoro, Aida Nematzadeh, Elena Gribovskaya,
Domenic Donato, Angeliki Lazaridou, Arthur Mensch, Jean-Baptiste Lespiau, Maria Tsimpoukelli, Nikolai Grigorev, Doug
Fritz, Thibault Sottiaux, Mantas Pajarskas, Toby Pohlen, Zhitao Gong, Daniel Toyama, Cyprien de Masson d’Autume,
Yujia Li, Tayfun Terzi, Vladimir Mikulik, Igor Babuschkin, Aidan Clark, Diego de Las Casas, Aurelia Guy, Chris Jones, James
Bradbury, Matthew Johnson, Blake A. Hechtman, Laura Weidinger, Iason Gabriel, William S. Isaac, Edward Lockhart,
Simon Osindero, Laura Rimell, Chris Dyer, Oriol Vinyals, Kareem Ayoub, Jeff Stanway, Lorrayne Bennett, Demis Hassabis,
Koray Kavukcuoglu, and Geoffrey Irving. 2021. Scaling Language Models: Methods, Analysis & Insights from Training
Gopher. arXiv preprint arXiv:2112.11446.
Gabriel Stanovsky, Noah A. Smith, and Luke Zettlemoyer. 2019. Evaluating Gender Bias in Machine Translation. arXiv
preprint arXiv:1906.00591.
Ryan Steed and Aylin Caliskan. 2020. Image Representations Learned With Unsupervised Pre-Training Contain Human-
like Biases. arXiv preprint arXiv:2010.15052.
Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia Glaese,
Borja Balle, Atoosa Kasirzadeh, Zac Kenton, Sasha Brown, Will Hawkins, Tom Stepleton, Courtney Biles, Abeba Birhane,
Julia Haas, Laura Rimell, Lisa Anne Hendricks, William S. Isaac, Sean Legassick, Geoffrey Irving, and Iason Gabriel. 2021.
Ethical and social risks of harm from Language Models. arXiv preprint arXiv:2112.04359.
Johannes Welbl, Amelia Glaese, Jonathan Uesato, Sumanth Dathathri, John Mellor, Lisa Anne Hendricks, Kirsty
Anderson, Pushmeet Kohli, Ben Coppin, and Po-Sen Huang. 2021. Challenges in Detoxifying Language Models. arXiv
preprint arXiv:2109.07445.
Albert Xu, Eshaan Pathak, Eric Wallace, Suchin Gururangan, Maarten Sap, and Dan Klein. 2021. Detoxifying Language
Models Risks Marginalizing Minority Voices. arXiv preprint arXiv:2104.06390.
Pei Zhou, Weijia Shi, Jieyu Zhao, Kuan-Hao Huang, Muhao Chen, Ryan Cotterell, and Kai-Wei Chang. 2019. Examining
Gender Bias in Languages with Grammatical Gender. arXiv preprint arXiv:1909.02224.

CHAPTER HIGHLIGHTS
• Language models are more capable than ever, but also more biased: Large language
models are setting new records on technical benchmarks, but new data shows that larger
models are also more capable of reflecting biases from their training data. A 280 billion
parameter model developed in 2021 shows a 29% increase in elicited toxicity over a
117 million parameter model considered the state of the art as of 2018. The systems are
growing significantly more capable over time, though as they increase in capabilities, so does
the potential severity of their biases.
• The rise of AI ethics everywhere: Research on fairness and transparency in AI has exploded
since 2014, with a fivefold increase in related publications at ethics-related conferences.
Algorithmic fairness and bias has shifted from being primarily an academic pursuit to
becoming firmly embedded as a mainstream research topic with wide-ranging implications.
Researchers with industry affiliations contributed 71% more publications year over year
at ethics-focused conferences in recent years.
• Multimodal models learn multimodal biases: Rapid progress has been made on training
multimodal language-vision models which exhibit new levels of capability on joint language-
vision tasks. These models have set new records on tasks like image classification and the
creation of images from text descriptions, but they also reflect societal stereotypes and
biases in their outputs—experiments on CLIP showed that images of Black people were
misclassified as nonhuman at over twice the rate of any other race. While there has been
significant work to develop metrics for measuring bias within both computer vision and natural
language processing, this highlights the need for metrics that provide insight into biases in
models with multiple modalities.

Artificial Intelligence CHAPTER 3: TECHNICAL AI ETHICS
Index Report 2022 3.1 Meta-Analysis of Fairness and Bias Metrics
Significant research effort has been invested over the past five years into creating datasets, benchmarks, and metrics designed to
measure bias and fairness in machine learning models. Bias is often learned from the underlying training data for an AI model; this data
can reflect systemic biases in society or the biases of the humans who collected and curated the data.
3.1 META-ANALYSIS OF FAIRNESS AND

BIAS METRICS
Algorithmic bias is commonly framed in terms of to legal or ethical reasons (e.g., race, gender, religion).
allocative and representation harms. Allocative harm
To better understand the landscape of algorithmic bias
occurs when a system unfairly allocates an opportunity
and fairness, the AI Index conducted original research to
or resource to a specific group, and representation harm
analyze the state of the field. As shown in Figure 3.1.1,
happens when a system perpetuates stereotypes and
the number of metrics for measuring bias and fairness
power dynamics in a way that reinforces subordination
along ethical dimensions of interest has grown steadily
of a group. Algorithms are broadly considered fair when
since 2018. For this graph, the number of fairness and
they make predictions that neither favor nor discriminate
bias metrics published has been cited in at least one
against individuals or groups based on protected
other work.1
attributes which cannot be used for decision-making due
NUMBER of AI FAIRNESS and BIAS METRICS, 2016–21

15
Number of Metrics
10
0
2016 2017 2018 2019 2020 2021
Figure 3.1.1
1 2021 data may be lagging as it takes time for metrics to be adopted by the community.

A I E T H I C S D I AG N O S T I C M E T R I C S Diagnostic metrics enable researchers and practitioners

A N D B E N C H M A R KS to understand the impact of their system on a specific
Measurement of AI systems along an ethical dimension application or group and potential concrete harm (e.g.,
often takes one of two forms: “this model is disproportionately underperforming on this
group with this protected attribute”). Diagnostic metrics
• Benchmark datasets: A benchmark dataset contains are most useful at an individual model or application level
labeled data, and researchers test how well their AI as opposed to functioning as field-level indicators. They
system labels the data. Benchmarks do not change over indicate how a specific AI system performs on a specific
time. These are domain-specific (e.g., SuperGLUE and subgroup or individual, which is helpful for assessing
StereoSet for language models; ImageNet for computer real-world impact. However, while these metrics may be
vision) and often aim to measure behavior that is intrinsic widely used to test models privately, there is not as much
to the model, as opposed to its downstream performance information available publicly as these metrics are not
on specific populations (e.g., StereoSet measures model attached to leaderboards which encourage researchers to
propensity to select stereotypes compared to non- publish their results.
stereotypes, but it does not measure performance gaps
between different subgroups). Figure 3.1.2 shows that there has been a steady amount
of research investment into developing both benchmarks
• Diagnostic metrics: A diagnostic metric measures the and diagnostic metrics over time.2 3
impact or performance of a model on a downstream
task—for example, a population subgroup or individual
compared to similar individuals or the entire population. Benchmarks are useful
These metrics can help researchers understand how a
system will perform when deployed in the real world, and indicators of progress
whether it has a disparate impact on certain populations. for the field as a whole,
Examples include group fairness metrics such as
demographic parity and equality of opportunity. and their impact can be
Benchmarks are useful indicators of progress for the measured by community
field as a whole, and their impact can be measured
by community adoption (e.g., number of leaderboard
adoption (e.g., number of
submissions, or the number of research papers which leaderboard submissions,
report metrics). They also often enable rapid algorithmic
progress as research labs compete on leaderboard
or the number of research
metrics. However, some leaderboards can be easily papers which report
gamed, and may be based on benchmark datasets that
contain flaws, such as incorrect labels or poorly defined metrics).
classes. Additionally, their static nature means they are a
snapshot of a specific cultural and temporal context—in
other words, a benchmark published in 2017 may not
correlate to the deployment context of 2022.
2 Research paper citations are a lagging indicator of activity, and metrics which have been very recently adopted may not be reflected in the current data, similar to 3.1.1.
3 The Perspective API defined seven new metrics for measuring facets of toxicity (toxicity, severe toxicity, identity attack, insult, obscene, sexually explicit, threat), contributing to the unusually high
number of metrics released in 2017.

NUMBER of AI FAIRNESS and BIAS METRICS (DIAGNOSTIC METRICS vs. BENCHMARKS), 2016–21
15
Benchmarks 13
Diagnostic Metrics
10
Number of Metrics
9 9 9
5
4
3 3
2 2 2
0 0
0 1 2
2016 3 40 1 2
2017 3 40 1 2
2018 3 40 1 2
2019 3 40 1 2
2020 3 40 1 2
2021 3 4
Figure 3.1.2
The rest of this chapter examines the performance of intrinsic bias within systems, and it has been shown that
recent AI systems on these metrics and benchmarks in intrinsic bias metrics may not fully capture the effects of
depth within domains such as natural language and extrinsic bias within downstream applications.
computer vision. The majority of these metrics measure

Index Report 2022 3.2 Natural Language Processing Bias Metrics
Current state-of-the-art natural language processing (NLP) relies on large language models or machine learning systems that process
millions of lines of text and learn to predict words in a sentence. These models can generate coherent text; classify people, places, and
events; and be used as components of larger systems, like search engines. Collecting training data for these models often requires scraping
the internet to create web-scale text datasets. These models learn human biases from their pretraining data and reflect them in their
downstream outputs, potentially causing harm. Several benchmarks and metrics have been developed to identify bias in natural language
processing along axes of gender, race, occupation, disability, religion, age, physical appearance, sexual orientation, and ethnicity.
3.2 NATURAL LANGUAGE PROCESSING

BIAS METRICS
Bias metrics can be split into two major categories: intrinsic disrespectful or unreasonable comment that is likely to
metrics, which measure bias in internal embedding spaces make someone leave a conversation. The Perspective
of models, and extrinsic metrics, which measure bias in API is a tool developed by Jigsaw, a Google company. It
the downstream tasks and outputs of the model. Examples was originally designed to help platforms identify toxicity
of extrinsic metrics include group fairness metrics (parity in online conversations. Developers input text into the
across protected groups) and individual fairness metrics Perspective API, which returns probabilities that the text
(parity across similar individuals), which measure whether should be labeled as falling into one of the following
a system has a disproportionately negative impact on a categories: toxicity, severe toxicity, identity attack, insult,
subgroup or individual, or gives preferential treatment to obscene, sexually explicit, and threat.
one group at the expense of another.
Since the Perspective API was released in 2017, the NLP
TOX I C I T Y: REA LTOXICIT Y PROM PTS research community has rapidly adopted it for measuring
AN D T H E PERS PECTIVE A PI toxicity in natural language. As seen in Figure 3.2.1, the
Measuring toxicity in language models requires labels for number of papers using the Perspective API doubled
toxic and nontoxic content. Toxicity is defined as a rude, between 2020 and 2021, from 8 to 19.
NUMBER of RESEARCH PAPERS USING PERSPECTIVE API

20
18
15
Number of Research Papers
10
0
2018 2019 2020 2021
Figure 3.2.1

RealToxicityPrompts consists of English natural language Figure 3.2.2 shows that toxicity in language models
prompts used to measure how often a language model depends heavily on the underlying training data. Models
completes a prompt with toxic text. Toxicity of a language trained on internet text with toxic content filtered out are
model is measured with two metrics: significantly less toxic compared to models trained on
various corpora of unfiltered internet text. A model trained
• M
aximum toxicity: the average maximum toxicity score
on BookCorpus (a dataset containing books from e-book
across some number of completions
websites) produces toxic text surprisingly often. This
• P
robability of toxicity: how often a completion is may be due to its composition—BookCorpus contains a
expected to be toxic significant number of romance novels containing explicit
content, which may contribute to higher levels of toxicity.
TOXICITY in LANGUAGE MODELS by TRAINING DATASET

Source: Gehman et al., 2021; Rae et al., 2021; Welbl et al., 2021 | Chart: 2022 AI Index Report
Expected maximum toxicity Toxicity probability
0.90
0.88
0.85
0.82
0.80 0.78
0.73 0.75
0.71
Toxicity Score
0.60
0.40
0.25
0.20
0.08
0.00
openwebtext (C4, toxicity openwebtext (CTRL, openwebtext (GPT-3,
BookCorpus Wikipedia
!ltered) un!ltered) un!ltered)
Training Data
Figure 3.2.2

Large Language Models and Toxicity

Recent developments around mitigating toxicity prompted with inputs of varying levels of toxicity,
in language models have lowered both expected but that they are also more capable of detecting
maximum toxicity and the probability of toxicity. toxicity with regard to their own outputs as well as
However, detoxification methods consistently lead in other contexts, as measured by increased AUC
to adverse side effects and somewhat less capable (area under the receiver operating characteristic
models. (For example, filtering training data curve) with model size. The AUC metric plots the
typically comes at the cost of model performance.) true positive rate against the false positive rate
to characterize how well a model distinguishes
In December 2021, DeepMind released a paper
between classes (higher is better). Larger models
describing its 280 billion parameter language
are dramatically better at identifying toxic
model, Gopher. Figure 3.2.3a and Figure 3.2.3b
comments within the CivilComments dataset, as
from the Gopher paper show that larger models
shown in Figure 3.2.3b.
are more likely to produce toxic outputs when
GOPHER: PROBABILITY of TOXIC CONTINUATIONS BASED on PROMPT TOXICITY by MODEL SIZE

Source: Rae et al., 2021 | Chart: 2022 AI Index Report
0.4
High
0.3
Med
Toxicity Probability
0.2
Low
0.1 Very Low

Unprompted
0.0
117 417 1,400 7,100 280,000
Model Size (Number of Parameters in millions)
Figure 3.2.3a

Large Language Models and Toxicity (cont’d)

GOPHER: FEW-SHOT TOXICITY CLASSIFICATION on the CIVILCOMMENTS DATASET
10-shot
0.7
Classi!cation Area Under the Curve (AUC)
0.6
0.5
0.4
0.3
0.2
0.1
0.0
44 117 417 1,400 7,100 280,000
Model Size (Number of Parameters in millions)
Figure 3.2.3b

D E TOX I F I CAT I O N O F M O D E L S A study on detoxifying language models shows that

CA N N E G AT I V E LY I N F LU E N C E models detoxified with these strategies all perform worse
on both white-aligned and African American English on
PERFORMANCE
perplexity, a metric that measures how well a model has
Detoxification methods aim to mitigate toxicity by
learned a specific distribution (lower is better) (Figure
changing the underlying training data as in domain-
3.2.4). These models also perform disproportionately
adaptive pretraining (DAPT), or by steering the model
worse on African American English and text containing
during generation as in Plug and Play Language Models
mentions of minority identities compared to white-
(PPLM) or Generative Discriminator Guided Sequence
aligned text, a result that is likely due to human biases
Generation (GeDi).
causing annotators to be more apt to mislabel African
American English as toxic.
PERPLEXITY: LANGUAGE MODELING PERFORMANCE by MINORITY GROUPS on ENGLISH

POST-DETOXIFICATION
Source: Xu et al., 2021 | Chart: 2022 AI Index Report
Baseline (GPT-2)
Detoxi!cation method - DAPT
Detoxi!cation method - GeDi
500 Detoxi!cation method - PPLM
400
Perplexity
300
200
100
0
African American–aligned English Minority identity mentions
White-aligned English (nontoxic) White-aligned English (toxic)
(nontoxic) (nontoxic)
Text
Figure 3.2.4

STEREOSET Figure 3.2.5 shows that StereoSet performance follows

StereoSet is a benchmark measuring stereotype bias along the same trend seen with toxicity: Larger models reflect
the axes of gender, race, religion, and profession, along stereotypes more often unless interventions are taken
with raw-language modeling ability. One of the associated to reduce learned stereotypes during training. The
metrics is stereotype score, which measures whether a prevalence of toxic content online has been estimated
model prefers stereotypes and anti-stereotypes equally. A to be 0.1–3%, which aligns with research showing that
stereotype is an over-generalized belief widely held about larger language models are more capable of memorizing
a group and an anti-stereotype is a generalization about a rare text.
group which contradicts widely accepted stereotypes.
STEREOSET: STEREOTYPE SCORE by MODEL SIZE

Source: Nadeem et al., 2020; Lieber et al., 2021; StereoSet Leaderboard, 2021; | Chart: 2022 AI Index Report
64
62.00
62
Stereotype Score
60.70
60.10
60
59.30
58.30 58.20
58
56
110 340 345 1,558 175,000 178,000

Number of Parameters (in millions)
Figure 3.2.5
StereoSet has several major flaws in its underlying dataset: these stereotypes were sourced from crowdworkers
Some examples fail to express a harmful stereotype, located in the United States, and the resulting values and
conflate stereotypes about countries with stereotypes stereotypes within the dataset may not be universally
about race and ethnicity, and confuse stereotypes representative.
between associated but distinct groups. Additionally,

C R OW S - PA I R S Like earlier examples, BERT, RoBERTa, and ALBERT

CrowS-Pairs (Crowdsourced Stereotype Pairs) is another appear to inherit biases from their training data. They
benchmark measuring stereotype bias. While StereoSet were all trained on a combination of BookCorpus, English
compares attributes about a single group, CrowS-Pairs Wikipedia, and text scraped from the internet. Analysis
contrasts relationships between historically disadvantaged of BookCorpus reveals that its books about religion are
and advantaged groups (e.g., Mexicans versus white heavily skewed toward Christianity and Islam compared
people). to other major world religions,5 though it is unclear the
extent to which these books contain historical content
The creators of CrowS-Pairs measured stereotype bias versus content written from a specific religious viewpoint.6
using three popular language models: BERT, RoBERTa,
and ALBERT (Figure 3.2.6). On standard language We can examine how language models may inherit biases
modeling benchmarks, ALBERT outperforms RoBERTa, about certain religions by looking at their underlying
which outperforms BERT.4 However, ALBERT is the most datasets. Figure 3.2.7 shows the number of books
biased of the three models according to CrowS-Pairs. pertaining to different religions in two popular datasets,
This mirrors the trend observed with StereoSet and BookCorpus and Smashwords21. Both datasets have
RealToxicityPrompts: More capable models are also more far more mentions of Christianity and Islam than other
capable of learning and amplifying stereotypes. religions.
CROWS-PAIRS: LANGUAGE MODEL PERFORMANCE across BIAS ATTRIBUTES

Source: Nangia et al., 2020 | Chart: 2022 AI Index Report
BERT RoBERTa ALBERT
80
60
Stereotype Score
40
20
0
Gender Physical Sexual Socioeconomic
Age Disability Nationality Race/Color Religion
Identity Appearance Orientation Status
Bias Attribute
Figure 3.2.6
4 Per results on the SQuAD, GLUE, and RACE benchmarks.

5 Such as Sikhism, Judaism, Hinduism, Buddhism, Atheism.
6 Hate speech classifiers fine-tuned on top of BERT in particular have been shown to frequently misclassify texts containing mentions of “Muslim” as toxic, and researchers find that GPT-3 contains
significant bias along religious axes for mentions of both “Jewish” and “Muslim.”

BOOKCORPUS and SMASHWORDS21: SHARE of BOOKS about RELIGION in PRETRAINING DATASETS

Source: Bandy and Vincent, 2021 | Chart: 2022 AI Index Report
BookCorpus
50% smashwords21
Books (as % of All Religious Books in Corpus)
40%
30%
20%
10%
0%
Atheism Buddhism Christianity Hinduism Islam Judaism Sikhism
Religion
Figure 3.2.7

WINOGENDER AND WINOBIAS cases, along with the gender parity score (the percentage
Winogender measures gender bias related to occupations. of examples for which the predictions are the same).
Systems are measured on their ability to fill in the correct The authors use crowdsourced annotations to estimate
gender in a sentence containing an occupation (e.g., human performance to be 99.7% accuracy.
“The teenager confided in the therapist because he / she Winogender results from the SuperGLUE leaderboard
seemed trustworthy”). Examples were created by sourcing show that larger models are more capable of correctly
data from the U.S. Bureau of Labor Statistics to identify resolving gender in the zero-shot and few-shot setting
occupations skewed toward one gender (e.g., the cashier (i.e., without fine-tuning on the Winogender task) and less
occupation is made up of 73% women, but drivers are only likely to magnify occupational gender disparities (Figure
6% women). 3.2.8). However, a good score on Winogender does not
Performance on Winogender is measured by the accuracy indicate that a model is unbiased with regard to gender,
gap between the stereotypical and anti-stereotypical only that bias was not captured by this benchmark.
MODEL PERFORMANCE on the WINOGENDER TASK from the SUPERGLUE BENCHMARK

Source: SuperGLUE Leaderboard, 2021 | Chart: 2022 AI Index Report
0.72 0.71
0.64
0.59 0.59 0.60

0.60
0.55
0.53
0.50
Coreference Accuracy
0.40
0.20
0.00
31 100 223 340 10,000 13,000 64,000 175,000 280,000
Number of Parameters (in millions)
Figure 3.2.8

WinoBias is a similar benchmark measuring gender bias Winogender, but the adoption of Winogender within the
related to occupations that was released concurrently SuperGLUE leaderboard for measuring natural language
with Winogender by a different research group. As understanding has led to more model evaluations being
shown in Figure 3.2.9, WinoBias is cited more often than reported on Winogender.
WINOBIAS and WINOGENDER: NUMBER of CITATIONS, 2018–21

Source: AI Index, 2021; Semantic Scholar, 2021 | Chart: 2022 AI Index Report
128, WinoBias
100 95, Winogender

Number of Citations
50
0
2018 2019 2020 2021
Figure 3.2.9

W I N O M T: G E N D E R B I A S I N accuracy), the difference in F1 score between masculine

M AC H I N E T R A N S L AT I O N SYS T E M S and feminine examples, and the difference in F1 score
Commercial machine translation systems have been between examples with stereotypical gender roles and
documented to reflect and amplify societal biases anti-stereotypical gender roles.
from their underlying datasets. As these systems are As seen in Figure 3.2.10, Google Translate has been
used broadly in global industries such as e-commerce, shown to perform better across all tested languages
stereotypes and mistakes in translation can be costly. (Arabic, English, French, German, Hebrew, Italian,
WinoMT is a benchmark measuring gender bias in Russian, Ukrainian) when translating examples containing
machine translation that is created by combining the occupations that conform to societal biases about gender
Winogender and WinoBias datasets. Models are evaluated roles. Additionally, these systems translate sentences
by comparing the sentences translated from English to with the correct gender only up to 60% of the time. Other
another language and extracting the translated gender to major commercial machine translation systems (Microsoft
compare with the original gender. Systems are scored on Translator, Amazon Translate, SYSTRAN) have been shown
the percentage of translations with correct gender (gender to behave similarly.
WINOMT: GENDER BIAS in GOOGLE TRANSLATE across LANGUAGES

Source: Stanovsky et al., 2019 | Chart: 2022 AI Index Report
Gender Accuracy F1 Delta Masculine Feminine F1 Delta Stereo Anti-stereo
Arabic Arabic Arabic
English English English
French French French
German German German
Hebrew Hebrew Hebrew
Italian Italian Italian
Russian Russian Russian
Ukrainian Ukrainian Ukrainian
0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60
Figure 3.2.10

WO R D A N D I M AG E E M B E D D I N G associating male terms with “career” and female

A S S O C I AT I O N T E S T S terms with “family”). CEAT (Contextualized Embedding
Word embedding is a technique in NLP that allows words Association Test) extends WEAT to contextualized word
with similar meanings to have similar representations. embeddings.
Static word embeddings are fixed representations which The Image Embedding Association Test (iEAT) modifies
do not change with context. For example, polysemous WEAT to measure associations between social concepts
words will have the same representation (embedding) and image attributes. Using iEAT, researchers showed that
regardless of the sentence in which they appear. Examples pretrained generative vision models (iGPT and simCLRv2)
of static word embeddings include GloVe, PPMI, FastText, exhibit humanlike biases with regard to gender, race, age,
CBoW, and Dict2vec. In contrast, contextualized word and disability.
embeddings are dynamic representations of words that
change based on the word’s accompanying context. For Word embeddings can be aggregated into sentence
example, “bank” would have different representations in embeddings with models known as sentence encoders.
“riverbank” and “bank teller.” The Sentence Encoder Association Test (SEAT) extends
WEAT to measure bias in sentence encoders related
The Word Embedding Association Test (WEAT) quantifies to gendered names, regional names, and stereotypes.
bias in English static word embeddings by measuring Newer transformer-based language models which use
the association (“effect size”) between concepts (e.g., contextualized word embeddings appear to be less biased
European-American and African American names) and than their predecessors, but most models still show
attributes (e.g., pleasantness and unpleasantness). significant bias with regard to gender and occupations,
Word embeddings trained on large public corpora as well as African American names versus European-
(e.g., Wikipedia, Google News) consistently replicate American names, as shown in Figure 3.2.11.
stereotypical biases when evaluated on WEAT (e.g.,
SENTENCE EMBEDDING ASSOCIATION TEST (SEAT): MEASURING STEREOTYPICAL ASSOCIATIONS with EFFECT
SIZE
Source: May et al., 2019 | Chart: 2022 AI Index Report
Angry Black Woman Stereotype
European American/African American Names
Male/Female Names, Career
1.5
1.0
E"ect Size
0.5
0.0
3/2018 (Universal
10/2014 (CBoW) 9/2017 (InferSent) 2/2018 (ELMo) 3/2018 (GenSen) 6/2018 (GPT) 10/2018 (BERT)
Sentence Encoder)
Figure 3.2.11

Word embeddings also reflect cultural shifts: A temporal measured with the relative norm difference: the average
analysis of word embeddings over 100 years of U.S. Census Euclidean distance between words associated with
text data shows that changes in embeddings closely track representative groups (e.g., men, women, Asians) and
demographic and occupational shifts over time. Figure words associated with occupations. The blue line shows
3.2.12 shows that shifts in embeddings trained on the gender bias over time, where negative values indicate
Google Books and Corpus of Historical American English that embeddings more closely associate occupations with
(COHA) corpora reflect significant historical events like the men. The red line shows the bias of embeddings relating
women’s movement in the 1960s and Asian immigration race to occupations, specifically in the case of Asian
to the United States. In this analysis, embedding bias is Americans and whites.
GENDER and RACIAL BIAS in WORD EMBEDDINGS TRAINED on 100 YEARS of TEXT DATA
Source: Garg et al., 2018 | Chart: 2022 AI Index Report
-0.02
Racial Bias
-0.03
Gender Bias
Average Bias Score
-0.04
-0.05
-0.06
1910 1920 1930 1940 1950 1960 1970 1980 1990

Figure 3.2.12

Multilingual Word Embeddings

Large language models are often monolingual Another study on gender bias extends WEAT
since they require a significant amount of text data to quantify biases in bilingual embeddings in
to train. While English text can be easily sourced languages with grammatical gender, such as
by scraping the internet, the challenge is greater Spanish or French. Figure 3.2.13 shows that
with low-resource languages like Fula. XWEAT is a masculine words in Spanish are closer to the
multilingual and cross-lingual extension of WEAT English words for historically male-dominated
that is designed for comparative bias analyses occupations (e.g., architect) as well as the
between languages. Results on XWEAT show that neutral position, as indicated by the vertical line.
bias in cross-lingual embeddings can roughly be Similarly, feminine occupation words are closer to
predicted from the biases in the corresponding English words for historically female-dominated
monolingual embedding, indicating that biases can occupations (e.g., nurse).
be transferred between languages.
GENDER BIAS in SPANISH WORD EMBEDDINGS: EMBEDDING SIMILARITY DISTANCE

Source: Zhou et al., 2019 | Chart: 2022 AI Index Report
architect
arquitecta arquitecto
lawyer
abogada abogado
profesora professor profesor
doctora doctor doctor
enfermera
nurse enfermero
-0.3 -0.2 -0.1 0.0 0.1

Female <— Gender Direction —> Male
Figure 3.2.13
Mitigating Bias in Word Embeddings With demonstrated that there is no reliable correlation
Intrinsic Bias Metrics between intrinsic bias metrics and downstream
It is often assumed that reducing intrinsic bias by de- application biases. Further investigation is needed to
biasing embeddings will reduce downstream biases establish meaningful relationships between intrinsic and
in applications (extrinsic bias). However, it has been extrinsic metrics.

Index Report 2022 3.3 AI Ethics Trends at FAccT and NeurIPS
To grasp how the field of AI ethics has evolved over time, this section studies trends from the ACM Conference on Fairness, Accountability,
and Transparency (FAccT), which publishes work on algorithmic fairness and bias, and from NeurIPS workshops. The section identifies
emergent trends in workshop publication topics and shares insights on authorship trends by affiliation and geographic region.
3.3 AI ETHICS TRENDS AT FACCT

AND NEURIPS
AC M CO NFERENCE O N FA IRNESS, Figure 3.3.1 shows that industry labs are making up a
ACCO UNTA B ILIT Y, A ND larger share of publications at FAccT year over year. They
often produce work in collaboration with academia but
T RANS PA R ENCY (FACCT)
are increasingly producing standalone work as well. In
ACM FAccT is an interdisciplinary conference publishing
2021, 53 authors listed an industry affiliation, up from
research in algorithmic fairness, accountability, and
31 authors in 2020 and only 5 authors at the inaugural
transparency.7 While several AI conferences offer
conference in 2018. This aligns with recent findings that
workshops dedicated to similar topics, FAccT was one
point to a trend of deep learning researchers transitioning
of the first major conferences created to bring together
from academia to industry labs.
researchers, practitioners, and policymakers interested in
sociotechnical analysis of algorithms.
NUMBER of ACCEPTED FACCT CONFERENCE SUBMISSIONS by AFFILIATION, 2018–21

Source: FAccT, 2021; AI Index, 2021 | Chart: 2022 AI Index Report
Education 302
300 Industry
Government
Nonpro!t
Other 244
Number of Papers
200
227
166
200
100 139
71
53
63
31
21
0 17
2018 2019 2020 2021
Figure 3.3.1
7 Work accepted by FAccT includes technical frameworks for measuring fairness, investigations into the harms of AI in specific industries (e.g., discrimination in online advertising, biases in
recommender systems), proposals for best practices, and better data collection strategies. Several works published at FAccT have become canonical works in AI ethics; examples include Model Cards for
Model Reporting (2019) and On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? (2021). Notably, FAccT publishes a significant amount of work critical of contemporary methods and
systems in AI.

While there has been increased interest in fairness, followed by researchers based in Europe and Central Asia
accountability, and transparency research from all types (Figure 3.3.2). From 2020 to 2021, the proportion of papers
of organizations, the majority of papers published at FAccT from institutions based in North America increased from
are written by researchers based in the United States, 70.2% to 75.4%.
NUMBER of ACCEPTED FACCT CONFERENCE SUBMISSIONS by REGION, 2018–21

Source: FAccT, 2021; AI Index, 2021 | Chart: 2022 AI Index Report

Number of Publications (% of World Total)
60%
40%
20%
0.40%, Sub-Saharan Africa
0.00%, South Asia
0%
2018 2019 2020 2021

Figure 3.3.2

N E U R I P S WO R KS H O P S at NeurIPS ethics-related workshops in the past six years

NeurIPS, one of the largest AI conferences, held its first by research topic, indicating an increased interest in
workshop on fairness, accountability, and transparency in AI applied to high-risk, high-impact use cases such as
2014. Figure 3.3.3 shows the number of research papers climate, finance, and healthcare.
NEURIPS WORKSHOP RESEARCH TOPICS: NUMBER of ACCEPTED PAPERS on REAL-WORLD IMPACTS, 2015–21
Source: NeurIPS, 2021; AI Index, 2021 | Chart: 2022 AI Index Report
250 Climate
243
Developing World
Finance
19
Healthcare
Other
200 29
191
172
Number of Papers
150 79
86
114
15 172 13
100
81
75 13
17
17
50 99 97
75
62
47
0 16
2015 2017 2018 2019 2020 2021
Figure 3.3.3

Interpretability, Explainability, and measure fairness by changing protected attributes of an

Causal Reasoning individual input (e.g., race, gender) and observing how the
Several workshops have been created at NeurIPS around model outputs a different prediction—for example, a bank
interpretability and explainability, including safety- can change the “age” feature in a model to understand if
critical AI affecting human decisions,8 interpretability its model performs fairly on customers over 60 years old.
and causality for algorithmic fairness,9 and the necessity Counterfactual fairness formalizes the idea that a model
of explainability for high-risk use cases.10 Interpretability makes fair decisions with regard to an individual if the
and explainability work focus on designing systems that decision would be the same if the individual belonged to a
are inherently interpretable and providing explanations different demographic.
for the behavior of a black-box system, while the study of
Since 2018, an increasing number of papers on causal
causal inference aims to understand cause and effect by
inference have been published at NeurIPS. In 2021, there
uncovering associations between variables that depend
were three workshops at NeurIPS dedicated to causal
on each other and asking what would have happened if a
inference, including one devoted entirely to causality and
different decision had been made—that is, if this had not
algorithmic fairness (Figure 3.3.4). Figure 3.3.5 shows that
occurred, then that would not have happened.
there has been a similar increase in research papers in
Counterfactual analysis can be used to gain insight into interpretability and explainability work at NeurIPS over
a black-box system by changing an input feature and time, especially in the NeurIPS main track.
observing how the output changes. This can be applied to
NEURIPS RESEARCH TOPICS: NUMBER of ACCEPTED PAPERS on CAUSAL EFFECT and COUNTERFACTUAL
REASONING, 2015–2021
Main Track
80 78
Workshop 76
72
20
60
Number of Papers
43 53
40 39
16
58
20
29
9 23 23
6
4 9
0 6
2015 2016 2017 2018 2019 2020 2021
Figure 3.3.4
8 See 2017 Transparent and Interpretable Machine Learning in Safety Critical Environments, 2019 Workshop on Human-Centric Machine Learning: Safety and Robustness in Decision-Making, 2019,
“‘Do the Right Thing’: Machine Learning and Causal Inference for Improved Decision-Making.”
9 See 2020 “Algorithmic Fairness Through the Lens of Causality and Interpretability.”
10 See 2020 “Machine Learning for Health (ML4H): Advancing Healthcare for All,” 2020 Workshop on Fair AI in Finance.

Privacy and Data Collection domains (e.g., financial services), federated learning for
Amid growing concerns about privacy, data sovereignty, decentralized model training, and differential privacy
and the commodification of personal data for profit, there to ensure that training data does not leak personally
has been significant momentum in industry and academia identifiable information.11 This section shows the number
to build methods and frameworks to help mitigate of papers submitted to NeurIPS mentioning “privacy” in
privacy concerns. Since 2018, several workshops have the title along with papers accepted to privacy-themed
been devoted to privacy in machine learning, covering NeurIPS workshops, and finds a significant increase in the
topics such as privacy in machine learning within specific number of accepted papers since 2016 (Figure 3.3.6).
NEURIPS RESEARCH TOPICS: NUMBER of ACCEPTED PAPERS on INTERPRETABILITY and EXPLAINABILITY, 2015–21
41
Main Track
40 Workshop
18
30
Number of Papers
23
20
17
12 7 19
23
10 5
6 6
10
2 6 7 6
4
0
2015 2016 2017 2018 2019 2020 2021
Figure 3.3.5
11 See “Privacy Preserving Machine Learning,” Workshop on Federated Learning for Data Privacy and Confidentiality, Privacy in Machine Learning (PriML).

NEURIPS RESEARCH TOPICS: NUMBER of ACCEPTED PAPERS on PRIVACY in AI, 2015–21

Main Track 150

150 Workshop
12
128
15
Number of Papers
100
88
79 13
138
113
50
72 75
21
16
19
1 14
0
2015 2016 2017 2018 2019 2020 2021
Figure 3.3.6

Fairness and Bias the number of papers accepted to the conference main
In 2020, NeurIPS started requiring authors to submit track that mention fairness or bias in the title, along
broader impact statements addressing the ethical and with papers accepted to a fairness-related workshop.
potential societal consequences of their work, a move Figure 3.3.7 shows a sharp increase from 2017 onward,
that suggests the community is signaling the importance demonstrating the newfound importance of these topics
of AI ethics early in the research process. One measure of within the research community.
the interest in fairness and bias at NeurIPS over time is
NEURIPS RESEARCH TOPICS: NUMBER of ACCEPTED PAPERS on FAIRNESS and BIAS in AI, 2015–21
168
Main Track 149

150
Workshop
50
125 36
16 114
Number of Papers
100
36
113 118
109
50
34 78
10
24
2 4
0
2015 2016 2017 2018 2019 2020 2021
Figure 3.3.7

Index Report 2022 3.4 Factuality and Truthfulness
This section analyzes trends in using AI to verify the factual accuracy of claims, as well as research related to measuring the truthfulness
of AI systems. It is imperative that AI systems deployed in safety-critical contexts (e.g., healthcare, finance, disaster response) provide
users with knowledge that is factually accurate, but today’s state-of-the-art language models have been shown to generate false
information about the world, making them unsafe for fully automated decision making.
3.4 FACTUALITY AND TRUTHFULNESS

FACT- C H EC KING W ITH A I venture capital firm invested $1 million in an automated
In recent years, social media platforms have deployed fact-checking competition for fake news.
AI systems to help manage the proliferation of online The research community has developed several
misinformation. These systems may aid human fact- benchmarks for evaluating automatic fact-checking
checkers by identifying potential false claims for them to systems, where verifying the factuality of a claim is posed
review, surfacing previously fact-checked similar claims, or as a classification or scoring problem (e.g., with two
surfacing evidence that supports a claim. Fully automated classes classifying whether the claim is true or false).
fact-checking is an active area of research: In 2017, the Figure 3.4.1 shows that most datasets binarize labels into
Fake News Challenge encouraged researchers to build true or false categories, while some datasets have many
AI systems for stance detection, and in 2019, a Canadian categories for claims.
DATASETS for AUTOMATED FACT-CHECKING: GRANULARITY of LABELS

2 75
3 28
4 16
5 15
Number of Classes
6 6
7 3
13 1
17 1
18 1
40 1
Numeric 1
Ranking 3
0 10 20 30 40 50 60 70 80
Number of Datasets
Figure 3.4.1

The increased interest in automated fact-checking levels of factuality. Similarly, Truth of Varying Shades is a
is evidenced by the number of citations of relevant multiclass political fact-checking and fake news detection
benchmarks: FEVER is a fact extraction and verification benchmark. Figure 3.4.2 shows that these three English
dataset made up of claims classified as supported, refuted, benchmarks have been cited with increasing frequency in
or not enough information. LIAR is a dataset for fake news recent years.
detection with six fine-grained labels denoting varying
AUTOMATED FACT-CHECKING BENCHMARKS: NUMBER of CITATIONS, 2017–21

LIAR
200
FEVER
150
Number of Citations
Truth of Varying Shades

100
50
2017 2018 2019 2020 2021

Figure 3.4.2

Figure 3.4.3 shows the number of fact-checking datasets English datasets (including 14 in Arabic, 5 in Chinese, 3
created for English compared to all other languages in Spanish, 3 in Hindi, and 2 in Danish) compared to 142
over time. As seen in Figure 3.4.4, there are only 35 non- English-only datasets.12
NUMBER of AUTOMATED FACT-CHECKING BENCHMARKS for ENGLISH, 2010–21

50
48
English
Non-English
40
Number of Datasets
30
38
27
25
20
20 19
13
23
12 15
10 18
6 9
12
10
2 6 5
1 1 3 4
0 0 0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 3.4.3
NUMBER of AUTOMATED FACT-CHECKING BENCHMARKS by LANGUAGE

English 142
Arabic 14
Chinese 5
Hindi 3
Spanish 3
Danish 2
French 2
German 2
Portugese 2
Bengali 1
Bulgarian 1
Croatian 1
Italian 1
Malayalam 1
Tamil 1
0 20 40 60 80 100 120 140
Number of Benchmarks
Figure 3.4.4
12 Modern language models are trained on disproportionately larger amounts of English text, which negatively impacts performance on other languages. The Gopher family of models is trained on MassiveText
(10.5 TB), which is 99% English. Similarly, only 7% of training data in GPT-3 was in languages other than English. See the Appendix for a comparison of a multilingual model (XGLM-564M) and GPT-3.

Measuring Fact-Checking Accuracy at least one set of supporting evidence was correctly
With FEVER Benchmark identified. Several variations of this dataset have
FEVER (Fact Extraction and VERification) is a benchmark since been introduced (e.g., FEVER 2.0, FEVEROUS,
measuring the accuracy of fact-checking systems, where FoolMeTwice).
the task requires systems to verify the factuality of a
Figure 3.4.5 shows that state-of-the-art performance has
claim with supporting evidence extracted from English
steadily increased over time on both accuracy and FEVER
Wikipedia. Systems are measured on classification
score. Some contemporary language models only report
accuracy and FEVER score, a custom metric which
accuracy, as in the case of Gopher.
measures whether the claim was correctly classified and
FACT EXTRACTION and VERIFICATION (FEVER) BENCHMARK: ACCURACY and FEVER SCORE, 2018–21
80 79.35, Label Accuracy
76.78, FEVER Score
75
Score
70
65
60
2018 2019 2020 2021
Figure 3.4.5

TOWAR D T RU THFUL L A NG UAGE “I don’t know” may be truthful but not informative). A
MO DE L S combination of metrics adapted from machine translation
While language models have become increasingly (BLEURT, ROUGE, and BLEU) are used to compare the
more coherent, they have also become more fluent at model answer to reference answers. In this setting, a small
generating factually incorrect statements and fabricating model called GPT-Judge is also learned to predict human
falsehoods. This increased fluency means they also have evaluations of truthfulness and informativeness for a
a greater capacity to perpetuate harm by, for instance, given answer. Alternatively, the task can be formulated
creating convincing conspiracy theories. as a multiple-choice task where models are evaluated on
their accuracy in selecting the correct answer.
Model Size and Truthfulness
TruthfulQA is a benchmark designed to evaluate the In the multiple-choice version of this task, initial
truthfulness of language models on question answering, experiments on GPT-Neo, GPT-2, T5 (UnifiedQA), and GPT-
with questions drawn from categories such as health, 3 showed that larger models provide more informative
law, finance, and politics. Questions were designed such answers but are not necessarily more truthful. Later
that humans might be prone to answer incorrectly due experiments on DeepMind’s Gopher model contradicted
to common misconceptions: For example, when asked, this finding: Figure 3.4.6 from the Gopher paper
“What happens if you smash a mirror?” GPT-3 responds, shows that accuracy improves with model size on the
“You will have seven years of bad luck.” multiple-choice task. This contradiction may be due to
the formulation of the TruthfulQA dataset, which was
The task is formulated as an open-ended generative collected adversarially against GPT-3 175-B, possibly
task, where systems are evaluated on the truthfulness explaining the lower performance of the GPT-3 family of
and informativeness of their answers (e.g., answering models.
TRUTHFULQA MULTIPLE-CHOICE TASK: TRUTHFUL and INFORMATIVE ANSWERS by MODEL

40%
Truthful and Informative Answers (%)
30%
20%
10%
0%
GPT-NEO-125M
GPT-NEO-1.3B
GPT-NEO-2.7B
GPT-NEO-6B
T5 60M
T5 220M
T5 770M
T5 2.8B
GPT-2 117M
GPT-2 1.5B
GPT3 350M
GPT3 1.3B
GPT3 6.7B
GPT3 175B
Gopher 1.4B
Gopher 7.1B
Gopher 280B
Gopher 280B
-10shot
Model
Figure 3.4.6

WebGPT was designed to improve the factual accuracy models using human-curated responses are called SFT
of GPT-3 by introducing a mechanism to search the Web (supervised fine-tuning). The baseline SFT is further
for sources to cite when providing answers to questions. fine-tuned using reinforcement learning from human
Similar to Gopher, WebGPT also shows more truthful feedback. This family is called PPO because it uses a
and informative results with increased model size. While technique called Proximal Policy Optimization. Finally,
performance improves compared to GPT-3, WebGPT PPO models are further enhanced and called InstructGPT.
still struggles with out-of-distribution questions, and its
Figure 3.4.7 shows the truthfulness of eight language
performance is considerably below human performance.
model families on the TruthfulQA generation task. Similar
However, since WebGPT cites sources and appears more
to the scaling effect observed in the Gopher family, the
authoritative, its untruthful answers may be more harmful
WebGPT and InstructGPT models yield more truthful and
as users may not investigate cited material to verify each
informative answers as they scale. The exception to the
source.
scaling trend is the supervised fine-tuned InstructGPT
InstructGPT models are a variant of GPT-3 and they use baseline, which corroborates observations from the
human feedback to train a model to follow instructions, TruthfulQA paper that the baseline GPT-3 family of models
created by fine-tuning GPT-3 on a dataset of human- underperforms with scale.
written responses to a set of prompts. The fine-tuned
TRUTHFULQA GENERATION TASK: TRUTHFUL and INFORMATIVE ANSWERS by MODEL

Source: Rae et al., 2021; Nakano, 2021; Ouyang, 2022 | Chart: 2022 AI Index Report
Truthful and Informative Answers (%)
60%
40%
20%
0%
PPO- netuned SFT 1.3B
PPO- netuned SFT 6B
PPO- netuned SFT 175B
InstructGPT 1.3B
GPT-NEO-125M
GPT-2 117M
GPT-2 1.5B
GPT-NEO-1.3B
GPT3 760M
WebGPT 760M
WebGPT 175B
InstructGPT 6B
InstructGPT 175B
GPT-NEO-2.7B
GPT-NEO-6B
T5 60M
T5 220M
T5 770M
T5 2.8B
GPT3 13B
GPT3 175B
WebGPT 13B
SFT 1.3B
SFT 6B
SFT 175B
Model
Figure 3.4.7

Multimodal Biases in Contrastive

Language-Image Pretraining (CLIP)
Techniques used in natural language processing “criminal,” and “suspicious person” to the FairFace
such as the transformer architecture have recently dataset classes resulted in images of Black
been adapted to the vision and multimodal people being misclassified as nonhuman at a
domains. General-purpose models such as CLIP, significantly higher rate than any other race (14%,
ALIGN, FLAVA, Florence, and Wu Dao 2 are compared to the next highest misclassification
trained on joint vision-language datasets compiled rate of 7.6% for images of Indians). People ages
from the internet and can be used for a wide range 20 years old and younger were also more likely to
of downstream vision tasks, such as classification. be assigned to crime-related classes compared to
all other age groups.
CLIP (Contrastive Language-Image Pretraining) is
a model that learns visual concepts from natural Gender Bias
language by training on 400 million image-text Probing CLIP with the Members of Congress
pairs scraped from the internet, and it is capable dataset shows that labels such as “nanny” and
of outperforming the best ImageNet-trained “housekeeper” were associated with women,
models on a variety of visual classification tasks. whereas labels such as “prisoner” and “mobster”
Like other models pretrained on internet corpora, were associated with men. Figure 3.4.8 shows
CLIP exhibits biases along gender, race, and age. the percentage of images in the Members of
However, while benchmarks exist for measuring Congress dataset that are attached to a certain
bias within computer vision and natural language, label by gender, reflecting similar gender biases
there are no well-established metrics for found in commercial image recognition systems.
measuring multimodal bias. This section provides Additionally, CLIP almost exclusively associates
insight into some ways that researchers have high-status occupation labels like “executive”
probed CLIP for bias. and “doctor” with men, and disproportionately
attaches labels related to physical appearance
Denigration Harm to women. These experiments show that design
Exploratory probes show that the design of decisions such as selecting the correct similarity
categories used in the model (i.e., ground-truth thresholds can have outsized impacts on model
labels) heavily influences the biases manifested performance and biases.
by CLIP. Probing the model by adding non-human
and crime-related classes such as “animal,”
“gorilla,” “chimpanzee,” “orangutan,” “thief,”


Language-Image Pretraining (CLIP) (cont’d)
BIAS in CLIP: FREQUENCY of IMAGE LABELS by GENDER

Source: Agarwal et al., 2021 | Chart: 2022 AI Index Report
Women Men
Lady Man
Woman Male
Female Face
Looking Player
Senior Citizen Black
Public Speaking Head
Spokesperson Facial Expression
Blonde Suit
Laughing Photo
Blazer Military O#cer
Bob Cut Walking
Magenta Photograph
Hot Elder
Black Hair Tie
Pixie Cut Display
Bangs Shoulder
Pink Frown
Man Man
Newsreader Kid
Woman Woman
Purple Necktie
Blouse Yellow
0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100%
Frequency (%) Frequency (%)
Figure 3.4.8


Language-Image Pretraining (CLIP) (cont’d)
Propagating Learned Bias Downstream This is problematic when CLIP is used for curating
CLIP has also been shown to learn historical biases datasets. Embeddings from CLIP were used to filter
and conspiracy theories from its internet-sourced the LAION-400M for high-quality image-text pairs;
training dataset. As one example of learned however, the biases learned by CLIP were shown to
historical bias, Figure 3.4.9 shows that CLIP assigns be propagated to LAION-400M, thus affecting any
higher similarity to “housewife with an orange future applications built with LAION-400M.
jumpsuit” to a picture of astronaut Eileen Collins.
RESULTS OF THE
CLIP-EXPERIMENTS
PERFORMED WITH THE
COLOR IMAGE OF THE
ASTRONAUT EILEEN
Source: Birthane et al., 2021
Figure 3.4.9
Underperformance on Non-English text, and its performance has not been evaluated
Languages on other languages.
CLIP can be extended to non-English languages
However, mBERT has performance gaps on low-
by replacing the original English text encoder
resource languages such as Latvian or Afrikaans,14
with a pretrained, multilingual model such as
which means that multilingual versions of CLIP
Multilingual BERT (mBERT) and fine-tuning
trained with mBERT will still underperform. Even
further. However, its documentation cautions
for high-resource languages, such as French and
against using the model for non-English
Spanish, there are still noticeable accuracy gaps
languages since CLIP was trained only on English
in gender and age classification.
14 While mBERT performs well on high-resource languages like French, on 30% of languages (out of 104 total languages) with lower pretraining resources, it performs worse than using no pretrained
model at all.

Index Report 2022
CHAPTER 4:
The Economy
and Education
Index Report 2022 CHAPTER 4: THE ECONOMY AND EDUCATION
CHAPTER 4:
Chapter Preview
Overview 141 4.3 CORPORATE ACTIVITY 160
Chapter Highlights 142 Industry Adoption 160
Global Adoption of AI 160
4.1 JOBS 143 AI Adoption by Industry and Function 161
AI Hiring 143 Type of AI Capabilities Adopted 162
AI Labor Demand 145 Consideration and Mitigation of Risks
Global AI Labor Demand 145 From Adopting AI 163
U.S. AI Labor Demand: By Skill Cluster 146
U.S. Labor Demand: By Sector 147 4.4 AI EDUCATION 165
U.S. Labor Demand: By State 147 CS Undergraduate Graduates in
AI Skill Penetration 149 North America 165
Global Comparison 149 New CS PhDs in North America 166
Global Comparison: By Industry 149 New CS PhDs by Specialty 166
Global Comparison: By Gender 150 New CS PhDs with AI/ML

and Robotics/Vision Specialties 167
New AI PhDs Employment in North America 168
4.2 INVESTMENT 151
Academia vs. Industry vs. Government 168
Corporate Investment 151
Diversity of New AI PhDs in North America 169
Startup Activity 152
By Gender 169
Global Trend 152
By Race/Ethnicity 170
Regional Comparison by Funding Amount 154
New International AI PhDs in North America 171
Regional Comparison by Newly Funded
AI Companies 156
Focus Area Analysis 158 ACCESS THE PUBLIC DATA

Overview
The growing use of artificial intelligence (AI) in everyday life, across
industries, and around the world generates numerous questions about
how AI is shaping the economy and education—and, conversely, how
the economy and education are adapting to AI. AI promises many
opportunities in workplace productivity, supply chain efficiency,
customized consumer experiences, and other areas. At the same time,
however, the technology gives rise to a number of concerns. How
do businesses adapt to recruiting and retaining AI talent? How is the
education system keeping pace with the demand for AI labor and the
need to understand AI’s impact on society? All of these questions and
more are inextricable from AI today.
This chapter examines the economy and education, using data

from Emsi Burning Glass, NetBase Quid, and LinkedIn to capture AI
trends in the global economy and data from the annual Computing
Research Association Taulbee Report to analyze trends in AI and
computer science PhD graduates. The chapter first examines AI’s
impact on jobs, including hiring, labor demand, and skill penetration
rate, followed by an analysis of corporate investments in AI—from
global trends to startup activity in the space, and the adoption of AI
technologies among industries. The final section discusses computer
science (CS) undergraduate graduates and PhD graduates who
specialize in AI-related disciplines.

CHAPTER HIGHLIGHTS
• New Zealand, Hong Kong, Ireland, Luxembourg, and Sweden are the countries or regions with
the highest growth in AI hiring from 2016 to 2021.
• In 2021, California, Texas, New York, and Virginia were states with the highest number of AI job
postings in the United States, with California having over 2.35 times the number of postings
as Texas, the second greatest. Washington, D.C., had the greatest rate of AI job postings
compared to its overall number of job postings.
• The private investment in AI in 2021 totaled around $93.5 billion—more than double the
total private investment in 2020, while the number of newly funded AI companies continues
to drop, from 1051 companies in 2019 and 762 companies in 2020 to 746 companies in 2021.
In 2020, there were 4 funding rounds worth $500 million or more; in 2021, there were 15.
• “Data management, processing, and cloud” received the greatest amount of private AI
investment in 2021—2.6 times the investment in 2020, followed by “medical and healthcare”
and “fintech.”
• In 2021, the United States led the world in both total private investment in AI and the number
of newly funded AI companies, three and two times higher, respectively, than China, the next
country on the ranking.
• Efforts to address ethical concerns associated with using AI in industry remain limited,
according to a McKinsey survey. While 29% and 41% of respondents recognize “equity and
fairness” and “explainability” as risks while adopting AI, only 19% and 27% are taking steps
to mitigate those risks.
• In 2020, 1 in every 5 CS students who graduated with PhD degrees specialized in artificial
intelligence/machine learning, the most popular specialty in the past decade. From 2010 to
2020, the majority of AI PhDs in the United States headed to industry while a small fraction
took government jobs.

Artificial Intelligence CHAPTER 4: THE ECONOMY AND EDUCATION
Index Report 2022 4.1 Jobs
4.1 JOBS
AI HIRING example, an index of 1.05 in December 2021 points to a
The AI hiring data draws on a dataset from LinkedIn hiring rate that is 5% higher than the average month in
of skills and jobs listings on the platform. It focuses 2016. LinkedIn makes month-to-month comparisons to
specifically on countries or regions where LinkedIn covers account for any potential lags in members updating their
at least 40% of the labor force and where there are at profiles. The index for a year is the number in December
least 10 AI hires each month. China and India were also of that year.
included due to their global importance, despite not The relative AI hiring index captures whether hiring of
meeting the 40% coverage threshold. Insights for these AI talent is growing faster than, equal to, or more slowly
countries may not provide as full a picture as others, and than overall hiring in a particular country or region.
should be interpreted accordingly. New Zealand has the highest growth in AI hiring—2.42
Figure 4.1.1 shows the 15 geographic areas with the times greater in 2021 compared with 2016, followed by
highest relative AI hiring index for 2021. The AI hiring Hong Kong (1.56), Ireland (1.28), Luxembourg (1.26),
rate is calculated as the percentage of LinkedIn members and Sweden (1.24). Moreover, many countries or regions
with AI skills on their profile or working in AI-related experienced a decrease in their AI hiring growth from
occupations who added a new employer in the same 2020 to 2021—indicating that the pace of change in the
period the job began, divided by the total number of AI hiring rate, against the rate of overall hiring, declined
LinkedIn members in the corresponding location. This over the last year, with the exception of Germany and
rate is then indexed to the average month in 2016; for Sweden (Figure 4.1.2).
RELATIVE AI HIRING INDEX by GEOGRAPHIC AREA, 2021

Source: LinkedIn, 2021 | Chart: 2022 AI Index Report
New Zealand 2.42
Hong Kong 1.56
Ireland 1.28
Luxembourg 1.26
Sweden 1.24
Netherlands 1.23
United Kingdom 1.20
China 1.18
United States 1.17
Australia 1.14
Canada 1.14
South Africa 1.13
Spain 1.10
Denmark 1.09
Germany 1.08
0.00 0.50 1.00 1.50 2.00 2.50

Relative AI Hiring Index
Figure 4.1.1

RELATIVE AI HIRING INDEX by GEOGRAPHIC AREA, 2016–21

Australia Canada China Denmark

1.24
1.08
1.00 1.18
1.07
0.00
Germany Hong Kong Ireland Luxembourg
1.15 1.14
Relative AI Hiring Index
1.00 1.13 1.16
0.00
Netherlands New Zealand 1.40 South Africa Spain
1.20
1.00 1.18 1.14
0.00
Sweden United Kingdom United States
1.24 1.16
1.00 1.12
0.00
2018 2020 2022 2018 2020 2022 2018 2020 2022 2018 2020 2022
Figure 4.1.2

AI LABOR DEMAND Global AI Labor Demand

To analyze demand for specific AI labor skills, Emsi Figure 4.1.3 shows that the percentage of AI job postings
Burning Glass mined millions of job postings collected among all job postings in 2021 was greatest in Singapore
from over 45,000 websites since 2010 and flagged all (2.33% of all job listings), followed by the United States
listings calling for AI skills. (0.90%), Canada (0.78%), and the United Kingdom
(0.74%). AI job postings increased in the United States,
Canada, Australia, and New Zealand from 2020 to
2021, while they declined in Singapore and the United
Kingdom.
AI JOB POSTINGS (% of ALL JOB POSTINGS) by GEOGRAPHIC AREA, 2013–21

Source: Emsi Burning Glass, 2021 | Chart: 2022 AI Index Report
2.50%
2.33%, Singapore
2.00%
AI Job Postings (% of All Job Postings)
1.50%
1.00%
0.78%, Canada
0.74%, United Kingdom
0.58%, Australia
0.50%
0.25%, New Zealand
2013 2014 2015 2016 2017 2018 2019 2020 2021

Figure 4.1.3

U.S. AI Labor Demand: By Skill Cluster The share of AI job postings

Figure 4.1.4 shows the U.S. labor demand from 2010
to 2021 by skill cluster. Each skill cluster consists of a among all job postings
list of AI-related skills; for example, the neural network
skill cluster includes skills like deep learning and
in 2021 was greatest for
convolutional neural networks.1 The share of AI job machine learning skills (0.6%
postings among all job postings in 2021 was greatest
for machine learning skills (0.6% of all job postings),
of all job postings), followed
followed by artificial intelligence (0.33%), neural by artificial intelligence
networks (0.16%), and natural language processing
(0.13%). Postings for AI jobs in machine learning and (0.33%), neural networks
artificial intelligence have significantly increased in (0.16%), and natural
the past couple of years, despite small declines in
both categories from 2019–2020. Machine learning language processing (0.13%).
jobs are at nearly three times the level, and artificial
intelligence jobs are at around 1.5 times the level they
each reached, respectively, in 2018.
AI JOB POSTINGS (% of ALL JOB POSTINGS) in the UNITED STATES by SKILL CLUSTER, 2010–21
0.57%, Machine Learning
0.50%
0.40%
0.33%, Arti cial Intelligence
0.30%
0.20%
0.15%, Neural Networks

0.13%, Natural Language Processing (NLP)
0.10% 0.11%, Robotics
0.10%, Visual Image Recognition
0.06%, Autonomous Driving
0.00%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 4.1.4
1 See the Appendix for a complete list of AI skills under each skill cluster.

U.S. Labor Demand: By Sector followed by professional, scientific, and technical

Figure 4.1.5 shows that 3.30% of all job postings in the services (2.59% of all listings), manufacturing (2.02%),
information sector in the United States were AI-related, and finance and insurance (1.81%).
AI JOB POSTINGS (% of ALL JOB POSTINGS) in the UNITED STATES by SECTOR, 2021
Information 3.30%
Professional, Scienti!c, and Technical Services 2.59%
Manufacturing 2.02%
Finance and Insurance 1.81%
Agriculture, Forestry, Fishing and Hunting 0.95%
Public Administration 0.95%
Educational Services 0.84%
Utilities 0.78%
Mining, Quarrying, and Oil and Gas Extraction 0.69%
Management of Companies and Enterprises 0.67%
Wholesale Trade 0.63%
Retail Trade 0.54%
Real Estate and Rental and Leasing 0.40%
Transportation and Warehousing 0.33%
Waste Management and Remediation Services 0.28%
0.00% 1.00% 2.00% 3.00%

Figure 4.1.5
U.S. Labor Demand: By State NUMBER of AI JOB POSTINGS in the UNITED STATES by STATE, 2021
Figure 4.1.6 breaks down the U.S.
AI labor demand by state. In 2021,
the top states posting AI jobs
AK ME
were California (80,238), Texas 481 693
(34,021), New York (24,494), and VT NH

552 1,211
Virginia (19,387). California, in first,
had over 2.35 times the number WA
19,253
ID
2,309
MT
629
ND
726
MN
7,453
IL
15,775
WI
5,142
MI
9,935
NY
24,494
RI
1,090
MA
18,430
of postings as Texas, the second

OR NV WY SD IA IN OH PA NJ CT
greatest. Proportionally, however, 4,393 2,253 340 693 2,388 3,899 9,593 10,806 10,503 4,291
Washington, D.C., had the greatest CA UT CO NE MO KY WV VA MD DE

80,238 3,631 11,189 1,385 5,314 2,107 706 19,387 7,745 1,764
rate of AI job postings compared to
its overall number of job postings AZ
9,606
NM
1,719
KS
2,058
AR
1,539
TN
5,219
NC
12,863
SC
2,259
DC
6,381
(Figure 4.1.7). That was followed by
OK LA MS AL GA
Virginia, Washington, Massachusetts, 1,462 1,437 986 3,605 14,283
and California. HI TX FL
866 34,021 14,972
Figure 4.1.6

AI JOB POSTINGS (TOTAL and % of ALL JOB POSTINGS) by U.S. STATE and DISTRICT, 2021
District of Columbia
2.00%
Virginia
1.50% Washington
Massachusetts California
Delaware New York
1.00% Maryland Georgia

New Mexico
Minnesota Illinois Texas
Vermont Utah
Missouri Arizona
Wyoming Maine Hawaii Arkansas
Alabama Ohio
0.50%
Mississippi Kansas Florida
Alaska
Indiana
Louisiana
0.00%
200 500 1,000 2,000 5,000 10,000 20,000 50,000 100,000
Number of AI Job Postings (Log Scale)
Figure 4.1.7

A I S K I L L P E N E T R AT I O N AI skills is measured as the sum of the penetration of

The AI skill penetration rate shows the prevalence each AI skill across occupations in a given country or
of AI skills across occupations, or the intensity with region, divided by the global average across the same
which LinkedIn members use AI skills in their jobs. It is occupations. For example, a relative penetration rate of
calculated by computing the frequencies of LinkedIn 2 means that the average penetration of AI skills in that
users’ self-added skills in a given area from 2015–2021, country or region is 2 times the global average across the
then reweighting those figures by using a statistical same set of occupations. Figure 4.1.8 shows that India
model to get the top 50 representative skills in that led the world in the rate of AI skill penetration—3.09
occupation. times the global average from 2015 to 2021—followed
by the United States (2.24) and Germany (1.7). After that
Global Comparison came China (1.56), Israel (1.52), and Canada (1.41).2
For global comparison, the relative penetration rate of
RELATIVE AI SKILL PENETRATION RATE by GEOGRAPHIC AREA, 2015–21

India 3.09
United States 2.24
Germany 1.70
China 1.56
Israel 1.52
Canada 1.41
United Kingdom 1.40
South Korea 1.28
Singapore 1.22
France 1.11
Spain 0.84
Australia 0.77
Switzerland 0.76
Italy 0.74
Sweden 0.60
0.00 0.50 1.00 1.50 2.00 2.50 3.00

Relative AI Skill Penetration Rate
Figure 4.1.8
Global Comparison: By Industry manufacturing, education, and finance (Figure 4.1.9).

India and the United States had the highest relative Israel and Canada are among the top seven countries
AI skill penetration across the board—leading the across all five industries, and Singapore holds the fourth
other countries or regions in skill penetration rates in position on the list.
software and IT services, hardware and networking,
2 Those included are a sample of eligible countries or regions with at least 40% labor force coverage by LinkedIn and at least 10 AI hires in any given month. China and India were also included in this
sample because of their increasing importance in the global economy, but LinkedIn coverage in these countries does not reach 40% of the workforce. Insights for these countries may not provide as full a
picture as in others, and should be interpreted accordingly.

RELATIVE AI SKILL PENETRATION RATE by INDUSTRY across GEOGRAPHIC AREA, 2015–21

Education Finance Hardware

India 3.96 India 3.08 India 2.92
United States 2.20 United States 1.89 United States 1.98
Israel 1.35 Germany 1.48 Israel 1.37
Germany 1.27 Canada 1.31 China 1.34
Canada 1.23 Singapore 1.28 Germany 1.22
South Korea 1.18 Israel 1.15 Canada 1.12
United Kingdom 0.91 Netherlands 1.12 Singapore 1.12
France 0.80 United Kingdom 1.00 South Korea 0.88
1.00 2.00 3.00 4.00 0.00 1.00 2.00 3.00 0.00 1.00 2.00 3.00
Manufacturing Software
India 2.39 India 3.02
United States 2.21 United States 2.18
Germany 1.66 China 1.58
Israel 1.31 Germany 1.38
Sweden 1.16 Israel 1.22
Singapore 1.04 Canada 1.20
Canada 0.99 Singapore 1.16
China 0.93 South Korea 1.13
0.00 1.00 2.00 0.00 1.00 2.00 3.00
Figure 4.1.9
Global Comparison: By Gender

Figure 4.1.10 shows the aggregated data from 2015 to 2021 of AI skills penetration by geographic area for female and
male talent. The data suggests that among the 15 countries listed, the AI skill penetration rates of females are higher than
those of males in India, Canada, South Korea, Australia, Finland, and Switzerland.
RELATIVE AI SKILL PENETRATION RATE by GENDER, 2015–21

India
United States
Canada
South Korea
China
Singapore
Germany
Israel
Australia
Finland
United Kingdom
Netherlands
Female
France Male
Switzerland
Brazil
0.00 1.00 2.00 3.00 4.00

Figure 4.1.10

Index Report 2022 4.2 Investment
This section on corporate AI activity draws on data from NetBase Quid, which aggregates over 6 million global public and private
company profiles, updated on a weekly basis, including metadata on investments, location of headquarters, and more. NetBase Quid
also applies natural language processing technology to search, analyze, and identify patterns in large, unstructured datasets, like
aggregated blogs, company and patent databases.
4.2 INVESTMENT
C O R P O R AT E I N V E S T M E N T investment through private investment (totaling around
Corporate investment in artificial intelligence, from $93.5 billion), followed by mergers and acquisitions
mergers and acquisitions to public offerings, is a key (around $72 billion), public offerings (around $9.5
contributor to AI research and development. It also billion), and minority stake (around $1.3 billion). In
contributes to AI’s impact on the economy. Figure 4.2.1 2021, investments from mergers and acquisitions grew
highlights overall global corporate investment in AI from by 3.3 times compared to 2020, led by two AI healthcare
2013–2021. In 2021, companies made the greatest AI companies and two cybersecurity companies.3
GLOBAL CORPORATE INVESTMENT in AI by INVESTMENT ACTIVITY, 2013–21

Source: NetBase Quid, 2021 | Chart: 2022 AI Index Report
176.47
Merger/Acquisition
Minority Stake
Private Investment
Total Investment (in billions of U.S. Dollars)
150
Public O"ering
72.03
119.54
21.53
100
44.06
66.98
20.42 93.54
50 47.33 45.58
22.49
29.17 46.00
19.83 35.26 42.57
14.99
5.23 22.78 23.09
8.98 15.73
0 7.94 9.58
2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 4.2.1
3 Nuance Communications (by Microsoft, $19.8 billion), Varian Medical Systems (Siemens, $17.2 billion), and Proofpoint (Thoma Bravo, $12.4 billion) in the United States, followed by Avast in the
Czech Republic (NortonLifeLock, $8.0 billion).

S TA R T U P ACT I V I T Y However, Figure 4.2.3 shows that the number of newly

The following section analyzes artificial intelligence and funded AI companies continues to drop, from 762
machine learning companies globally that have received companies in 2020 to 746 companies in 2021—the
more than $1.5 million in investment from 2013 to 2021. third year of a decline that started in 2018. The largest
private investments in 2021 have been led by two data
Global Trend management companies and two robotics/autonomous
In 2021, global private investment in AI totaled around driving companies.4
$93.5 billion, which is more than double the total
private investment in 2020 (Figure 4.2.2). That marks
the greatest year-over-year increase since 2014 (when In 2021, global private
investment from 2013 to 2014 more than doubled). investment in AI totaled
Among companies that disclosed the amount of funding,
around $93.5 billion,
the number of AI funding rounds that ranged from
$100 million to $500 million more than doubled in 2021 which is more than
compared to 2020, while funding rounds that were
between $50 million and $100 million more than doubled
double the total private
as well (Table 4.2.1). In 2020, there were only four funding investment in 2020.
rounds worth $500 million or more; in 2021, that number
grew to 15. Companies attracted significantly higher
investment in 2021, as the average private investment
deal size in 2021 was 81.1% higher than in 2020.
PRIVATE INVESTMENT in AI, 2013–21

100
93.54
Total Investment (in bIllions of U.S. Dollars)
80
60
40
20
0
2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 4.2.2
4 The largest private investments have been Databricks (United States), Beijing Horizon Robotics Technology (China), Oxbotica Limited (United Kingdom), and Celonis (Germany).

NUMBER of NEWLY FUNDED AI COMPANIES in the WORLD, 2013–21

1200
1000
Number of Companies
800 746
600
400
200
0
2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 4.2.3
Funding Size 2020 2021 Total
Over $1 billion 3 5 8
$500 million – $1 billion 1 10 11
$100 million – $500 million 93 235 328
$ 50 million – $100 million 85 194 279
Under $50 million 2,102 2,120 4,222
Undisclosed 354 395 749
Total 2,638 2,959 5,597
Table 4.2.1

Regional Comparison by Funding Amount Kingdom ($10.8 billion), India ($10.77 billion), and Israel
In 2021, as captured in Figure 4.2.4, the United States ($6.1 billion). Notably, U.S. private investment in AI
led the world in overall private investment in funded AI companies from 2013–2021 was more than double the
companies—at approximately $52.9 billion—over three total in China, which itself was about six times the total
times the next country on the list, China ( $17.2 billion). investment from the United Kingdom in the same period.
In third place was the United Kingdom ($4.65 billion), Broken out by geographic area, as shown in Figure
followed by Israel ($2.4 billion) and Germany ($1.98 4.2.6, the United States, China, and the European Union
billion). Figure 4.2.5 shows that when combining total all grew their investments from 2020 to 2021, with the
private investment from 2013 to 2021, the same ranking United States leading China and the European Union by
applies: U.S. investment totaled $149 billion and Chinese 3.1 and 8.2 times the investment amount, respectively.
investment totaled $61.9 billion, followed by the United
PRIVATE INVESTMENT in AI by GEOGRAPHIC AREA, 2021

United States 52.88
China 17.21
United Kingdom 4.65
Israel 2.41
Germany 1.98
Canada 1.87
France 1.55
India 1.35
Australia 1.25
South Korea 1.10
Singapore 0.93
Spain 0.89
United Arab Emirates 0.82
Hong Kong 0.63
Portugal 0.52
0 10 20 30 40 50
Figure 4.2.4

PRIVATE INVESTMENT in AI by GEOGRAPHIC AREA, 2013–21

United States 149.0
China 61.9
United Kingdom 10.8
India 10.8
Israel 6.1
Canada 5.7
Germany 4.4
France 3.9
Singapore 2.3
Japan 2.2
Australia 1.8
South Korea 1.8
Spain 1.3
Hong Kong 1.2
Netherlands 1.2
0 20 40 60 80 100 120 140 160
Figure 4.2.5
PRIVATE INVESTMENT in AI by GEOGRAPHIC AREA, 2013–21

52.87, United States
50
40
30
20 17.21, China
10
6.42, European Union
2013 2014 2015 2016 2017 2018 2019 2020 2021

Figure 4.2.6

Regional Comparison by Newly Funded shows a similar trend (Figure 4.2.8).

AI Companies
However, the number of newly funded AI companies has
This section breaks down the investment data by the
declined in both the United States and China since 2018
number of newly funded AI companies in each region. For
and 2019 (Figure 4.2.9). Despite that downward trend,
2021, Figure 4.2.7 shows that the United States led with
the United States still leads in the number of newly
299 companies, followed by China with 119, the United
funded companies, with 299 funded in 2021, followed by
Kingdom with 49, and Israel with 28. The gaps between
China (119) and the European Union (96).
each are significant. Aggregated data from 2013 to 2021
NUMBER of NEWLY FUNDED AI COMPANIES by GEOGRAPHIC AREA, 2021

United States 299
China 119
United Kingdom 49
Israel 28
France 27
Germany 25
India 23
Canada 22
South Korea 19
Netherlands 16
Australia 12
Japan 12
Switzerland 12
Singapore 10
0 50 100 150 200 250 300

Number of Companies
Figure 4.2.7

NUMBER of NEWLY FUNDED AI COMPANIES by GEOGRAPHIC AREA, 2013–21 (SUM)

United States 3,234
China 940
United Kingdom 427
Israel 264
France 242
Canada 239
Japan 172
India 169
Germany 154
Singapore 99
Australia 77
South Korea 74
Switzerland 65
Spain 56
0 500 1,000 1,500 2,000 2,500 3,000 3,500

Number of Companies
Figure 4.2.8
NUMBER of NEWLY FUNDED AI COMPANIES by GEOGRAPHIC AREA, 2013–21

500
400
Number of Companies
300
299, United States
200
119, China
100 106, European Union
0
2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 4.2.9

Focus Area Analysis Aggregated data in Figure 4.2.11 shows that in the last
Private AI investment also varies by focus area. According five years, the medical and healthcare category received
to Figure 4.2.10, the greatest private investment in AI in the largest private investment globally ($28.9 billion);
2021 was in data management, processing, and cloud followed by data management, processing, and cloud
(around $12.2 billion). Notably, this was 2.6 times the ($26.9 billion); fintech ($24.9 billion); and retail ($21.95
investment in 2020 (around $4.69 billion) as two of the billion). Moreover, Figure 4.2.12 shows the overall trend
four largest private investments made in 2021 are data in private investment by industries from 2017–2021 and
management companies. In second place was private reveals a steady increase in AV, cybersecurity and data
investment in medical and healthcare ($11.29 billion), protection, fitness and wellness, medical and healthcare,
followed by fintech ($10.26 billion), AV ($8.09 billion), and and semiconductor industries.
semiconductors ($6.0 billion).
PRIVATE INVESTMENT in AI by FOCUS AREA, 2020 vs. 2021 PRIVATE INVESTMENT in AI by FOCUS AREA, 2017–21 (SUM)
Source: NetBase Quid, 2021 | Chart: 2022 AI Index Report Source: NetBase Quid, 2021 | Chart: 2022 AI Index Report
Data Management, Processing, Cloud Medical and Healthcare 28.92

Medical and Healthcare Data Management, Processing, Cloud 26.91
Fintech Fintech 24.92
AV Retail 21.95
Semiconductor 2021 Industrial Automation, Network 19.03
Industrial Automation, Network 2020 AV 18.60
Retail Semiconductor 11.91
Fitness and Wellness Fitness and Wellness 10.75
NLP, Customer Support NLP, Customer Support 8.45
Energy, Oil, and Gas Ed Tech 8.16
Cybersecurity, Data Protection Energy, Oil, and Gas 7.82
Drones Facial Recognition 7.49
Marketing, Digital Ads Cybersecurity, Data Protection 7.30
HR Tech Marketing, Digital Ads 7.20
Facial Recognition AR/VR 4.46
Insurtech HR Tech 4.27
Agritech Drones 3.91
Sales Enablement Sales Enablement 3.54
AR/VR Insurtech 3.29
Ed Tech Agritech 3.13
Geospatial Legal Tech 2.62
Legal Tech Geospatial 1.82
Entertainment Music, Video Content 1.33
Music, Video Content Entertainment 1.31
VC VC 1.20
0 2 4 6 8 10 12 0 5 10 15 20 25 30
Total Investment (in billions of U.S. Dollars) Total Investment (in billions of U.S. Dollars)
Figure 4.2.10 Figure 4.2.11

PRIVATE INVESTMENT in AI by FOCUS AREA, 2017–21

Agritech AR/VR AV Cybersecurity, Data Protection Data Management, Processing, Cloud

10 8.09
2.65 12.17
1.43 1.26
0
Drones Ed Tech Energy, Oil, and Gas Entertainment Facial Recognition
10
2.48 2.78 1.87

0.71
0 1.24
Fintech Fitness and Wellness Geospatial HR Tech Industrial Automation, Network

10.26
10
5.56 5.85
0.93 2.18
0
Insurtech Legal Tech Marketing, Digital Ads Medical and Healthcare Music, Video Content
11.29
10
1.77 2.31
0.91 0.47
0
NLP, Customer Support Retail Sales Enablement Semiconductor VC

10
5.74 6.00
4.01
1.28
0.29
0
2018 2020 2018 2020 2018 2020 2018 2020 2018 2020
Figure 4.2.12

Index Report 2022 4.3 Corporate Activity
4.3 CORPORATE ACTIVITY

I N D U S T RY A D O P T I O N Global Adoption of AI
This section on corporate AI activity draws on McKinsey’s Figure 4.3.1 shows AI adoption by organizations globally,
“The State of AI in 2021” report from December 2021. The broken out by geographic area. In 2021, India led with
report based its conclusions on a global online survey 65% adoption, followed by “Developed Asia-Pacific”
of 1,843 participants conducted earlier in 2021. Survey (64%), “Developing markets (incl. China, MENA)” (57%),
respondents came from a range of industries, companies, and North America (55%). The average adoption rate
functional specialties, tenures, and regions of the world— across all geographies was 56%, up 6% from 2020.
and each provided answers to questions about the state Notably, “Developing markets (incl. China, MENA)”
of artificial intelligence today. registered a 21% increase from 2020, and India had an
8% increase from 2020.
AI ADOPTION by ORGANIZATIONS in the WORLD, 2020-21

Source: McKinsey & Company, 2021 | Chart: 2022 AI Index Report
All geographies 56%
50%
Developed 64%
Asia-Paci!c
61%
Developing markets 57%

(incl. China, MENA)
45%
Europe 51% 2021

47% 2020
India 65%
57%
Latin America 47%
41%
North America 55%
51%
0% 10% 20% 30% 40% 50% 60% 70%

% of Respondents
Figure 4.3.1

AI Adoption by Industry and Function (45%), followed by service operations for financial
Figure 4.3.2 shows AI adoption by industry and function services (40%), service operations for high tech/
in 2021. The greatest adoption was in product and/or telecommunications (34%), and risk function for
service development for high tech/telecommunications financial services (32%).
AI ADOPTION by INDUSTRY and FUNCTION, 2021

Product and/or Strategy and Supply-chain
Human Resources Manufacturing Marketing and Sales Risk Service Operations
Service Development Corporate Finance Management
All Industries 9% 12% 20% 23% 13% 25% 9% 13%
Automotive and
Assembly
11% 26% 20% 15% 4% 18% 6% 17%
Business, Legal, and

Professional Services
14% 8% 28% 15% 13% 26% 8% 13%
Industry
Consumer
Goods/Retail
2% 18% 22% 17% 1% 15% 4% 18%
Financial Services 10% 4% 24% 20% 32% 40% 13% 8%
Healthcare
Systems/Pharma and 9% 11% 14% 29% 13% 17% 12% 9%
Medical Products
High Tech/Telecom 12% 11% 28% 45% 16% 34% 10% 16%
% of Respondents (Function)
Figure 4.3.2

Type of AI Capabilities Adopted telecommunications industry (34%), followed by robotic

With respect to the type of AI capabilities embedded process automation for both the financial services and
in standard business processes in 2021, as indicated automotive and assembly industry (33%) and natural
by Figure 4.3.3, the highest rate of embedding was in language text understanding for financial services (32%).
natural language text understanding for the high tech/
AI CAPABILITIES EMBEDDED in STANDARD BUSINESS PROCESSES, 2021

NL Speech Recom- Reinforce- Robotic

Computer Deep Facial Knowledge NL NL Text Un- Physical Transfer Virtual
Un- mender ment Process Simulations
Vision Learning Regonition Graphs Generation derstanding Robotics Learning Agents
derstanding Systems Learning Automation
All Industries 23% 19% 11% 17% 12% 14% 24% 12% 17% 16% 26% 17% 12% 23%
Automotive and
Assembly
15% 14% 9% 16% 3% 11% 12% 24% 12% 5% 33% 27% 6% 12%
Industry
Business, Legal, and

29% 24% 15% 20% 23% 18% 19% 13% 22% 27% 31% 18% 21% 19%
Professional Services
Consumer
Goods/Retail
23% 12% 14% 17% 11% 13% 14% 4% 8% 8% 16% 9% 1% 15%
Financial Services 17% 16% 11% 16% 12% 18% 32% 4% 13% 16% 33% 12% 12% 28%
Healthcare
Systems/Pharma and 30% 25% 12% 19% 10% 8% 26% 28% 22% 13% 28% 22% 19% 31%
Medical Products
High Tech/Telecom 28% 22% 6% 17% 17% 18% 34% 5% 19% 15% 23% 14% 11% 25%
% of Respondents (AI Capability)

Figure 4.3.3

Consideration and Mitigation of Risks Figure 4.3.5 shows risks from AI that organizations are
From Adopting AI taking steps to mitigate. Cybersecurity was the most
The risk from adopting AI that survey respondents frequent response (47% of respondents), followed by
identified as most relevant in 2021 was cybersecurity regulatory compliance (36%), personal/individual privacy
(55% of respondents), followed by regulatory compliance (28%), and explainability (27%). It is worth noting the
(48%), explainability (41%), and personal/individual gaps between risks that organizations find relevant and
privacy (41%) (Figure 4.3.4). Fewer organizations found risks that organizations take steps to mitigate—a gap of 10
AI risks from cybersecurity to be relevant in 2021 than percentage points with equity and fairness (29% to 19%),
in 2020, declining from just over 60% of respondents 12 percentage points with regulatory compliance (48%
expressing concern in 2020 to 55% in 2021. Concern over to 36%), 13 percentage points with personal/individual
AI regulatory compliance risks, meanwhile, remained privacy (41% to 28%), and 14 percentage points with
virtually unchanged from 2020. explainability (41% to 27%).
RISKS from ADOPTING AI that ORGANIZATIONS CONSIDER RELEVANT, 2019–21

60%
55%, Cybersecurity
50%
48%, Regulatory Compliance
41%, Explainability
% of Respondents
40% 41%, Personal/Individual Privacy
35%, Organizational Reputation
30% 29%, Equity and Fairness

26%, Workforce/Labor Displacement
21%, Physical Safety

20%
14%, National Security

10%
9%, Political Stability
0%
2019 2020 2021
Figure 4.3.4

RISKS from ADOPTING AI that ORGANIZATIONS TAKE STEPS to MITIGATE, 2019–21

50%
47%, Cybersecurity
40%
36%, Regulatory Compliance

% of Respondents
30%
28%, Personal/Individual Privacy
27%, Explainability
22%, Organizational Reputation

20%
19%, Equity and Fairness
16%, Physical Safety
15%, Workforce/Labor Displacement
10%
8%, National Security
4%, Political Stability

0%
2019 2020 2021
Figure 4.3.5

Index Report 2022 4.4 AI Education
The following section draws on data from the annual Computing Research Association (CRA) Taulbee Survey. For the latest survey
featured in this section, CRA collected data in Fall 2020 by reaching out to over 200 PhD-granting departments in the United States and
Canada. Results are published in May 2021. The CRA survey documents trends in student enrollment, degree production, employment
of graduates, and faculty salaries in academic units in the United States and Canada that grant doctoral degrees in computer science
(CS), computer engineering (CE), or information (I). Academic units include departments of CS and CE or, in some cases, colleges or
schools of information or computing.
4.4 AI EDUCATION
C S U N D E R G R A D UAT E G R A D UAT E S The number of new CS undergraduate graduates at
I N N O R T H A M E R I CA doctoral institutions in North America has grown 3.5
In North America, most AI-related courses are offered times from 2010 to 2020 (Figure 4.4.1). More than 31,000
as part of the CS curriculum at the undergraduate level. undergraduates completed CS degrees in 2020—an
11.60% increase from the number in 2019.
NUMBER of NEW CS UNDERGRADUATE GRADUATES at DOCTORAL INSTITUTIONS in NORTH AMERICA, 2010–20

Source: CRA Taulbee Survey, 2021 | Chart: 2022 AI Index Report
31.84
Number of New CS Undergraduate Graduates (in thousands)
30
25
20
15
10
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
Figure 4.4.1

N E W C S P H D S I N N O R T H A M E R I CA
The following sections show the trend of CS PhD graduates In 2020, 1 in every 5
in North America with a focus on those with AI-related
specialties.5 The CRA survey includes 20 specialties in total, CS students who
two of which are directly related to the field of AI: artificial
graduated with PhD
intelligence/machine learning (AI/ML) and robotics/vision.
New CS PhDs by Specialty

degrees specialized
In 2020, 1 in every 5 CS students who graduated with in AI/ML, the most
PhD degrees specialized in AI/ML, the most popular
specialty in the past decade (Figure 4.4.2). It is also the
popular specialty
speciality that exhibits the most significant growth from in the past decade.
2010 to 2021, relative to 18 other specializations (Figure
4.4.3). Robotics/vision is also among the most popular
CS specialties of PhD graduates in 2020, registering a
1.4 percentage point change in the share of total new CS
PhDs in the past 11 years.
NEW CS PHDS (% of TOTAL) in the UNITED STATES by SPECIALITY, 2020

Arti!cial Intelligence/Machine Learning 21.00%

Software Engineering 7.30%
Security/Information Assurance 7.10%
Theory and Algorithms 7.00%
Databases/Information Retrieval 6.50%
Robotics/Vision 6.30%
Networks 6.10%
Other 5.80%
Human-Computer Interaction 5.30%
Graphics/Visualization 4.90%
Operating Systems 4.30%
Programming Languages/Compilers 3.40%
Informatics: Biomedical/Other Science 3.30%
High Performance Computing 2.90%
Information Systems 2.70%
Hardware/Architecture 2.40%
Social Computing/Social Informatics/CSCW 2.00%
Computing Education 0.70%
Information Science 0.50%
Scienti!c/Numerical Computing 0.50%
0% 5% 10% 15% 20%
% of New CS PhDs
Figure 4.4.2
5 New CS PhDs in this section include PhD graduates from academic units (departments, colleges, or schools within universities) of computer science in the United States.

PERCENTAGE POINT CHANGE in NEW CS PHDS in the UNITED STATES by SPECIALTY, 2010–20
Arti!cial Intelligence/Machine Learning 6.80%

Human-Computer Interaction 2.50%
Security/Information Assurance 2.10%
Information Systems 1.50%
Robotics/Vision 1.40%
High Performance Computing 0.60%
Databases/Information Retrieval 0.10%
Information Science 0.00%
Social Computing/Social Informatics/CSCW -0.10%
Other -0.50%
Theory and Algorithms -0.60%
Operating Systems -0.60%
Software Engineering -1.20%
Graphics/Visualization -1.50%
Informatics: Biomedical/Other Science -1.60%
Hardware/Architecture -1.70%
Scienti!c/Numerical Computing -1.90%
Programming Languages/Compilers -2.00%
Networks -4.10%
-4% -2% 0% 2% 4% 6% 8%
Percentage Point Change in New CS PhDs
Figure 4.4.3
New CS PhDs with AI/ML and Robotics/Vision Specialties

Between 2010 and 2020, the number of CS PhD graduates with AI/ML and robotics/vision specialities grew by 72.05%
and 50.91%, respectively. The slight decrease in the total number for both specialties from 2019 to 2020 may be due to
the impact of the COVID-19 pandemic.
NEW CS PHDS with AI/ML and ROBOTICS/VISION SPECIALTY NEW CS PHDS (% of TOTAL) with AI/ML and
in the UNITED STATES, 2010–20 ROBOTICS/VISION SPECIALTY in the UNITED STATES, 2010–20
Source: CRA Taulbee Survey, 2021 | Chart: 2022 AI Index Report Source: CRA Taulbee Survey, 2021 | Chart: 2022 AI Index Report
350
20% 21.00%
300
New CS PhDs (% of Total)
Number of Graduates
250 15%
286
277
200 266
233 218
10%
150 171
144 138
161 138
159
100
5% 6.30%
50 91 83
68 65 74 76 72 75 66
55 46
0 0%
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
Arti!cial Intelligence/Machine Learning Arti!cial Intelligence/Machine Learning

Robotics/Vision Robotics/Vision

N E W A I P H D S E M P LOY M E N T I N America who chose to work in the industry dipped

N O R T H A M E R I CA slightly, with its share dropping from 65.7% in 2019 to
Where do new AI PhDs choose to work following 60.2% in 2020, whereas the share of new AI PhDs who
graduation? This section analyzes the employment went into academia and government changed little
trends of new AI PhDs across North America in academia, (Figure 4.4.5a and Figure 4.4.5b). Note that the 2020 data
industry, and government.6 may be impacted by the increasing number of new AI
PhDs who went abroad upon graduation, a number that
Academia vs. Industry vs. Government grew from 19 in 2019 to 32 in 2020.
In 2020, the share of new AI PhD graduates in North
EMPLOYMENT of NEW AI PHDS to ACADEMIA, GOVERNMENT, EMPLOYMENT of NEW AI PHDS (% of TOTAL) to ACADEMIA,
or INDUSTRY in NORTH AMERICA, 2010–20 GOVERNMENT, or INDUSTRY in NORTH AMERICA, 2010–20
Source: CRA Taulbee Survey, 2021 | Chart: 2022 AI Index Report Source: CRA Taulbee Survey, 2021 | Chart: 2022 AI Index Report
250
60%
65 60.24%
200 73
61 50%
New AI PhDs (% of Total)
Number of New AI PhDs
63
150 60 40%
47
72 43 30%
51
100 63 42 24.02%
180
162
153 20%
134
116
50 101
85
76
64
74 77 10%
1.97%
0 0%
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
Academia Government Industry Academia Government Industry

6 New AI PhDs in this section include PhD graduates who specialize in artificial intelligence from academic units (departments, colleges, or schools within universities) of computer science,
computer engineering, and information in the United States and Canada.

D I V E R S I T Y O F N E W A I P H D S I N N O R T H A M E R I CA
By Gender
Figure 4.4.6 shows that the share of new female AI and CS PhDs in North America remains low and has changed little
from 2010 to 2020.
FEMALE NEW AI and CS PHDS (% of TOTAL NEW AI and CS PHDS) in NORTH AMERICA, 2010–20
20.20%, AI
20%
Female New AI PhDs (% of All New AI PhDs)
19.90%, CS
15%
10%
5%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
Figure 4.4.6

By Race/Ethnicity past 11 years. Figure 4.4.8 shows all PhDs awarded in the
According to Figure 4.4.7, among the new AI PhDs from 2010 to United States to U.S. residents across departments of CS, CE,
2020 who are U.S. residents, the largest percentage has been and information between 2010 and 2020. In the past 11 years,
non-Hispanic white and Asian—65.2% and 18.8% on average. the share of new white (non-Hispanic) PhDs has changed little,
By comparison, around 1.5% were Black or African American while the percentage of new Black or African American (non-
(non-Hispanic) and 2.9% were Hispanic on average over the Hispanic) and Hispanic computing PhDs is significantly lower.
NEW U.S. AI RESIDENT PHDS (% of TOTAL) by RACE/ETHNICITY, 2010–20

80%
70%
New U.S. AI Resident PhDs (% of Total)
60%
50% 50.86%, White (Non-Hispanic)
40%
30% 30.17%, Asian
20%
8.62%, Unknown
6.90%, Hispanic
10% 1.72%, Black or African American (Non-Hispanic)
0.86%, Native Hawaiian or Paci c Islander
0.86%, Multiracial (Non-Hispanic)
0.00%, Native American or Alaskan Native
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
Figure 4.4.7
NEW COMPUTING PHDS, U.S. RESIDENT (% of TOTAL) by RACE/ETHNICITY, 2010–20

70%
60%
New Computing PhDs, U.S. Resident (% of Total)
57.50%, White (Non-Hispanic)
50%
40%
30%
24.80%, Asian
20%
7.80%, Unknown
10% 4.20%, Hispanic
3.60%, Black or African American (Non-Hispanic)
1.80%, Multiracial (Non-Hispanic)
0.10%, Native American or Alaskan Native
0%
0.10%, Native Hawaiian or Paci c Islander
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
Figure 4.4.8

N E W I N T E R N AT I O N A L A I P H D S I N of all computing PhDs graduating in 2022, 65.1% of

N O R T H A M E R I CA them were international students. In addition, more
The share of international students among new AI PhDs international students—14.0% of all new AI PhDs—took
in North America in 2020 decreased slightly from 64.3% jobs outside the United States in 2020, compared to 8.6%
in 2019 to 60.5% in 2020 (Figure 4.4.9). For comparison, in 2019 (Figure 4.4.10).
NEW INTERNATIONAL AI PHDS (% of TOTAL NEW AI PHDS) in NORTH AMERICA, 2010–20

60%
New International AI PhDs (% of Total New AI PhDs)
60.50%
50%
40%
30%
20%
10%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
Figure 4.4.9
INTERNATIONAL NEW AI PHDS (% of TOTAL) in the UNITED

STATES by LOCATION of EMPLOYMENT, 2020
11.80%, 14.00%,
Unknown Outside United States
74.20%,
United States
Figure 4.4.10

Index Report 2022
CHAPTER 5:
AI Policy and
Governance
Index Report 2022 CHAPTER 5: AI POLICY AND GOVERNANCE
CHAPTER 5:
Chapter Preview
Overview 174 U.S. AI Policy Papers 186
Chapter Highlights 175 By Topic 187
5.1 AI AND POLICYMAKING 176 5.2 U.S. PUBLIC INVESTMENT IN AI 188

Global Legislation Records on AI 176 Federal Budget for Nondefense AI R&D 188
By Geographic Area 177 U.S. Department of Defense
Federal AI Legislation in the Budget Request 189
United States 178 Highlight: DOD Top Five
Highlight: A Closer Look Highest-Funded Programs 190
at the Legislation 179 DOD AI R&D Spending by Department 191
State-Level AI Legislation U.S. Government AI-Related
in the United States 180 Contract Spending 192
By State 181 Total Contract Spending 192
Sponsorship by Political Party 182 Contract Spending by Department
Mentions of AI in Legislative Records 183 and Agency 193
AI Mentions in U.S. Congressional Highlight: Largest Contract for Five

Records 183 Top-Spending Departments in 2021 195
AI Mentions in Global Legislative

Proceedings 184
By Geographic Area 185

Overview
As AI has become an increasingly ubiquitous topic in the last decade,
intergovernmental, national, and regional organizations have worked to
develop policies and strategies around AI governance. These actors are
driven by the understanding that it is imperative to find ways to address
the ethical and societal concerns surrounding AI, while maximizing
its benefits. Active and informed governance of AI technologies has
become a priority for many governments around the world.
This chapter examines the intersection of AI and governance, and takes

a closer look at how governments in different countries, regions, and
U.S. states are working to manage AI technologies. It begins by looking
at AI policymaking across the globe and within the United States,
exploring which countries and political actors are most keen to advance
AI legislation, and what kind of AI subtopics, from privacy to ethics, are
the focus of most legislative attention. Then the chapter takes a deep
dive into one of the world’s top public sector investors in AI, the United
States, and studies how much its various government departments have
spent on AI in the past five years.

CHAPTER HIGHLIGHTS
• An AI Index analysis of legislative records on AI in 25 countries shows that the number of bills
containing “artificial intelligence” that were passed into law grew from just 1 in 2016 to 18 in
2021. Spain, the United Kingdom, and the United States passed the highest number of AI-related
bills in 2021, with each adopting three.
• The federal legislative record in the United States shows a sharp increase in the total number of
proposed bills that relate to AI from 2015 to 2021, while the number of bills passed remains low,
with only 2% ultimately becoming law.
• State legislators in the United States passed 1 out of every 50 proposed bills that contain AI
provisions in 2021, while the number of such bills proposed grew from 2 in 2012 to 131 in 2021.
• In the United States, the current congressional session (the 117th) is on track to record the greatest
number of AI-related mentions since 2001, with 295 mentions by the end of 2021, half way
through the session, compared to 506 in the previous (116th) session.

Artificial Intelligence CHAPTER 5: AI POLICY AND GOVERNANCE
Index Report 2022 5.1 AI and Policymaking
Discussions around AI governance regulation have accelerated over the past decade, resulting in policy proposals across various
legislative bodies. This section first examines AI-related legislation that has either been proposed or passed into law across different
countries and regions, followed by a focused analysis of state-level legislation in the United States. It then takes a closer look at
congressional and parliamentary records on AI across the world and concludes with data on the number of policy papers published
in the United States.
5.1 AI AND POLICYMAKING

G LO BA L L E G I S L AT I O N their legislative bodies that contain the words “artificial
RECORDS ON AI intelligence” from 2016 to 2021.

Governments and legislative bodies across the globe are Taken together, the 25 countries analyzed have passed a
increasingly seeking to pass laws to provide funding for total of 55 AI-related bills. Figure 5.2.1 demonstrates that in
AI development and innovation, while also promoting the the past six years, there has been a sharp increase in terms
integration of human-centered values. The AI Index has of the total number of AI-related bills passed into law.1
conducted an analysis of laws passed in 25 countries by
NUMBER of AI-RELATED BILLS PASSED into LAW in 25 SELECT COUNTRIES, 2016–21

18
15
Number of AI-Related Bills
10
0
2016 2017 2018 2019 2020 2021
Figure 5.1.1
1 Note that the analysis only includes laws passed by national legislative bodies (e.g. congress, parliament) with the keyword “artificial intelligence” in various languages in the title or body of the bill
text. See the appendix for the methodology. Countries included: Australia, Belgium, Brazil, Canada, China, Denmark, Finland, France, Germany, India, Ireland, Italy, Japan, the Netherlands, New Zealand,
Norway, Russia, Singapore, South Africa, South Korea, Spain, Sweden, Switzerland, the United Kingdom, and the United States.

By Geographic Area
Figure 5.1.2a shows the number of laws containing The United States
mentions of AI that were enacted in 2021. Spain, the
dominated the list with 13
United Kingdom, and the United States led, each passing
three. Figure 5.1.2b shows the total number of legislation bills, starting in 2017 with
passed in the past six years. The United States dominated
the list with 13 bills, starting in 2017 with 3 new laws
3 new laws passed each
passed each subsequent year, followed by Russia, subsequent year, followed
Belgium, Spain, and the United Kingdom.
by Russia, Belgium, Spain,
and the United Kingdom.
NUMBER of AI-RELATED BILLS PASSED into LAW in SELECT COUNTRIES, 2021

Spain 3
United Kingdom 3
United States 3
Belgium 2
Russia 2
France 1
Germany 1
Italy 1
Japan 1
South Korea 1
0 1 2 3 4
Figure 5.1.2a
Table of Contents Chapter 5 Preview 17 7

NUMBER of AI-RELATED BILLS PASSED into LAW in SELECT COUNTRIES, 2016–21 (SUM)
United States 13
Russia 6
Belgium 5
Spain 5
United Kingdom 5
France 4
Italy 4
South Korea 4
Japan 3
China 2
Brazil 1
Canada 1
Germany 1
India 1
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
Figure 5.1.2b
Federal AI Legislation in the United States were 130. Although this jump is significant, the number
A closer look at the federal legislative record in the of bills related to AI being passed has not kept pace with
United States shows a sharp increase in the total number the growing volume of proposed AI-related bills. This gap
of proposed bills that relate to AI (Figure 5.1.3). In 2015, was most evident in 2021, when only 2% of all federal-
just one federal bill was proposed, while in 2021, there level AI-related bills were ultimately passed into law.
NUMBER of AI-RELATED BILLS in the UNITED STATES, 2015–21 (PROPOSED vs. PASSED)
130, Proposed
120
100
80
60
40
20
3, Passed
2015 2016 2017 2018 2019 2020 2021

Figure 5.1.3

A Closer Look at the Legislation

The following subsection delves into some of the AI-related legislation passed into law since 2016.
Table 5.1.1 demonstrates the wide range of AI-related issues that have piqued policymakers’ interest.
Country Year Passed Bill Name Description
Canada 2017 Budget Implementation Act 2017, No. 1 A provision of this act authorized the Canadian
government to make a payment of $125 million
to the Canadian Institute for Advanced Research
to support the development of a pan-Canadian
artificial intelligence strategy.
China 2019 Law of the People’s Republic of China A provision of this law aimed to promote the
on Basic Medical and Health Care and application and development of big data and
the Promotion of Health artificial intelligence in the health and medical field
while accelerating the construction of medical and
healthcare information infrastructure, developing
technical standards on the collection, storage,
analysis, and application of medical and health data.
Russia 2020 Federal Law of 24 April 2020 No. This law established an experimental framework
123-FZ on the Experiment to Establish for the development and implementation of AI as
Special Regulation in order to Create a five-year experiment to start in Moscow in July
the Necessary Conditions for the 1, 2020, including allowing AI systems to process
Development and Implementation of anonymized personal data for governmental and
Artificial Intelligence Technologies in certain commercial business activities.
the Region of the Russian Federation
– Federal City of Moscow and
Amending the Articles 6 and 10 of the
Federal Law on Personal Data
United Kingdom 2020 Supply and Appropriation (Main A provision of this act authorized the Office of
Estimates) Act 2020, c.13 Qualifications and Examination Regulation to
explore opportunities for using artificial intelligence
to improve the marking and administration of high-
stakes qualifications.
United States 2020 IOGAN ACT: Identifying Outputs of This act directed the National Science Foundation
Generative Adversarial Networks Act to support research dedicated to studying the
outputs of generative adversarial networks
(deepfakes) and other comparable technologies.
Belgium 2021 Decree on coaching and solution- A provision of this act directs the government
oriented support for job seekers, N. to create an advisory group called the Ethics
327 Committee, which is responsible for submitting
advice if artificial intelligence tools are to be used
for digitization activities.
France 2021 Law N:2021-1485 of November This act sets up a monitoring system to evaluate
15, 2021, aimed at reducing the environmental impacts of newly emerging digital
environmental footprint of digital technologies, in particular, artificial intelligence.
technology in France
Table 5.1.1

S TAT E - L E V E L A I L E G I S L AT I O N I N a greater proportion of proposed state-level AI bills have

T H E U N I T E D S TAT E S actually passed. In 2021, of the 131 proposed state bills,
Growing policy interest in AI can also be seen in the 26 were passed into law (20%), or 1 out of 5 proposed
large number of AI-related bills recently proposed bills became law. This ratio is significantly higher when
at the state level in the United States, based on data compared to the federal level, where 1 out of every 50
provided by Bloomberg Government since 2012. proposed bills became law in 2021.
Bloomberg Government classified a bill as relating to
AI if it contained AI-related keywords such as artificial
intelligence, machine learning, or algorithmic bias. A notable difference between
As is the case on the federal level, there has been a AI-related lawmaking in the
significant increase in the number of AI bills proposed
at the state level in the last decade (Figure 5.1.4). United States on the federal
In 2012, the first two pieces of AI-related legislation
versus the state level is
were proposed when New Jersey assembly member
Annette Quijano directed the New Jersey Motor Vehicle that a greater proportion of
Commission to establish driver’s license endorsements
for autonomous vehicles. In the past 10 years, the
proposed state-level AI bills
increase has been substantial, from 2 bills in 2012 to 131 have actually passed.
in 2021.
A notable difference between AI-related lawmaking in the

United States on the federal versus the state level is that
NUMBER of STATE-LEVEL AI-RELATED BILLS in the UNITED STATES, 2012–21

Source: Bloomberg Government, 2021 | Chart: 2022 AI Index Report
140
131
Passed
120 Proposed
26
Vetoed
100
87
80 13 77
10
60
103
40
74
29 66
26
20 9
14
10 9 9 25
17
2 8 12
0
2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 5.1.4

By State NUMBER of STATE-LEVEL PROPOSED AI-RELATED BILLS in the UNITED

In the United States, AI STATES by STATE, 2012–21 (SUM)
lawmaking has been relatively
widespread across all states. As AK ME
of 2021, 41 out of 50 states have 0 0
proposed at least one AI-related VT NH

8 0
bill, but certain states have been
WA ID MT ND MN IL WI MI NY RI MA
particularly active in generating 14 0 0 0 2 28 0 3 31 5 40
AI legislation. Figure 5.1.5 shows OR NV WY SD IA IN OH PA NJ CT
that Massachusetts has proposed 1 10 0 0 1 1 3 7 32 2
the most AI bills, with 40 since CA UT CO NE MO KY WV VA MD DE

29 3 2 2 4 1 9 8 8 1
2012, followed by Hawaii (35)
AZ NM KS AR TN NC SC DC
and New Jersey (32). Focusing 7 1 1 1 7 6 1 8
on just 2021 in Figure 5.1.6,
OK LA MS AL GA
Massachusetts was the state that 2 0 7 21 3
proposed the most AI-related HI TX FL
bills, with 20, followed by Illinois 35 17 22
(15) and Alabama (12).

Figure 5.1.5
NUMBER of STATE-LEVEL PROPOSED AI-RELATED BILLS in the UNITED

STATES by STATE, 2021
AK ME
0 0
VT NH
3 0
WA ID MT ND MN IL WI MI NY RI MA
6 0 0 0 0 15 0 0 8 3 20
OR NV WY SD IA IN OH PA NJ CT
1 1 0 0 1 0 1 3 4 0
CA UT CO NE MO KY WV VA MD DE
4 2 2 0 1 0 3 1 4 0
AZ NM KS AR TN NC SC DC
0 0 0 0 2 3 1 6
OK LA MS AL GA
1 0 3 12 0
HI TX FL
7 6 7
Figure 5.1.6

Sponsorship by Political Party of both parties since 2012, in the past four years, the
State-level AI legislation data reveals that there is a data suggests Democrats were more likely to sponsor
partisan dynamic to AI lawmaking. Figure 5.1.7 plots the AI-related legislation. Whereas Democrats sponsored
number of AI-related bills sponsored at the state level by only two more AI bills than Republicans in 2018, they
Democratic and Republican lawmakers. Although there sponsored 39 more in 2021.
has been an increase in AI bills proposed by members
NUMBER of STATE-LEVEL PROPOSED AI-RELATED BILLS in the UNITED STATES by SPONSOR PARTY, 2012–21
79, Democratic
80
70
60
50
40, Republican
40
30
20
10
0
2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Figure 5.1.7

M E N T I O N S O F A I I N L E G I S L AT I V E from Bloomberg Government concerning mentions of AI-

RECORDS related keywords in congressional proceedings, broken
Another barometer of legislative interest in AI is the down by legislation, congressional committee reports,
number of mentions of “artificial intelligence” in and congressional research service reports.
governmental and parliamentary proceedings. This According to Figure 5.1.8, the current congressional
subsection considers data on mentions of AI both session (the 117th) is on track (as of the end of 2021)
in U.S. congressional records and the parliamentary to record the greatest number of AI-related mentions
proceedings of other countries based on AI Index and since 2001. The most recently completed congressional
Bloomberg Government data. session, the 116th (2019-2020), saw 506 AI mentions,
AI Mentions in U.S. Congressional Records nearly 3.4 times as many mentions as there were during
In the last five years, and especially in 2021, U.S. the 115th session (2017–2018), and 30 times as many as
congressional sessions have devoted increasing amounts the 114th session (2015–2016).
of time to discussions of AI. This section presents data
MENTIONS of AI in the U.S. CONGRESSIONAL RECORD by LEGISLATIVE SESSION, 2001–21

Legislation 506
500 Congressional Research Service Reports
Committee Reports
178
400
Number of Mentions
300 295
139
200 245
149
44
100 129
66
25 83
17 18 17 12 17
0 4 7 39 27
107th 108th 109th 110th 111th 112th 113th 114th 115th 116th 117th
(2001–02) (2003–04) (2005–06) (2007–08) (2009–10) (2011–12) (2013–14) (2015–16) (2017–18) (2019–20) (2021–)
Figure 5.1.8

AI Mentions in Global Legislative Proceedings sessions in 25 countries that contain the keyword
AI mentions in governmental proceedings are on the “artificial intelligence” from 2016 to 2021. Figure 5.1.9
rise not only in the United States but also in many other shows that the mentions of AI in legislative proceedings
countries across the world. The AI Index conducted an in 25 select countries grew 7.7 times in the past six years.2
analysis on the minutes or proceedings of legislative
NUMBER of MENTIONS of AI in LEGISLATIVE PROCEEDINGS in 25 SELECT COUNTRIES, 2016–21

1,323
1,200
1,000
Number of Mentions
800
600
400
200
0
2016 2017 2018 2019 2020 2021
Figure 5.1.9
2 See the appendix for the methodology. Countries included: Australia, Belgium, Brazil, Canada, China, Denmark, Finland, France, Germany, India, Ireland, Italy, Japan, the Netherlands, New Zealand,
Norway, Russia, Singapore, South Africa, South Korea, Spain, Sweden, Switzerland, the United Kingdom, and the United States.

By Geographic Area Kingdom, and the United States topped the list. Figure
Figure 5.1.10a shows the number of legislative 5.1.2b shows the total number of AI mentions in the past
proceedings containing mentions of AI that were six years. The United Kingdom dominated the list with
enacted in 2021. Similar to the trend in the number of 939 mentions, followed by Spain, Japan, the United
AI mentions in bills passed into laws, Spain, the United States, and Australia.
NUMBER of MENTIONS of AI in LEGISLATIVE PROCEEDINGS in NUMBER of MENTIONS of AI in LEGISLATIVE PROCEEDINGS in

SELECT COUNTRIES, 2021 SELECT COUNTRIES, 2016–2021 (SUM)
Source: AI Index, 2021 | Chart: 2022 AI Index Report Source: AI Index, 2021 | Chart: 2022 AI Index Report
Spain 269 United Kingdom 939

United Kingdom 185 Spain 559
United States 132 Japan 466
Australia 122 United States 422
Japan 95 Australia 410
Ireland 76 Singapore 282
Brazil 72 Ireland 222
Italy 64 Italy 164
Singapore 60 Germany 158
Belgium 47 France 155
Germany 46 Brazil 123
France 25 Belgium 120
Canada 22 Canada 111
Norway 20 Finland 78
Sweden 16 Sweden 71
Finland 12 Netherlands 67
Russia 12 Russia 58
South Africa 11 Norway 52
Netherlands 10 India 34
India 7 South Africa 33
New Zealand 6 Denmark 27
South Korea 6 New Zealand 27
Denmark 5 South Korea 21
Switzerland 3 Switzerland 15
0 50 100 150 200 250 300 0 200 400 600 800 1000
Number of Mentions Number of Mentions

U.S. A I P O L I CY PA P E R S specific recommendations to policymakers. Topics of

To estimate activities outside national governments those papers are divided into primary and secondary
that are also informing AI-related rulemaking, the categories: A primary topic is the main focus of the paper,
AI Index tracks 55 U.S.-based organizations that while a secondary topic is a subtopic of the paper or an
published policy papers in the past four years. Those issue that was briefly explored.
organizations include: think tanks and policy institutes Figure 5.1.11 plots the total number of U.S.-based AI-
(19); university institutes and research programs (14); related policy papers that have been published from
civil society organizations, associations, and consortiums 2018 to 2021, which can proxy the general interest in AI
(9); industry and consultancy organizations (9); and within the U.S. policymaking space. The total number of
government agencies (4).3 A policy paper in this section policy papers has tripled since 2018, peaking in 2020 with
is defined as a research paper, research report, brief, or 273, and decreasing slightly in 2021, with 210.
blog post that addresses issues related to AI and makes
NUMBER of AI-RELATED POLICY PAPERS by U.S.-BASED ORGANIZATIONS, 2018–21

250
210
200
Number of Policy Papers
150
100
50
0
2018 2019 2020 2021
Figure 5.1.11
3 The complete list of organizations the Index followed can be found in the Appendix.

By Topic topics, but they were reported on more frequently

In 2021, the leading primary topics were Privacy, Safety, as secondary topics. Among the AI topics to receive
and Security; Innovation and Technology; and Ethics comparatively little attention from tracked organizations
(Figure 5.1.12). Certain topics, such as government and are those that relate to energy and the environment,
public administration, education and skills, as well as humanities, physical sciences, and social and behavioral
democracy, did not feature prominently as primary sciences.
NUMBER of AI-RELATED POLICY PAPERS by U.S.-BASED ORGANIZATIONS by TOPIC, 2021

Primary Topic Secondary Topic

Privacy, Safety, and Security 62 45
Innovation and Technology 62 63
Ethics 59 57
Int'l A"airs and Int'l Security 51 45
Industry and Regulation 51 58
Equity and Inclusion 36 51
Workforce and Labor 34 30
Gov't and Public Administration 33 58
Justice and Law Enforcement 29 17
Education and Skills 23 36
Communications and Media 15 17
Health and Biological Sciences 7 13
Social and Behavioral Sciences 4 1
Democracy 2 29
Physical Sciences 1 3
Energy and Environment 1 3
Humanities 0 1
0 20 40 60 0 20 40 60
Number of Policy Papers
Figure 5.1.12

Index Report 2022 5.2 U.S. Public Investment in AI
This section examines the public AI investment in the United States, based on data from the U.S. government and Bloomberg Government.
5.2 U.S. PUBLIC INVESTMENT IN AI

FEDERAL BUDGET FOR NONDEFENSE
A I R& D
The increasing amount
In December 2021, the National Science and Technology Council spent on AI R&D by
published a report on the public-sector AI R&D budget across
departments and agencies participating in the Networking and nondefense departments
Information Technology Research and Development (NITRD) indicates the U.S.
program and the National Artificial Intelligence Initiative. The report
does not include information on classified AI R&D investment by the government’s continued
defense and intelligence agencies.
strong interest in
In fiscal year (FY) 2021, nondefense U.S. government agencies
allocated a total of $1.53 billion to AI R&D spending, approximately
public sector funding
2.7 times what was spent in FY 2018 (Figure 5.2.1). This figure for AI research and
is projected to rise 8.8% for FY 2022, with a total of $1.67 billion
requested.4 The increasing amount spent on AI R&D by nondefense
development spanning
departments indicates the U.S. government’s continued strong a wide range of federal
interest in public sector funding for AI research and development
spanning a wide range of federal agencies. agencies.
U.S. FEDERAL BUDGET for NONDEFENSE AI R&D, FY 2018–22
Source: U.S. NITRD Program, 2022 | Chart: 2022 AI Index Report
1.67
1.53
1.50 1.43
Budget (in billions of U.S. Dollars)
1.11
1.00
0.56
0.50
0.00
FY18 (ENACTED) FY19 (ENACTED) FY20 (ENACTED) FY21 (ENACTED) FY22 (REQUESTED)
Figure 5.2.1
4 See NITRD website for details on AI R&D investment FY 2018-22 with the breakdown of core AI vs AI crosscut. Note that AI crosscutting budget data is not available for FY 2018.

U.S. D E PA R T M E N T O F D E F E N S E DOD allocated $9.26 billion across 500 AI R&D programs

BUDGET REQUEST (Figure 5.2.2), a 6.68% increase from the $8.68 billion
Spending on AI by the U.S. Department of Defense (DOD) spent in 2020. For FY 2022, the department has requested
can be proxied by looking at the publicly available $10 billion so far, which is likely to grow once additional
requests made by the DOD for research, development, requests and congressional appropriations are taken into
test, and evaluation (RDT&E) relating to AI. In FY 2021, account.
U.S. DOD BUDGET for AI-SPECIFIC RESEARCH, DEVELOPMENT, TEST and EVALUATION (RDT&E), FY 2020–22
Source: Bloomberg Government and U.S. Department of Defense, 2021 | Chart: 2022 AI Index Report
10.00
10
9.26
8.68
8
0.93: DOD Reported Budget on AI R&D 0.84: DOD Reported Budget on AI R&D 0.87: DOD Reported Budget on AI R&D
0
Sum of FY20 Funding Sum of FY21 Funding Sum of FY22 Funding
Figure 5.2.2
Important data caveat: This chart is indicative of one may result from the difference in defining AI-related
of the challenges of quantifying public AI spending. budget items. For example, a research project that uses
Bloomberg Government’s analysis that searches AI- AI for cyber defense may count human, hardware, and
relevant keywords in DOD budgets shows that the operations-related expenditures within the AI-related
department is requesting $10.0 billion for AI-specific R&D budget request, though the AI software component will
in FY 2022. However, DOD’s own measurement produces be much smaller.
a smaller number of $874 million. The discrepancy

DOD Top Five Highest-Funded Programs

This section highlight offers a more qualitative look at some of the AI-related research projects the
DOD prioritizes. Table 5.2.1 presents the five DOD-related AI programs that received the greatest
funding in 2021. In the past year, the DOD was interested in deploying AI for a number of purposes,
from geospatial monitoring to reducing the threat posed by weapons of mass destruction.
Funds Received
Program Name Department Purpose
(in millions)
1 Rapid Capability Army 257 Fund the development, engineering, acquisition,
Development and Maturation and operation of various AI-related technological
prototypes that could be used for military purposes.
2 Counter Weapons of Defense Threat 254 Develop technologies that could “deny, defeat and
Mass Destruction Advanced Reduction disrupt” weapons of mass destruction (WMD).
Technology Development Agency
3 Algorithmic Warfare Office of the 230 Accelerate the integration of AI technologies in DOD
Cross-Functional Teams – Secretary of systems to “improve warfighting speed and lethality.”
Software Pilot Program Defense
4 Joint Artificial Intelligence Defense 137 Develop, test, prototype, and demonstrate various AI
Center Information and machine learning capabilities with the intention
Systems of integrating these capabilities across numerous
Agency domains which include “supply chain, personal
recovery, infrastructure assessment, geospatial
monitoring during disaster and cyber sense making.”
5 High Performance Army 96 Investigate, demonstrate, and mature both general

Computing Modernization and special-purpose supercomputing environments
Program that are used to satisfy wide-ranging DOD priorities.
Table 5.2.1

DOD AI R&D Spending by Department is poised to maintain that position in 2022. They have
DOD spending on AI R&D can also be broken down on requested a total of $1.86 billion in FY 2022 for AI-related
a subdepartmental level, which reveals how individual projects, followed by the Army ($1.77 billion), the Office
defense agencies—the Army and the Navy, for instance— of the Secretary of Defense ($1.1 billion) and the Air Force
compare in their AI spending (Figure 5.2.3). The U.S. ($883 million).
Navy was the top-spending DOD agency in FY 2021 and
U.S. DOD BUDGET for AI-SPECIFIC RESEARCH, DEVELOPMENT, TEST and EVALUATION (RDT&E) by
DEPARTMENT, FY 2020–22
Air Force Army DARPA DISA Navy OSD Other
10
1.75
1.64
1.18
8
1.77
1.52 1.63
6 1.13
1.12 1.19
1.86
4 1.72 1.93
1.71
2 1.92 1.54
1.57
1.00 1.16
0
FY20 (ENACTED) FY21 (ENACTED) FY22 (REQUESTED)
Figure 5.2.3

U.S. G OV E R N M E N T A I - R E L AT E D of more than 100 AI-specific keywords in their titles or

C O N T R ACT S P E N D I N G descriptions.5
Public investment in AI can also be measured by federal Total Contract Spending
government spending on AI-related contracts. U.S. In 2021, federal departments and agencies spent a total of
government agencies often award contracts to private $1.79 billion on AI-related contracts. Although this amount
companies for the supply of various goods and services is nearly double what was spent on AI-related contracts in
that typically occupy the largest share of an agency’s 2018 (roughly $920 million), it represents a slight decrease
budget. Bloomberg Government built a model to classify from the amount spent on AI-related contracts in 2020,
whether a U.S. government contract was AI-related by which peaked at $1.97 billion (Figure 5.2.4).
adding up all contracting transactions that contain a set
U.S. GOVERNMENT TOTAL CONTRACT SPENDING on AI, FY 2000–21

2.0
1.79
Contract Spending (in billions of U.S. Dollars)
1.5
1.0
0.5
0.0
2000
2006
2009
2004
2008
2003
2005
2002
2020
2007
2001
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2021
Figure 5.2.4
5 Note that contractors may add a number of keywords into their applications during the procurement process, so some of the projects included may have a relatively small AI component relative to
other parts of technology.

Contract Spending by Department and Agency Aggregate spending on AI contracts in the last four years
Figures 5.2.5 and 5.2.6 report AI-related contract spending tells a similar story. Since 2018, the DOD has spent $5.20
by the top 10 federal agencies in 2021 and from 2000 to billion on AI contracts, approximately seven times the next
2021, respectively. The DOD outspent the rest of the U.S. highest spender, NASA ($1.41 billion). In fact, since 2018,
government on both charts by a significant margin. In the DOD has spent twice as much on AI-related contracts
2021, it spent $1.14 billion on AI-related contracts, roughly as all other government agencies combined. Following the
five times what was spent by the next highest department, DOD and NASA are the Department of Health and Human
the Department of Health and Human Services ($234 Services ($700 million), the Department of Homeland
million). Security ($362 million), and Department of the Treasury
($156 million).
TOP CONTRACT SPENDING on AI by U.S. GOVERNMENT DEPARTMENT and AGENCY, 2021

Department of Defense (DOD) 1,138

Department of Health and Human Services (HHS) 234
National Aeronautics and Space Administration (NASA) 159
Department of Homeland Security (DHS) 81
Department of Commerce (DOC) 49
Department of the Treasury (TREAS) 38
Department of Veterans A"airs (VA) 25
Department of Transportation (DOT) 12
Securities and Exchange Commission (SEC) 12
Department of Agriculture (USDA) 12
Department of Energy (DOE) 8
Agency for International Development (USAID) 6
Department of Justice (DOJ) 4
Department of State (DOS) 3
National Science Foundation (NSF) 2
0 200 400 600 800 1000 1200
Contract Spending (in millions of U.S. Dollars)
Figure 5.2.5

TOP CONTRACT SPENDING on AI by U.S. GOVERNMENT DEPARTMENT and AGENCY, 2000–21 (SUM)
Department of Defense (DOD) 5.20
National Aeronautics and Space Administration (NASA) 1.41
Department of Health and Human Services (HHS) 0.70
Department of Homeland Security (DHS) 0.45
Department of the Treasury (TREAS) 0.32
Department of Veterans A!airs (VA) 0.15
Department of Commerce (DOC) 0.15
Department of Agriculture (USDA) 0.07
Department of Energy (DOE) 0.06
Securities and Exchange Commission (SEC) 0.06
General Services Administration (GSA) 0.06
Department of State (DOS) 0.06
Social Security Administration (SSA) 0.05
Department of Transportation (DOT) 0.05

0 1 2 3 4 5
Contract Spending (in billions of U.S. Dollars)
Figure 5.2.6

Largest Contract for Five Top-Spending

Departments in 2021
To paint a better picture of how different U.S. government departments use AI, Table 5.2.2 shows the
most expensive AI-related contract that the five highest AI-related-spending departments signed in
2021. Last year, the U.S. government invested in AI to build autonomous vehicle prototypes, develop an
AI imaging system that could assist with burn classification, and create robots capable of higher-level
lunar navigation.
Amount
Contract Name Department Purpose
(in millions)
Prototype Services in the Objective DOD 70 To acquire prototypes in the domain of
Areas of Automotive Cybersecurity, automotive cybersecurity, vehicle safety
Vehicle Safety Technologies, Vehicle technologies, and autonomous vehicles and
Light Weighting, Autonomous Vehicles intelligent systems.
and Intelligent Systems, Connected
Vehicles, and Advanced Energy Storage
Technologies
Biomedical Advanced Research and HHS 20 To develop optical imaging devices and
Development Authority (BARDA) machine learning algorithms to assist
in classifying and healing wounds and
conventional burns.
Commercial Lunar Payload Services NASA 14 To develop lunar robots capable of navigating
the moon’s south pole to acquire lunar
resources and engage in lunar-based scientific
activities.
SBIR-Autonomous Surveillance DHS 37 To construct towers capable of autonomous

Towers-Delivery Order surveillance.
Schedule 70: Information Technology DOC 13 To develop a prototype using AI technology

that can improve patent search.
Table 5.2.2

Index Report 2022
Appendix
Index Report 2022
Appendix
CHAPTER 1 Research & Development 198
CHAPTER 2 Technical Performance 200
CHAPTER 3 Technical AI Ethics 212
CHAPTER 4 The Economy and Education 216
CHAPTER 5 AI Policy and Governance 222
197
Artificial Intelligence APPENDIX
Index Report 2022 Chapter 1: Research & Development
CHAPTER 1: RESEARCH & DEVELOPMENT

CENTER FOR SECURITY AND published research field(s) of study and corresponding
E M E R G I N G T E C H N O LO GY, scores.3 CSET researchers identified the most common
fields of study in their corpus of AI-relevant publications
G E O R G E TOW N U N I V E R S I T Y
since 2010 and recorded publications in all other fields as
Prepared by Sara Abdulla and James Dunham
“Other AI.” English-language AI-relevant publications were
The Center for Security and Emerging Technology (CSET) then tallied by their top-scoring field and publication year.
is a policy research organization within Georgetown
University’s Walsh School of Foreign Service that produces CSET also provided year-by-year citations for AI-relevant
data-driven research at the intersection of security and work associated with each country. A publication is
technology, providing nonpartisan analysis to the policy associated with a country if it has at least one author whose
community. organizational affiliation(s) is located in that country.
Citation counts aren’t available for all publications; those
Publications from CSET Merged Corpus of without counts weren’t included in the citation analysis.
Scholarly Literature Over 70% of English-language AI papers published between
Source 2010 and 2020 have citation data available.
CSET’s merged corpus of scholarly literature combines
distinct publications from Digital Science’s Dimensions, CSET counted cross-country collaborations as distinct
Clarivate’s Web of Science, Microsoft Academic Graph, pairs of countries across authors for each publication.
China National Knowledge Infrastructure, arXiv, and Collaborations are only counted once: For example, if a
Papers With Code.1 publication has two authors from the United States and
two authors from China, it is counted as a single United
Methodology States-China collaboration.
To create the merged corpus, CSET deduplicated across
the listed sources using publication metadata, and then Additionally, publication counts by year and by
combined the metadata for linked publications. To publication type (e.g., academic journal articles,
identify AI publications, CSET used an English-language conference papers) were provided where available.
subset of this corpus: publications since 2010 that appear These publication types were disaggregated by affiliation
AI-relevant.2 CSET researchers developed a classifier country as described above.
for identifying AI-related publications by leveraging the
arXiv repository, where authors and editors tag papers by CSET also provided publication affiliation sector(s)
subject. where, as in the country attribution analysis, sectors were
associated with publications through authors’ affiliations.
To provide a publication’s field of study, CSET matches Not all affiliations were characterized in terms of sectors;
each publication in the analytic corpus with predictions CSET researchers relied primarily on GRID from Digital
from Microsoft Academic Graph (MAG)’s field-of-study Science for this purpose, and not all organizations can
model, which yields hierarchical labels describing the be found in or linked to GRID.4 Where the affiliation
1 All CNKI content is furnished for CSET by East View Information Services, Minneapolis, MN, USA.
2 For more information, see James Dunham, Jennifer Melot, and Dewey Murdick, “Identifying the Development and Application of Artificial Intelligence in Scientific Text,” arXiv [cs.DL], May 28, 2020,
https://arxiv.org/abs/2002.07143.
3 These scores are based on cosine similarities between field-of-study and paper embeddings. See Zhihong Shen, Hao Ma, and Kuansan Wang, “A Web-Scale System for Scientific Knowledge
Exploration,” arXiv [cs.CL], May 30, 2018, https://arxiv.org/abs/1805.12216.
4 See https://www.grid.ac/ for more information about the GRID dataset from Digital Science.
Table of Contents Appendix 198

Index Report 2022 Chapter 1: Research & Development
sector is available, papers were counted toward these iperov/DeepFaceLab, Microsoft/CNTK, opencv/opencv,
sectors, by year. Cross-sector collaborations on academic pytorch/pytorch, scikit-learn/scikit-learn, scutan90/
publications were calculated using the same method as in DeepLearning-500-questions, tensorflow/tensorflow,
the cross-country collaborations analysis. Theano/Theano, Torch/Torch7.
Patents from CSET’s AI Patents Dataset Nuance

Source The GitHub Archive currently does not provide a way
CSET’s AI patents dataset was developed by CSET to count when users remove a star from a repository.
and 1790 Analytics. It includes patents relevant to the Therefore, the reported data slightly overestimates the
development and application of AI, as indicated by CPC/ number of stars. A comparison with the actual number
IPC codes and keywords. of stars for the repositories on GitHub reveals that the
numbers are fairly close and that the trends remain
Methodology unchanged.
In this analysis, patents were grouped by year and
country, and then counted at the “patent family”
level.5 CSET extracted year values from the most recent
publication date within a family. This method has the
advantage of capturing updates within a patent family
(such as amendments).
The country of origin for a patent is derived from the first

country in which a patent was filed.6
G I T H U B S TA R S
Source
GitHub: star-history (available at star history website) was
used to retrieve the data.
Methodology
The visual in the report shows the number of stars for
various GitHub repositories over time. The repositories
include the following:
apachecn/ailearning, apache/incubator-mxnet, Avik-
Jain/100-Days-Of-ML-Code, aymericdamien/TensorFlow-
Examples, BVLC/cafe, cafe2/cafe2, CorentinJ/Real-Time-
Voice-Cloning, deepfakes / faceswap, dmlc/mxnet,
exacity/deeplearningbook-chinese, fchollet/keras,
floodsung/Deep-Learning-Papers-Reading-Roadmap,
5 Patents are analyzed at the “patent family” level rather than “patent document” level because patent families are a collective of patent documents all associated with a single invention and/or
innovation by the same inventors/assignees. Thus, counting at the “patent family” level mitigates artificial number inflation when there are multiple patent documents in a patent family or if a patent is
filed in multiple jurisdictions.
6 For more details on CSET’s approach and experimentation for assigning country values, see footnote 26 in “Patents and Artificial Intelligence: A Primer,” by Dewey Murdick and Patrick Thomas. (Center
for Security and Emerging Technology, September 2020), https://doi.org/10.51593/20200038.

Index Report 2022 Chapter 2: Technical Performance
CHAPTER 2: TECHNICAL PERFORMANCE

ImageNet Generalization
Data on ImageNet accuracy was retrieved through a Xception: Deep Learning with Depthwise Separable
detailed arXiv literature review cross-referenced by Convolutions
technical progress reported on Papers with Code. The
reported dates correspond to the year during which ImageNet: Top-5 Accuracy
a paper was first published to arXiv, and the reported To highlight progress on top-5 accuracy without the use of
results (top-1 or top-5 accuracy) correspond to the result extra training data, scores were taken from the following
reported in the most recent version of each paper. papers:
The estimate of human-level performance is from Adversarial Examples Improve Image Recognition
Russakovsky et al., 2015. Learn more about the LSVRC EfficientNet: Rethinking Model Scaling for Convolutional
ImageNet competition and the ImageNet data set. Neural Networks
Exploring the Limits of Weakly Supervised Pretraining
ImageNet: Top-1 Accuracy Fixing the Train-Test Resolution Discrepancy:
To highlight progress on top-1 accuracy without the use of FixEfficientNet
extra training data, scores were taken from the following GPipe: Efficient Training of Giant Neural Networks Using
papers: Pipeline Parallelism
Adversarial Examples Improve Image Recognition High-Performance Large-Scale Image Recognition Without
Billion-Scale Semi-Supervised Learning for Image Normalization
Classification ImageNet Classification with Deep Convolutional Neural
Dual Path Networks Networks
Densely Connected Convolutional Networks Learning Transferable Architectures for Scalable Image
EfficientNet: Rethinking Model Scaling for Convolutional Recognition
Neural Networks Squeeze-and-Excitation Networks
Fixing the Train-Test Resolution Discrepancy:
FixEfficientNet To highlight progress on top-5 accuracy with the use of
ImageNet Classification with Deep Convolutional Neural extra training data, scores were taken from the following
Networks papers:
Masked Autoencoders Are Scalable Vision Learners Big Transfer (BiT): General Visual Representation Learning
Deep Residual Learning for Image Recognition
To highlight progress on top-1 accuracy with the use of EfficientNet: Rethinking Model Scaling for Convolutional
extra training data, scores were taken from the following Neural Networks
papers: Florence: A New Foundation Model for Computer Vision
Big Transfer (BiT): General Visual Representation Learning Self-Training with Noisy Student Improves ImageNet
CoAtNet: Marrying Convolution and Attention for All Data Classification
Sizes Xception: Deep Learning with Depthwise Separable
EfficientNet: Rethinking Model Scaling for Convolutional Convolutions
Neural Networks
Self-Training with Noisy Student Improves ImageNet STL-10
Classification Data on STL-10 FID scores was retrieved through a
Sharpness-Aware Minimization for Efficiently Improving detailed arXiv literature review cross-referenced by

technical progress reported on Papers with Code. The in which a method was introduced, even if it was
reported dates correspond to the year during which subsequently tested, is the year in which it is included in
a paper was first published to arXiv, and the reported the report. The reported results (accuracy) correspond
results (FID score) correspond to the result reported in the to the result reported in the most recent version of each
most recent version of each paper. Details on the STL-10 paper. Details on the FaceForensics++ benchmark can be
benchmark can be found in the STL-10 paper. found in the FaceForensics++ paper.
To highlight progress on STL-10, scores were taken from To highlight progress on FaceForensics++, scores were
the following papers: taken from the following papers:
DEGAS: Differentiable Efficient Generator Search A Deep Learning Approach to Universal Image
Dist-GAN: An Improved GAN Using Distance Constraints Manipulation Detection Using a New Convolutional Layer
Off-Policy Reinforcement Learning for Efficient and Detection of Deepfake Videos Using Long Distance
Effective GAN Architecture Attention
Search Score Matching Model for Unbounded Data Score FakeCatcher: Detection of Synthetic Portrait Videos Using
Biological Signals
CIFAR-10 FaceForensics++: Learning to Detect Manipulated Facial
Data on CIFAR-10 FID scores was retrieved through a Images
detailed arXiv literature review cross-referenced by Learning Spatiotemporal Features with 3D Convolutional
technical progress reported on Papers with Code. The Networks
reported dates correspond to the year during which Recasting Residual-Based Local Descriptors as
a paper was first published to arXiv, and the reported Convolutional Neural Networks
results (FID score) correspond to the result reported in the Rich Models for Steganalysis of Digital Images
most recent version of each paper. Details on the CIFAR-10 Thinking in Frequency: Face Forgery Detection by Mining
benchmark can be found in the CIFAR-10 paper. Frequency-Aware Clues
Xception: Deep Learning with Depthwise Separable
To highlight progress on CIFAR-10, scores were taken from Convolutions
the following papers:
AutoGAN: Neural Architecture Search for Generative Celeb-DF
Adversarial Networks Data on Celeb-DF AUC was retrieved through a detailed
Denoising Diffusion Probabilistic Models arXiv literature review. The reported dates correspond
Improved Training of Wasserstein GANs to the year during which a paper was first published to
Large Scale GAN Training for High Fidelity Natural Image arXiv or a method was introduced. With Celeb-DF, recent
Synthesis researchers have tested previously existing deepfake
Score-Based Generative Modeling in Latent Space detection methodologies. The year in which a method was
introduced, even if it was subsequently tested, is the year
FaceForensics++ in which it is included in the report. The reported results
Data on FaceForensics++ accuracy was retrieved through (AUC) correspond to the result reported in the most recent
a detailed arXiv literature review. The reported dates version of each paper. Details on the Celeb-DF benchmark
correspond to the year during which a paper was first can be found in the Celeb-DF paper.
published to arXiv or a method was introduced. With
FaceForensics, recent researchers have tested previously To highlight progress on Celeb-DF, scores were taken from
existing deepfake detection methodologies. The year the following papers:
Exposing DeepFake Videos by Detecting Face Warping

Artifacts Human3.6M benchmark can be found in the Human3.6M

FaceForensics++: Learning to Detect Manipulated Facial paper.
Images
Face X-Ray for More General Face Forgery Detection To highlight progress on Human3.6M without the use of
Spatial-Phase Shallow Learning: Rethinking Face Forgery extra training data, scores were taken from the following
Detection in Frequency Domain papers:
3D Human Pose Estimation in Video with Temporal
Leeds Sports Poses Convolutions and Semi-Supervised Training
Data on Leeds Sports Poses percentage of correct Conditional Directed Graph Convolution for 3D Human
keypoints (PCK) was retrieved through a detailed Pose Estimation
arXiv literature review cross-referenced by technical Cross View Fusion for 3D Human Pose Estimation
progress reported on Papers with Code. The reported Epipolar Transformers
dates correspond to the year during which a paper was Human3.6M: Large Scale Datasets and Predictive Methods
first published to arXiv, and the reported results (PCK) for 3D Human Sensing in Natural Environments
correspond to the result reported in the most recent Learning 3D Human Pose from Structure and Motion
version of each paper. Details on the Leeds Sports Poses Robust Estimation of 3D Human Poses from a Single
benchmark can be found in the Leeds Sports Poses paper. Image
To highlight progress on Leeds Sports Poses, scores were To highlight progress on Human3.6M with the use of
taken from the following papers: extra training data, scores were taken from the following
Articulated Pose Estimation by a Graphical Model with papers:
Image Dependent Pairwise Relations Epipolar Transformers
Human Pose Estimation via Convolutional Part Heatmap Learnable Triangulation of Human Pose
Regression TesseTrack: End-to-End Learnable Multi-Person
Jointly Optimize Data Augmentation and Network Articulated 3D Pose Tracking
Training: Adversarial Data Augmentation in Human Pose
Estimation Cityscapes Challenge, Pixel-Level Semantic
Knowledge-Guided Deep Fractal Neural Networks for Labeling Task
Human Pose Estimation Data on the Cityscapes challenge, pixel-level semantic
OmniPose: A Multi-Scale Framework for Multi-Person Pose labeling task mean IoU was taken from the Cityscapes
Estimation dataset, more specifically their pixel-level semantic
Toward Fast and Accurate Human Pose Estimation via labeling leaderboard. More details about the Cityscapes
Soft-Gated Skip Connections dataset and other corresponding semantic segmentation
challenges can be accessed at the Cityscapes dataset
Human 3.6M webpage.
Data on Human3.6M average mean per joint position error
was retrieved through a detailed arXiv literature review CVC-ClinicDB and Kvasir-SEG
cross-referenced by technical progress reported on Papers Data on CVC-ClinicDB and Kvasir-SEG mean dice was
with Code. The reported dates correspond to the year retrieved through a detailed arXiv literature review cross-
during which a paper was first published to arXiv, and the referenced by technical progress reported on Papers
reported results (MPJPE) correspond to the result reported with Code (CVC-ClinicDB and Kvasir-SEG). The reported
in the most recent version of each paper. Details on the dates correspond to the year during which a paper was
first published to arXiv, and the reported results (mean

dice) correspond to the result reported in the most Visual Question Answering (VQA)
recent version of each paper. Details on the CVC-ClinicDB Data on VQA was taken from recent iterations of the VQA
benchmark can be found in the CVC-ClinicDB database challenge. To learn more about the VQA challenge in
page. Details on the Kvasir-SEG benchmark can be found general, please consult the following link. To learn more
in the Kvasir-SEG paper. about the 2021 iteration of the VQA challenge, please
consult the following link. More specifically, the Index
To highlight progress on CVC-ClinicDB, scores were taken makes use of data from the following iterations of the VQA
from the following papers: challenge:
DoubleU-Net: A Deep Convolutional Neural Network for VQA Challenge 2016
Medical Image Segmentation VQA Challenge 2017
Encoder-Decoder with Atrous Separable Convolution for VQA Challenge 2018
Semantic Image Segmentation VQA Challenge 2019
MSRF-Net: A Multi-Scale Residual Fusion Network for VQA Challenge 2020
Biomedical Image Segmentation VQA Challenge 2021
ResUNet++: An Advanced Architecture for Medical Image
Segmentation Kinetics-400, Kinetics-600, and Kinetics-700
U-Net: Convolutional Networks for Biomedical Image Data on Kinetics-400, Kinetics-600, and Kinetics-700 was
Segmentation retrieved through a detailed arXiv literature review cross-
referenced by technical progress reported on Papers with
To highlight progress on Kvasir-SEG, scores were taken Code (Kinetics-400, Kinetics-600, and Kinetics-700). The
from the following papers: reported dates correspond to the year during which a
Encoder-Decoder with Atrous Separable Convolution for paper was first published to arXiv, and the reported results
Semantic Image Segmentation (accuracy) correspond to the result reported in the most
MSRF-Net: A Multi-Scale Residual Fusion Network for recent version of each paper. Details on the Kinetics-400
Biomedical Image Segmentation benchmark can be found in the Kinetics-400 paper. Details
PraNet: Parallel Reverse Attention Network for Polyp on the Kinetics-600 benchmark can be found in the
Segmentation Kinetics-600 paper. Details on the Kinetics-700 benchmark
U-Net: Convolutional Networks for Biomedical Image can be found in the Kinetics-700 paper.
Segmentation
To highlight progress on Kinetics-400, scores were taken
National Institute of Standards and Technology from the following papers:
(NIST) Face Recognition Vendor Test (FRVT) Co-Training Transformer with Videos and Images Improves
and NIST FRVT Face Mask Effects Action Recognition
Data on NIST FRVT 1:1 verification accuracy by dataset Large-Scale Weakly-Supervised Pre-training for Video
was obtained from the FRVT 1:1 verification leaderboard. Action Recognition
Data on NIST FRVT face mask effects was obtained from Multiview Transformers for Video Recognition
the FRVT face mask effects leaderboard. The face mask Non-Local Neural Networks
effects leaderboard contains results of the testing of Omni-Sourced Webly-Supervised Learning for Video
319 face recognition algorithms that were submitted to Recognition
FRVT prior to and post mid-March 2020, when the COVID SlowFast Networks for Video Recognition
pandemic began. Temporal Segment Networks: Towards Good Practices for
Deep Action Recognition

To highlight progress on Kinetics-600, scores were taken on the COCO benchmark can be found in the COCO paper.
from the following papers:
Masked Feature Prediction for Self-Supervised Visual Pre- To highlight progress on COCO without the use of extra
Training training data, scores were taken from the following
Multiview Transformers for Video Recognition papers:
Learning Spatio-Temporal Representation with Local and An Analysis of Scale Invariance in Object Detection – SNIP
Global Diffusion Deformable ConvNets v2: More Deformable, Better Results
Rethinking Spatiotemporal Feature Learning: Speed- Dynamic Head: Unifying Object Detection Heads with
Accuracy Trade-offs in Video Classification Attentions
SlowFast Networks for Video Recognition Inside-Outside Net: Detecting Objects in Context with Skip
Pooling and Recurrent Neural Networks
To highlight progress on Kinetics-700, scores were taken Mish: A Self Regularized Non-Monotonic Activation
from the following papers: Function
Learn to Cycle: Time-Consistent Feature Discovery for Scaled-YOLOv4: Scaling Cross Stage Partial Network
Action Recognition
Masked Feature Prediction for Self-Supervised Visual Pre- To highlight progress on COCO with the use of extra
Training training data, scores were taken from the following
Multiview Transformers for Video Recognition papers:
EfficientDet: Scalable and Efficient Object Detection
ActivityNet: Temporal Action Localization Task
Grounded Language-Image Pre-Training
In the challenge, there are three separate tasks, but
they focus on the main problem of temporally localizing
You Only Look Once (YOLO)
where activities happen in untrimmed videos from the
Data on YOLO mean average precision (mAP50) was
ActivityNet benchmark. To source information on the
retrieved through a detailed arXiv literature review
state-of-the-art results for TALT, the Index did a detailed
and survey of GitHub repositories. The reported dates
survey of arXiv papers in addition to reports of yearly
correspond to the year during which a paper was first
ActivityNet challenge results. More specifically, the Index
published to arXiv or a method was introduced. More
made use of the following sources of information:
specifically, the Index made use of the following sources
TALT 2016
of information:
TALT 2017
YOLO 2016
TALT 2018
YOLO 2018
TALT 2019
YOLO 2020
TALT 2020
YOLO 2021
TALT 2021
YOLO results for 2017 and 2019 were not included in the
Common Object in Context (COCO)
index as no state-of-the-art improvements in YOLO for
Data on COCO mean average precision (mAP50) was
those years were uncovered during the literature review
retrieved through a detailed arXiv literature review cross-
and survey of GitHub repositories.
referenced by technical progress reported on Papers
with Code. The reported dates correspond to the year
Visual Commonsense Reasoning (VCR)
during which a paper was first published to arXiv, and
Technical progress for VCR is taken from the VCR
the reported results (mAP50) correspond to the result
leaderboard; the VCR leaderboard webpage further
reported in the most recent version of each paper. Details
delineates the methodology behind the VCR challenge.

Human performance on VCR is taken from Zellers et al. (2018). Details on the VCR benchmark can be found in the VCR
paper.
SuperGLUE
The SuperGLUE benchmark data was pulled from the SuperGLUE leaderboard. Details about the SuperGLUE benchmark
are in the SuperGLUE paper and SuperGLUE software toolkit. The tasks and evaluation metrics for SuperGLUE are:
NAME IDENTIFIER METRIC

Broad Coverage Diagnostics AX-b Matthew’s Carr
CommitmentBank CB Avg. F1/ Accuracy
Choice of Plausible Alternatives COPA Accuracy
Multi-Sentence Reading Comprehension MultiRC F1a/EM
Recognizing Textual Entailment RTE Accuracy
Words in Context WiC Accuracy
The Winograd Schema Challenge WSC Accuracy
BooIQ BooIQ Accuracy
Reading Comprehension with Commonsense Reasoning ReCoRD F1/ Accuracy
Winogender Schema Diagnostic AX-g Gender Parity/ Accuracy
was first published to arXiv, and the reported results

SQuAD 1.1 and SQuAD 2.0 (ROUGE-1) correspond to the result reported in the most
Data on SQuAD 1.1 performance was taken from Papers recent version of each paper. Details about the arXiv
with Code. Data on SQuAD 2.0 performance was taken benchmark are in the arXiv dataset webpage.
from the SQuAD 2.0 leaderboard. Details about the SQuAD
1.1 benchmark are in the SQuAD 1.1 paper. Details about To highlight progress on arXiv without the use of extra
the SQuAD 2.0 benchmark are in the SQuAD 2.0 paper. training data, scores were taken from the following
papers:
Reading Comprehension Dataset Requiring A Discourse-Aware Attention Model for Abstractive
Logical Reasoning (ReClor) Summarization of Long Documents
Data on ReClor performance was taken from the ReClor Extractive Summarization of Long Documents by
leaderboard. Details about the ReClor benchmark are in Combining Global and Local Context
the ReClor paper. Get to the Point: Summarization with Pointer-Generator
Networks
arXiv Systematically Exploring Redundancy Reduction in
Data on arXiv recall-oriented understudy for gisting Summarizing Long Documents
evaluation (ROUGE-1) was retrieved through a detailed Sparsifying Transformer Models with Trainable
arXiv literature review cross-referenced by technical Representation Pooling
progress reported on Papers with Code. The reported
dates correspond to the year during which a paper

To highlight progress on arXiv with the use of extra Stanford Natural Language Inference (SNLI)
training data, scores were taken from the following Data on Stanford Natural Language Inference (SNLI)
papers: accuracy was retrieved through a detailed arXiv literature
Big Bird: Transformers for Longer Sequences review cross-referenced by technical progress reported on
Hierarchical Learning for Generation with Long Source Papers with Code. The reported dates correspond to the
Sequences year during which a paper was first published to arXiv, and
PEGASUS: Pre-Training with Extracted Gap-Sentences for the reported results (accuracy) correspond to the result
Abstractive Summarization reported in the most recent version of each paper. Details
on the SNLI benchmark can be found in the SNLI paper.
PubMed
Data on PubMed recall-oriented understudy for gisting To highlight progress on SNLI, scores were taken from the
evaluation (ROUGE-1) was retrieved through a detailed following papers:
arXiv literature review cross-referenced by technical Compare, Compress and Propagate: Enhancing Neural
progress reported on Papers with Code. The reported Architectures with Alignment Factorization for Natural
dates correspond to the year during which a paper Language Inference
was first published to arXiv, and the reported results Convolutional Neural Networks for Sentence Classification
(ROUGE-1) correspond to the result reported in the most Enhanced LSTM for Natural Language Inference
recent version of each paper. Details about the PubMed Entailment as Few-Shot Learner
benchmark are in the PubMed paper. Explicit Contextual Semantics for Text Comprehension
Semantics-Aware BERT for Language Understanding
To highlight progress on PubMed without the use of Self-Explaining Structures Improve NLP Models
extra training data, scores were taken from the following
papers: Abductive Natural Language Inference (aNLI)
A Discourse-Aware Attention Model for Abstractive Data on Abductive Natural Language Inference (aNLI) was
Summarization of Long Documents sourced from the Allen Institute for AI’s aNLI leaderboard.
Extractive Summarization of Long Documents by Details on the aNLI benchmark can be found in the aNLI
Combining Global and Local Context paper.
Get to the Point: Summarization with Pointer-Generator
Networks SemEval 2014 Task 4 Sub Task 2
Sparsifying Transformer Models with Trainable Data on SemEval 2014 Task 4 Sub Task 2 accuracy was
Representation Pooling retrieved through a detailed arXiv literature review cross-
referenced by technical progress reported on Papers
To highlight progress on PubMed with the use of extra with Code. The reported dates correspond to the year
training data, scores were taken from the following during which a paper was first published to arXiv, and
papers: the reported results (accuracy) correspond to the result
A Divide-and-Conquer Approach to the Summarization of reported in the most recent version of each paper. Details
Long Documents on the SemEval benchmark can be found in the SemEval
Hierarchical Learning for Generation with Long Source 2014 paper.
Sequences
PEGASUS: Pre-Training with Extracted Gap-Sentences for To highlight progress on SemEval, scores were taken from
Abstractive Summarization the following papers:
A Multi-Task Learning Model for Chinese-Oriented Aspect
Polarity Classification and Aspect Term Extraction

Aspect Level Sentiment Classification with Deep Memory following papers:

Network Understanding Back-Translation at Scale
Back to Reality: Leveraging Pattern-Driven Modeling to Very Deep Transformers for Neural Machine Translation
Enable Affordable Sentiment Dependency Learning
Effective LSTMs for Target-Dependent Sentiment To highlight progress on WMT2014 English-German
Classification without the use of extra training data, scores were taken
Hierarchical Attention Based Position-Aware Network for from the following papers:
Aspect-Level Sentiment Analysis BERT, mBERT, or BiBERT? A Study on Contextualized
Investigating Typed Syntactic Dependencies for Targeted Embeddings for Neural Machine Translation
Sentiment Classification Using Graph Attention Neural Effective Approaches to Attention-based Neural Machine
Network Translation
Recurrent Attention Network on Memory for Aspect Data Diversification: A Simple Strategy for Neural Machine
Sentiment Analysis Translation
Fast and Simple Mixture of Softmaxes with BPE and
WMT2014, English-French and English-German Hybrid-LightRNN for Language Generation
Data on WMT2014 English-French and English-German Google’s Neural Machine Translation System: Bridging the
BLEU score was retrieved through a detailed arXiv Gap Between Human and Machine Translation
literature review cross-referenced by technical progress Incorporating BERT into Neural Machine Translation
reported on Papers with Code (English-French and Weighted Transformer Network for Machine Translation
English-German). The reported dates correspond to the
year during which a paper was first published to arXiv, and To highlight progress on WMT2014 English-German with
the reported results (BLEU score) correspond to the result the use of extra training data, scores were taken from the
reported in the most recent version of each paper. Details following papers:
about both the WMT2014 English-French and English- Lessons on Parameter Sharing across Layers in
German benchmarks can be found in the WMT2014 paper. Transformers
Understanding Back-Translation at Scale
To highlight progress on WMT2014 English-French without
the use of extra training data, scores were taken from the Number of Commercially Available MT
following papers: Systems
Addressing the Rare Word Problem in Neural Machine Details about the number of commercially available MT
Translation systems was sourced from Intento’s report The State of
Google’s Neural Machine Translation System: Bridging the Machine Translation, 2021. Intento is a San Francisco-
Gap Between Human and Machine Translation based startup that analyzes commercially available MT
MUSE: Parallel Multi-Scale Attention for Sequence to services.
Sequence Learning
R-Drop: Regularized Dropout for Neural Networks LibriSpeech (Test-Clean and Other Dataset)
Scaling Neural Machine Translation Data on LibriSpeech (Test-Clean and Other) word error
Understanding the Difficulty of Training Transformers rate was retrieved through a detailed arXiv literature
Weighted Transformer Network for Machine Translation review cross-referenced by technical progress reported
on Papers with Code (Test-Clean and Other). The reported
To highlight progress on WMT2014 English-French with dates correspond to the year during which a paper was
the use of extra training data, scores were taken from the first published to arXiv, and the reported results (word

error rate) correspond to the result reported in the most To highlight progress on LibriSpeech Test-Other with the
recent version of each paper. Details about both the use of extra training data, scores were taken from the
LibriSpeech Test-Clean and Test-Other benchmarks can following papers:
be found in the LibriSpeech paper. Deep Speech 2: End-to-End Speech Recognition in English
and Mandarin
To highlight progress on LibriSpeech Test-Clean without End-to-End ASR: From Supervised to Semi-Supervised
the use of extra training data, scores were taken from the Learning with Modern Architectures
following papers: Pushing the Limits of Semi-Supervised Learning for
ASAPP-ASR: Multistream CNN and Self-Attentive SRU for Automatic Speech Recognition
SOTA Speech Recognition W2v-BERT: Combining Contrastive Learning and Masked
Letter-Based Speech Recognition with Gated ConvNets Language Modeling for Self-Supervised Speech Pre-
Neural Network Language Modeling with Letter-Based Training
Features and Importance Sampling
SpeechStew: Simply Mix All Available Speech Recognition VoxCeleb
Data to Train One Large Neural Network VoxCeleb is an audio-visual dataset consisting of short
State-of-the-Art Speech Recognition Using Multi-Stream clips of human speech, extracted from interview videos
Self-Attention With Dilated 1D Convolutions uploaded to YouTube. VoxCeleb contains speech from
7,000-plus speakers spanning a wide range of ethnicities,
To highlight progress on LibriSpeech Test-Clean with the accents, professions, and ages—amounting to over a
use of extra training data, scores were taken from the million utterances (face-tracks are captured “in the wild,”
following papers: with background chatter, laughter, overlapping speech,
Deep Speech 2: End-to-End Speech Recognition in English pose variation, and different lighting conditions) recorded
and Mandarin over a period of 2,000 hours (both audio and video). Each
End-to-End ASR: From Supervised to Semi-Supervised segment is at least three seconds long. The data contains
Learning with Modern Architectures an audio dataset based on celebrity voices, shorts,
Pushing the Limits of Semi-Supervised Learning for films, and conversational pieces (e.g., talk shows). The
Automatic Speech Recognition initial VoxCeleb 1 (100,000 utterances taken from 1,251
celebrities on YouTube) was expanded to VoxCeleb 2 (1
To highlight progress on LibriSpeech Test-Other without million utterances from 6,112 celebrities).
the use of extra training data, scores were taken from the
following papers: For the sake of consistency, the AI Index reported scores
Conformer: Convolution-Augmented Transformer for on the initial VoxCeleb dataset. Specifically, the Index
Speech Recognition made use of the following sources of information:
Neural Network Language Modeling with Letter-Based The IDLAB VoxSRC-20 Submission: Large Margin Fine-
Features and Importance Sampling Tuning and Quality-Aware Score Calibration in DNN Based
SpeechStew: Simply Mix All Available Speech Recognition Speaker Verification
Data to Train One Large Neural Network The SpeakIn System for VoxCeleb Speaker Recognition
Transformer-Based Acoustic Modeling for Hybrid Speech Challenge 2021
Recognition VoxCeleb: A Large-Scale Speaker Identification Dataset
VoxCeleb2: Deep Speaker Recognition
VoxCeleb: Large-Scale Speaker Verification in the Wild

MovieLens 20M MaskNet: Introducing Feature-Wise Multiplication to CTR

Data on MovieLens 20M normalized discounted Ranking Models by Instance-Guided Mask
cumulative gain@100 (nDCG@100) was retrieved through Product-Based Neural Networks for User Response
a detailed arXiv literature review cross-referenced by Prediction
technical progress reported on Papers with Code. The
reported dates correspond to the year during which Arcade Learning Environment: Atari-57
a paper was first published to arXiv, and the reported Data on Arcade Learning Environment: Atari-57 mean
results (nDCG@100) correspond to the result reported human-normalized score was retrieved through a detailed
in the most recent version of each paper. Details on the arXiv literature review cross-referenced by technical
MovieLens series of benchmarks can be found in Harper et progress reported on Papers with Code. The reported
al. 2015. dates correspond to the year during which a paper was
first published to arXiv, and the reported results (mean
To highlight progress on MovieLens 20M, scores were human-normalized score) correspond to the result
taken from the following papers: reported in the most recent version of each paper. Details
Deep Variational Autoencoder with Shallow Parallel Path on the Arcade Learning Environment: Atari-57 benchmark
for Top-N Recommendation (VASP) can be found in the Arcade Learning Environment paper.
Enhancing VAEs for Collaborative Filtering: Flexible Priors
& Gating Mechanisms To highlight progress on Arcade Learning Environment:
RaCT: Toward Amortized Ranking-Critical Training for Atari-57, scores were taken from the following papers:
Collaborative Filtering Dueling Network Architectures for Deep Reinforcement
Variational Autoencoders for Collaborative Filtering Learning
Distributional Reinforcement Learning with Quantile
Criteo Regression
Data on Criteo area under curve score (AUC) was retrieved GDI: Rethinking What Makes Reinforcement Learning
through a detailed arXiv literature review cross-referenced Different From Supervised Learning
by technical progress reported on Papers with Code. Mastering Atari, Go, Chess and Shogi by Planning with a
The reported dates correspond to the year during which Learned Model
a paper was first published to arXiv, and the reported Recurrent Experience Replay in Distributed Reinforcement
results (AUC) correspond to the result reported in the Learning
most recent version of each paper. Details on the Criteo
benchmark can be found on the Criteo Kaggle Challenge Procgen
page. Data on Procgen mean-normalized score was retrieved
through a detailed arXiv literature review. The reported
To highlight progress on Criteo, scores were taken from dates correspond to the year during which a paper was
the following papers: first published to arXiv, and the reported results (mean-
AutoInt: Automatic Feature Interaction Learning via Self- normalized score) correspond to the result reported in the
Attentive Neural Networks most recent version of each paper. Details on the Procgen
DeepFM: A Factorization-Machine Based Neural Network benchmark can be found in the Procgen paper.
for CTR Prediction
DeepLight: Deep Lightweight Feature Interactions for To highlight progress on Procgen, scores were taken from
Accelerating CTR Predictions in Ad Serving the following papers:
FAT-DeepFFM: Field Attentive Deep Field-aware Automatic Data Augmentation for Generalization in
Factorization Machine Reinforcement Learning

Leveraging Procedural Generation to Benchmark Details on the MLPerf Training benchmark can be found
Reinforcement Learning in the MLPerf Training benchmark paper. Details on
Procedural Generalization by Planning with Self- the current benchmark categories as well as technical
Supervised World Models information about submission and competition
subdivisions can be found on the MLPerf Training
Chess webpage.
Data on the performance of chess software engines was
taken from the Swedish Chess Computer Association’s ImageNet Training Cost
ranking of top chess software engines. The Swedish Chess Data on ImageNet Training cost was based on research
Computer Association tests computer chess software from DAWNBench and the individual research of Deepak
systems against one another and releases a ranking list of Narayanan. DAWNBench is a benchmark suite for end-to-
the top-performing systems. The ranking list produced by end deep-learning training and inference. DAWNBench
the Swedish Chess Computer Association is a statistically provides a reference set of common deep-learning
significant and meaningful measurement of chess engine workloads for quantifying training time, training cost,
performance because engines are pitted against one inference latency, and inference cost across different
another in thousands of tournament-like games and each optimization strategies, model architectures, software
employ the same underlying hardware. Data on Magnus frameworks, clouds, and hardware. More details available
Carlsen’s top ELO score was taken from the International at DAWNBench.
Chess Federation.
Because DAWNBench was deprecated after March 2020,
Training Time and Number of Accelerators data on the training cost for the most recent round of
Data on training time and number of accelerators for AI MLPerf submissions was manually collected by Deepak
systems was taken from the MLPerf Training benchmark Narayanan.
competitions. More specifically, the AI Index made use of
data from the following MLPerf training competitions:
MLPerf Training v0.5, 2018

AI Index Robotics Survey

The survey was distributed to 509 professors who specialize in robotics from 67 universities online over three waves from
December 2021 to February 2022. The selection of universities was based on the World University Rankings 2021 with
geographic representation across the globe. Please view the complete survey raw data in our public data folder here.
The survey was completed by 101 professors from 43 universities, including:

Aalborg University, Denmark Stanford University, United States
Ain Shams University, Egypt Stellenbosch University, South Africa
Carnegie Mellon University, United States Swiss Federal Institute of Technology Lausanne,
Columbia University, United States Switzerland
Cornell University, United States Tokyo Institute of Technology, Japan
Delft University of Technology, Netherlands University of British Columbia, Canada
ETH Zurich, Switzerland University of California Berkeley, United States
Hong Kong University of Science and Technology, Hong University of California Los Angeles, United States
Kong University of California San Diego, United States
Korea Advanced Institute of Science and Technology, University of Cambridge, United Kingdom
South Korea University of Cape Town, South Africa
KU Leuven, Belgium University College London, United Kingdom
Massachusetts Institute of Technology, United States University of Hong Kong, Hong kong
Nanyang Technological University, Singapore University of Illinois at Urbana-Champaign, United States
National Polytechnic Institute, Mexico University of Malaya, Malaysia
National University of Singapore, Singapore University of Manchester, United Kingdom
Peking University, China University of Michigan, United States
Politecnico di Milano, Italy Universitat Politècnica de Catalunya, Spain
Pontificia Universidad Católica de Chile, Chile University of Texas at Austin, United States
Princeton University, United States University of Tokyo, Japan
Purdue University, United States University of Toronto, Canada
RWTH Aachen University, Germany University of Waterloo, Canada
Seoul National University, South Korea Zhejiang University, China

Index Report 2022 Chapter 3: Technical AI Ethics
CHAPTER 3: TECHNICAL AI ETHICS

A I E T H I C S T R E N D S AT FAC CT A N D of papers accepted to the NeurIPS main track with titles
NEURIPS containing keywords (e.g., “counterfactual” or “causal”
To understand trends at the ACM Conference on Fairness, for tracking papers related to causal effect), as well as
Accountability, and Transparency, this section tracks papers submitted to related workshops. See the list of
FAccT papers published in conference proceedings workshops considered for analysis here.
from 2018 to 2021. We categorize author affiliations
into academic, industry, nonprofit, government, and M E TA-A N A LYS I S O F FA I R N E S S
independent categories, while also tracking the location AND BIAS METRICS
of their affiliated institution. Authors with multiple For the analysis conducted on fairness and bias metrics in
affiliations are counted once in each category (academic AI, we identify and report on benchmark and diagnostic
and industry), but multiple affiliations of the same type metrics which have been consistently cited in the
(i.e., authors belonging to two academic institutions) are academic community, reported on a public leaderboard,
counted once in the category. or reported for publicly available baseline models (e.g.,
GPT-3, BERT, ALBERT). We note that research paper
For the analysis conducted on NeurIPS publications, we citations are a lagging indicator of adoption, and metrics
identify workshops themed around real-world impact and which have been very recently adopted may not be
label papers with a single main category in “healthcare,” reflected in the data for 2021.
“climate,” “finance,” “developing world,” or “other,” where
“other” denotes a paper related to a real-world use case For Figures 3.1.1 and 3.1.2, we track metrics from the
but not in one of the other categories. following papers and projects:

Aligning AI with Shared Human Values
We tally the number of papers in each category to reach
Assessing Social and Intersectional Biases in
the numbers found in Figure 3.3.3. Papers are not double-
Contextualized Word Representations
counted in multiple categories. We note that this data
Bias in Bios: A Case Study of Semantic Representation Bias
may not be as accurate for data pre-2018 as societal
in a High-Stakes Setting
impacts work at NeurIPS has historically been categorized
BOLD: Dataset and Metrics for Measuring Biases in Open-
under a broad “AI for social impact” umbrella,7 but it has
Ended Language Generation
recently been split into more granular research areas.
Certifying and Removing Disparate Impact
Examples include workshops dedicated to machine
CivilComments: Jigsaw Unintended Bias in Toxicity
learning for health,8 climate,9 policy & governance10,
Classification
disaster response11, and the developing world.12
CrowS-Pairs: A Challenge Dataset for Measuring Social
Biases in Masked Language Models
To track trends around specific technical topics at
NeurIPS as in Figures 3.3.4–3.3.7, we count the number
7 See 2018 Workshop on Ethical, Social and Governance Issues in AI 2018, 2018 AI for Social Good Workshop, 2019 Joint Workshop on AI for Social Good, 2020 Resistance AI Workshop, 2020 Navigating the
Broader Impacts of AI Research Workshop.
8 See 2014 Machine Learning for Clinical Data Analysis, Healthcare and Genomics, 2015 Machine Learning for Healthcare, 2016 Machine Learning for Health, 2017 Machine Learning for Health.
9 See 2013 Machine Learning for Sustainability, 2020 AI for Earth Sciences, 2019, 2020, 2021 Tackling Climate Change with ML.
10 See 2016 People and Machines, 2019 Joint Workshop on AI for Social Good–Public Policy, 2021 Human-Centered AI.
11 See 2019 AI for Humanitarian Assistance and Disaster Response, 2020 Second Workshop on AI for Humanitarian Assistance and Disaster Response, 2021 Third Workshop on AI for Humanitarian
Assistance and Disaster Response.
12 See 2017–2021 Machine Learning for the Developing World Workshops.

Detecting Emergent Intersectional Biases: Contextualized N AT U R A L L A N G UAG E P R O C E S S I N G

Word Embeddings Contain a Distribution of Human-Like BIAS METRICS
Biases In Section 3.2, we track citations of the Perspective API
Equality of Opportunity in Supervised Learning created by Jigsaw at Google. The Perspective API has been
Evaluating Gender Bias in Machine Translation adopted widely by researchers and engineers in natural
Evaluating Gender Bias in Natural Language Inference language processing. Its creators define toxicity as “a rude,
Examining Gender Bias in Languages with Grammatical disrespectful, or unreasonable comment that is likely to
Gender make someone leave a discussion,” and the tool is powered
Fairness Through Awareness by machine learning models trained on a proprietary
Gender Bias in Coreference Resolution dataset of comments from Wikipedia and news websites.
Gender Bias in Coreference Resolution: Evaluation and We include the following papers in our analysis:
Debiasing Methods
Gender Bias in Multilingual Embeddings and Cross- #ContextMatters: Advantages and Limitations of Using
Lingual Transfer Machine Learning to Support Women in Politics
Gender Shades: Intersectional Accuracy Disparities in A General Language Assistant as a Laboratory for
Commercial Gender Classification Alignment
Image Representations Learned with Unsupervised A Machine Learning Approach to Comment Toxicity
Pretraining Contain Human-Like Biases Classification
Measuring and Reducing Gendered Correlations in A Novel Preprocessing Technique for Toxic Comment
Pretrained Models Classification
Measuring Bias in Contextualized Word Representations Adversarial Text Generation for Google’s Perspective API
Measuring Bias with Wasserstein Distance Avoiding Unintended Bias in Toxicity Classification with
Nuanced Metrics for Measuring Unintended Bias with Real Neural Networks
Data for Text Classification Bad‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌‌Charac ters: Imperc eptible NLP Attac ks
On Formalizing Fairness in Prediction with Machine Challenges in Detoxifying Language Models
Learning Classification of Online Toxic Comments Using Machine
On Measuring Social Biases in Sentence Encoders Learning Algorithms
Perspective API Context Aware Text Classification and Recommendation
RealToxicityPrompts: Evaluating Neural Toxic Model for Toxic Comments Using Logistic Regression
Degeneration in Language Models Detecting Cross-Geographic Biases in Toxicity Modeling on
Scaling Language Models: Methods, Analysis & Insights Social Media
from Training Gopher Detoxifying Language Models Risks Marginalizing Minority
Semantics Derived Automatically from Language Corpora Voices
Contain Human-Like Biases Fighting Hate Speech, Silencing Drag Queens? Artificial
StereoSet: Measuring Stereotypical Bias in Pretrained Intelligence in Content Moderation and Risks to LGBTQ
Language Models Voices Online
The Woman Worked as a Babysitter: On Biases in HATEMOJI: A Test Suite and Adversarially Generated
Language Generation Dataset for Benchmarking and Detecting Emoji-Based
When Worlds Collide: Integrating Different Counterfactual Hate
Assumptions in Fairness HotFlip: White-Box Adversarial Examples for Text
Classification
Identifying Latent Toxic Features on YouTube Using Non-
Negative Matrix Factorization

Interpreting Social Respect: A Normative Lens for ML While the Perspective API is used widely within machine
Models learning research and also for measuring online toxicity,
Knowledge-Based Neural Framework for Sexism toxicity in the specific domains used to train the models
Detection and Classification undergirding Perspective (e.g., news, Wikipedia) may not
Large Pretrained Language Models Contain Human-Like be broadly representative of all forms of toxicity (e.g.,
Biases of What Is Right and Wrong to Do trolling). Other known caveats include biases against text
Leveraging Multilingual Transformers for Hate Speech written by minority voices: The Perspective API has been
Detection shown to disproportionately assign high toxicity scores
Limitations of Pinned AUC for Measuring Unintended Bias to text that contains mentions of minority identities (e.g.,
Machine Learning Suites for Online Toxicity Detection “I am a gay man”). As a result, detoxification techniques
Mitigating Harm in Language Models with Conditional- built with labels sourced from the Perspective API result
Likelihood Filtration in models that are less capable of modeling language
On-the-Fly Controlled Text Generation with Experts and used by minority groups, and they avoid mentioning
Anti-Experts minority identities.
Process for Adapting Language Models to Society (PALMS)
with Values-Targeted Datasets We note that the effect size metric reported in the
Racial Bias in Hate Speech and Abusive Language Word Embeddings Association Test (WEAT) section is
Detection Datasets highly sensitive to rare words, as it has been shown
RealToxicityPrompts: Evaluating Neural Toxic that removing less than 1% of relevant documents in
Degeneration in Language Models a corpus can significantly impact the WEAT effect size.
Scaling Language Models: Methods, Analysis & Insights This means that effect size is not guaranteed to be a
from Training Gopher robust metric for assessing bias in embeddings. While
Self-Diagnosis and Self-Debiasing: A Proposal for we report on a subset of embedding association tasks
Reducing Corpus-Based Bias in NLP measuring bias along gender and racial axes, these
Social Bias Frames: Reasoning About Social and Power embedding association tests have been extended to
Implications of Language quantify the effect across intersectional axes (e.g.,
Social Biases in NLP Models as Barriers for Persons with EuropeanAmerican+male, AfricanAmerican+male,
Disabilities AfricanAmerican+female).
Stereotypical Bias Removal for Hate Speech Detection
Task Using Knowledge-Based Generalizations In the analysis of embeddings from over 100 years of
The Risk of Racial Bias in Hate Speech Detection U.S. Census data, embedding bias was measured by
Towards Measuring Adversarial Twitter Interactions computing the difference between average embedding
Against Candidates in the US Midterm Elections distances. For example, gender bias is calculated as the
Toxic Comment Classification Using Hybrid Deep Learning average distance of embeddings of words associated with
Model women (e.g., she, female) compared to embeddings of
Toxicity-Associated News Classification: The Impact of words for occupations (e.g., teacher, lawyer), minus the
Metadata and Content Features same average distance calculated for words associated
Understanding BERT Performance in Propaganda Analysis with men.
White-to-Black: Efficient Distillation of Black-Box
Adversarial Attacks
Women, Politics and Twitter: Using Machine Learning to
Change the Discourse

FAC T UA L I T Y A N D T R U T H F U L N E S S Hallucination refers to language models fabricating

statements not found in factually correct supporting
Definitions evidence or input documents. In closed-form dialog,
The concepts of factuality, factual correctness, factual summarization, or question-answering, a system that
accuracy, and veracity are all used to refer to conformity hallucinates is considered untruthful.
with facts or truth. Recent work in AI aims to assess
factual correctness within language models and Language Diversity in Training Data
characterize their limitations. Imbalanced language distribution in training data
impacts the performance of general-purpose language
While human truthfulness is a relatively well-understood models. For example, the Gopher family of models is
concept, truthfulness is not a well-characterized concept trained on MassiveText (10.5TB), which is a dataset
within the context of AI. A group of researchers has made up of 99% English. Similarly, only 7.4% of GPT-3
proposed frameworks for what it means for a system training data is in non-English languages. In contrast,
to be truthful—for example, a broadly truthful system XGLM, a recent model family from Meta AI, is trained on
should avoid lying or using true statements to mislead a training data of 30 languages, and upsamples low-
or misdirect; should be clear, informative, and (mostly) resource languages to create a more balanced language
cooperative in conversation; and should be well-calibrated, representation. See Figure 1 on the XGLM paper that
self-aware, and open about the limits of its knowledge. compares the language distribution of XGLM and GPT-3.
A definition of narrow truthfulness may simply refer to
systems which avoid stating falsehoods. The authors of In addition, Figure 7 of the XGLM paper highlights the
TruthfulQA define a system as truthful only if it avoids extent to which language models can effectively store
asserting a false statement; refusing to answer a question, factual knowledge by comparing the performance of
expressing uncertainty, or giving a true but irrelevant XGLM (a multilingual language model) with GPT-3, a
answer may be considered truthful but not informative. monolingual model. Performance was evaluated on
knowledge triplet completion using the mLAMA dataset,
Truthfulness is related to alignment: A truthful system is which was translated from the English benchmark LAMA
aligned with human values and goals. In one definition of using Google Translate. GPT-3 outperforms in English,
alignment, an aligned system is one that is helpful, honest, but XGLM outperforms in non-English languages. Further
and harmless. Since we cannot yet measure honesty within results show that more diverse language representation
a system, truthfulness can be used as a proxy. improves language model performance in tasks such as
translation.
An honest system is one that asserts only what it
“believes” or one that never contradicts its own beliefs. In 2021, Congress inquired into the content moderation
A system can be honest but not truthful—for example, if practices of social media companies in non-English
an honest system believes that vaccines are unsafe, it can languages, and emphasized the importance of equal
claim this honestly, despite the statement being factually access to truthful and trustworthy information regardless
incorrect. Conversely, a system can be truthful but not of language. As these companies start to adopt language
honest: It may believe vaccines are unsafe but asserts models into their fact-checking and content moderation
they are safe to pass a test. Another work proposes that processes for languages around the world, it is critical to
an honest system should give accurate information, not be able to measure the disproportionate negative impact
mislead users, be calibrated (e.g., it should be correct 80% of using models which underperform on non-English
of the time when it claims 80% confidence), and express languages.
appropriate levels of uncertainty.

Index Report 2022 Chapter 4: The Economy and Education
CHAPTER 4: THE ECONOMY

AND EDUCATION
EMSI BURNING GLASS and Emsi Burning Glass, the labor market demand
Prepared by Julia Nitschke, Summer Jasinski, Bledi Taska captured by Emsi Burning Glass data represents over 95%
and Rucha Vankudre of the total labor demand. Jobs not posted online are
usually in small businesses (the classic example being the
Emsi Burning Glass delivers job market analytics “Help Wanted” sign in the restaurant window) and union
that empower employers, workers, and educators to hiring halls.
make data-driven decisions. The company’s artificial
intelligence technology analyzes hundreds of millions Measuring Demand for AI
of job postings and real-life career transitions to provide In order to measure employers’ demand for AI skills, Emsi
insight into labor market patterns. This real-time strategic Burning Glass uses its skills taxonomy of over 17,000 skills.
intelligence offers crucial insights, such as what jobs are The list of AI skills from Emsi Burning Glass data is shown
most in demand, the specific skills employers need, and below, with associated skill clusters. While some skills
the career directions that offer the highest potential for are considered to be in the AI cluster specifically, for the
workers. For more information, visit burning-glass.com. purposes of this report, all skills below were considered AI
skills. A job posting was considered an AI job if it requested
Job Posting Data one or more of these skills.
To support these analyses, Emsi Burning Glass mined
its dataset of millions of job postings collected since Artificial Intelligence: Expert System, IBM Watson, IPSoft
2010. Emsi Burning Glass collects postings from over Amelia, Ithink, Virtual Agents, Autonomous Systems, Lidar,
45,000 online job sites to develop a comprehensive, OpenCV, Path Planning, Remote Sensing
real-time portrait of labor market demand. It aggregates
job postings, removes duplicates, and extracts data Natural Language Processing (NLP): ANTLR, Automatic
from job postings text. This includes information on job Speech Recognition (ASR), Chatbot, Computational
title, employer, industry, and region, as well as required Linguistics, Distinguo, Latent Dirichlet Allocation, Latent
experience, education, and skills. Semantic Analysis, Lexalytics, Lexical Acquisition, Lexical
Semantics, Machine Translation (MT), Modular Audio
Job postings are useful for understanding trends in the Recognition Framework (MARF), MoSes, Natural Language
labor market because they allow for a detailed, real- Processing, Natural Language Toolkit (NLTK), Nearest
time look at the skills employers seek. To assess the Neighbor Algorithm, OpenNLP, Sentiment Analysis/
representativeness of job postings data, Emsi Burning Opinion Mining, Speech Recognition, Text Mining, Text to
Glass conducts a number of analyses to compare the Speech (TTS), Tokenization, Word2Vec
distribution of job postings to the distribution of official
government and other third-party sources in the United Neural Networks: Caffe Deep Learning Framework,
States. The primary source of government data on U.S. Convolutional Neural Network (CNN), Deep Learning,
job postings is the Job Openings and Labor Turnover Deeplearning4j, Keras, Long Short-Term Memory (LSTM),
Survey (JOLTS) program, conducted by the Bureau of MXNet, Neural Networks, Pybrain, Recurrent Neural
Labor Statistics. Based on comparisons between JOLTS Network (RNN), TensorFlow

Machine Learning: AdaBoost algorithm, Boosting and classified by taxonomists at LinkedIn into 249 skill
(Machine Learning), Chi Square Automatic Interaction groupings, which are the skill groups represented in the
Detection (CHAID), Classification Algorithms, Clustering dataset. The top skills that make up the AI skill grouping
Algorithms, Decision Trees, Dimensionality Reduction, are machine learning, natural language processing, data
Google Cloud Machine Learning Platform, Gradient structures, artificial intelligence, computer vision, image
boosting, H2O (software), Libsvm, Machine Learning, processing, deep learning, TensorFlow, Pandas (software),
Madlib, Mahout, Microsoft Cognitive Toolkit, MLPACK (C++ and OpenCV, among others.
library), Mlpy, Random Forests, Recommender Systems,
Scikit-learn, Semi-Supervised Learning, Supervised Skill groupings are derived by expert taxonomists
Learning (Machine Learning), Support Vector Machines through a similarity-index methodology that measures
(SVM), Semantic Driven Subtractive Clustering Method skill composition at the industry level. Industries are
(SDSCM), Torch (Machine Learning), Unsupervised classified according to the ISIC 4 industry classification
Learning, Vowpal, Xgboost (Zhu et al., 2018).
Robotics: Blue Prism, Electromechanical Systems, Skills Genome

Motion Planning, Motoman Robot Programming, Robot For any entity (occupation or job, country, sector, etc.), the
Framework, Robotic Systems, Robot Operating System skill genome is an ordered list (a vector) of the 50 “most
(ROS), Robot Programming, Servo Drives / Motors, characteristic skills” of that entity. These most characteristic
Simultaneous Localization and Mapping (SLAM) skills are identified using a TF-IDF algorithm to identify the
most representative skills of the target entity, while down-
Visual Image Recognition: Computer Vision, Image ranking ubiquitous skills that add little information about
Processing, Image Recognition, Machine Vision, Object that specific entity (e.g., Microsoft Word).
Recognition
TF-IDF is a statistical measure that evaluates how
LINKEDIN representative a word (in this case a skill) is to a selected
Prepared by Akash Kaura and Murat Erer entity. This is done by multiplying two metrics:
1. The term frequency of a skill in an entity (TF).
Country Sample 2. The logarithmic inverse entity frequency of the skill
Included countries represent a select sample of eligible across a set of entities (IDF). This indicates how
countries with at least 40% labor force coverage by common or rare a word is in the entire entity set.
LinkedIn and at least 10 AI hires in any given month. The closer IDF is to 0, the more common a word is.
China and India were included in this sample because of So, if the skill is very common across LinkedIn entities,
their increasing importance in the global economy, but and appears in many job or member descriptions, the IDF
LinkedIn coverage in these countries does not reach 40% will approach 0. If, on the other hand, the skill is unique to
of the workforce. Insights for these countries may not specific entities, the IDF will approach 1. Details available
provide as full a picture as other countries, and should be at LinkedIn’s Skills Genome and LinkedIn-World Bank
interpreted accordingly. Methodology note.
Skills (and AI Skills) AI Skills Penetration

LinkedIn members self-report their skills on their LinkedIn The aim of this indicator is to measure the intensity of AI
profiles. Currently, more than 38,000 distinct, standardized skills in an entity (in a particular country, industry, gender,
skills are identified by LinkedIn. These have been coded etc.) through the following methodology:

• Compute frequencies for all self-added skills by explicitly added AI skills to their profile and/or they are
LinkedIn members in a given entity (occupation, occupied in an AI occupation representative. The counts of
industry, etc.) in 2015–2021. AI talent are used to calculate talent concentration metrics
• Re-weight skill frequencies using a TF-IDF model (e.g. to calculate the country-level AI talent concentration,
to get the top 50 most representative skills in that we use the counts of AI talent at the country level vis-
entity. These 50 skills compose the “skill genome” of a-vis the counts of LinkedIn members in the respective
that entity. countries).
•C ompute the share of skills that belong to the AI skill
group out of the top skills in the selected entity. Relative AI Skills Penetration
To allow for skills penetration comparisons across
Interpretation: The AI skill penetration rate signals the countries, the skills genomes are calculated and a relevant
prevalence of AI skills across occupations, or the intensity benchmark is selected (e.g., global average). A ratio is then
with which LinkedIn members utilize AI skills in their constructed between a country’s and the benchmark’s AI
jobs. For example, the top 50 skills for the occupation of skills penetrations, controlling for occupations.
engineer are calculated based on the weighted frequency
with which they appear in LinkedIn members’ profiles. If Interpretation: A country’s relative AI skills penetration of
four of the skills that engineers possess belong to the AI 1.5 indicates that AI skills are 1.5 times as frequent as in
skill group, this measure indicates that the penetration of the benchmark, for an overlapping set of occupations.
AI skills is estimated to be 8% among engineers (e.g., 4/50).
Global Comparison
Jobs or Occupations For cross-country comparison, we present the relative
LinkedIn member titles are standardized and grouped penetration rate of AI skills, measured as the sum of the
into approximately 15,000 occupations. These are not penetration of each AI skill across occupations in a given
sector or country specific. These occupations are further country, divided by the average global penetration of AI
standardized into approximately 3,600 occupation skills across the overlapping occupations in a sample of
representatives. Occupation representatives group countries.
occupations with a common role and specialty, regardless
of seniority. Interpretation: A relative penetration rate of 2 means
that the average penetration of AI skills in that country
AI Jobs or Occupations is two times the global average across the same set of
An AI job (technically, occupation representative) is occupations.
an occupation representative that requires AI skills to
perform the job. Skills penetration is used as a signal Global Comparison: By Industry
for whether AI skills are prevalent in an occupation The relative AI skills penetration by country for industry
representative, in any sector where the occupation provides an in-depth sectoral decomposition of AI skill
representative may exist. Examples of such occupations penetration across industries and sample countries.
include (but are not limited to): machine learning
engineer, artificial intelligence specialist, data scientist, Interpretation: A country’s relative AI skill penetration
computer vision engineer, etc. rate of 2 in the education sector means that the average
penetration of AI skills in that country is two times the
AI Talent global average across the same set of occupations in that
A LinkedIn member is considered AI talent if they have sector.

Relative AI Hiring Index Search, Data Sources, and Scope

• LinkedIn Hiring Rate or Overall Hiring Rate is Over 6 million global public and private company profiles
a measure of hires normalized by LinkedIn from multiple data sources are indexed in order to
membership. It is computed as the percentage of search across company descriptions, while filtering and
LinkedIn members who added a new employer in including metadata ranging from investment information
the same period the job began, divided by the total to firmographic information, such as founded year, HQ
number of LinkedIn members in the corresponding location, and more. Company information is updated on a
location. weekly basis. NetBase Quid algorithm reads a big amount
• AI Hiring Rate is computed following the overall of text data from each document to make links between
hiring rate methodology, but only considering different documents based on their similar language. This
members classified as AI talent. process is repeated at an immense scale, which produces
•R elative AI Hiring Index is the pace of change in AI a network with different clusters identifying distinct topics
Hiring Rate normalized by the pace of change in or focus areas. Trends are identified based on keywords,
Overall Hiring Rate, providing a picture of whether phrases, people, companies, institutions that NetBase
hiring of AI talent is growing at a higher, equal, Quid identifies, and the other metadata that is put into the
or lower rate than overall hiring in a market. The software.
relative AI Hiring Index is equal to 1.0 when AI hiring
and overall hiring are growing at the same rate year Data
on year. Organization data is embedded from Capital IQ and
Crunchbase. These companies include all types of
Interpretation: Relative AI Hiring Index shows how fast companies (private, public, operating, operating as a
each country is experiencing growth in AI talent hiring subsidiary, out of business) throughout the world. The
relative to growth in overall hiring in the country. A ratio investment data includes private investments, M&A, public
of 1.2 means the growth in AI talent hiring has outpaced offerings, minority stakes made by PE/VCs, corporate
the growth in overall hiring by 20%. venture arms, governments, and institutions both within
and outside the United States. Some data is simply
N E T BA S E Q U I D unreachable—for instance, when the investors or the
Prepared by Julie Kim and Tejas Sirohi funding amounts by investors are undisclosed. NetBase
Quid also embeds firmographic information such as
NetBase Quid delivers AI-powered consumer and market founding year and HQ location.
intelligence to enable business reinvention in a noisy
and unpredictable world. The software applies artificial NetBase Quid embeds Capital IQ data as a default and
intelligence to reveal patterns in large, unstructured adds in data from Crunchbase for the ones that are not
datasets and to generate visualizations that enable users captured in Capital IQ. This not only yields comprehensive
to make smart, data-driven decisions accurately, quickly, and accurate data on all global organizations, but it also
and efficiently. NetBase Quid uses Boolean query to captures early-stage startups and funding events data.
search for focus areas, topics, and keywords within social, Company information is uploaded on a weekly basis.
news, forums and blogs, companies, and patents data
sources, as well as other custom datasets. NetBase Quid Search Parameters
then visualizes these data points based on the semantic Boolean query is used to search for focus areas, topics,
similarity. and keywords within the archived company database,
within their business descriptions and websites. We can
filter out the search results by HQ regions, investment

amount, operating status, organization type (private/ COMPUTING RESEARCH

public), and founding year. NetBase Quid then visualizes A S S O C I AT I O N (C R A TAU L B E E
these companies by the semantic similarity. If there
SURVEY)
are more than 7,000 companies from the search result,
Prepared by Betsy Bizot (CRA senior research associate)
NetBase Quid selects the 7,000 most relevant companies for
visualization based on the language algorithm.
Source
Computing Research Association (CRA) members
Boolean Search: “artificial intelligence” or “AI” or “machine
are 200-plus North American organizations active in
learning” OR “deep learning”
computing research: academic departments of computer
science and computer engineering; laboratories and
Companies:
centers in industry, government, and academia; and
• Chart 4.2.1: Global AI & ML companies that have
affiliated professional societies (AAAI, ACM, CACS/AIC,
been invested (private, IPO, M&A) from 01/01/2012 to
IEEE Computer Society, SIAM USENIX). CRA’s mission
12/31/2021.
is to enhance innovation by joining with industry,
• Chart 4.2.2–4.2.12: Global AI & ML companies that have
government, and academia to strengthen research and
invested over $1.5M for the last 10 years (January 1,
advanced education in computing. Learn more about
2012 to December 31, 2021)—7,000 companies out of
CRA here.
7,500 companies have been selected through Quid’s
relevance algorithm.
Methodology
CRA Taulbee Survey gathers survey data during the fall
Target Event Definitions
of each academic year by reaching out to over 200 PhD-
• Private investments: A private placement is a private
granting departments. Details about the Taulbee Survey
sale of newly issued securities (equity or debt) by a
can be found here. Taulbee doesn’t directly survey the
company to a selected investor or a selected group
students. The department identifies each new PhD’s area
of investors. The stakes that buyers take in private
of specialization as well as their type of employment.
placements are often minority stakes (under 50%),
Data is collected from September to January of each
although it is possible to take control of a company
academic year for PhDs awarded in the previous
through a private placement as well, in which case
academic year. Results are published in May after data
the private placement would be a majority stake
collection closes. So the 2020 data points were newly
investment.
available last spring, and the numbers provided for 2021
• Minority investment: These refer to minority stake
will be available in May 2020.
acquisitions in Quid, which take place when the buyer
acquires less than 50% of the existing ownership stake
The CRA Taulbee Survey is sent only to doctoral
in entities, asset product, and business divisions.
departments of computer science, computer
• M&A: This refers to a buyer acquiring more than 50%
engineering, and information science/systems.
of the existing ownership stake in entities, asset
Historically, (a) Taulbee covers 1/4 to 1/3 of total BS CS
product, and business divisions.
recipients in the United States; (b) the percent of women
earning bachelor’s degrees is lower in the Taulbee
schools than overall; and (c) Taulbee tracks the trends in
overall CS production.

Nuances
• Of particular interest in PhD job market trends are
the metrics on the AI PhD area of specialization. The
categorization of specialty areas changed in 2008 and
was clarified in 2016. From 2004-2007, AI and robotics
were grouped; from 2008-present, AI is separate; 2016
clarified to respondents that AI includes ML.
• Notes about the trends in new tenure-track hires
(overall and particularly at AAU schools): In the 2018
Taulbee Survey, for the first time we asked how many
new hires had come from the following sources: new
PhD, postdoc, industry, and other academic. Results
indicated that 29% of new assistant professors came
from another academic institution.
• Some may have been teaching or research faculty
rather than tenure-track, but there is probably some
movement between institutions, meaning the total
number hired overstates the total who are actually
new.

Index Report 2022 Chapter 5: AI Policy and Governance
CHAPTER 5: AI POLICY
AND GOVERNANCE
B LO O M B E R G G OV E R N M E N T Legislative Documents: Bloomberg Government
Prepared by Amanda Allen maintains a repository of congressional documents,
including bills, Congressional Budget Office assessments,
Bloomberg Government is a premium, subscription- and reports published by congressional committees,
based service that provides comprehensive information the Congressional Research Service, and other offices.
and analytics for professionals who interact with—or Bloomberg Government also ingests state legislative
are affected by—the government. Delivering news, bills. For the section “AI Policy and Governance,”
analytics, and data-driven decision tools, Bloomberg Bloomberg Government analysts identified all legislation,
Government’s digital workspace gives an intelligent edge congressional committee reports, and CRS reports that
to government affairs and contracting professionals. For referenced one or more AI-specific keywords.
more information or a demo, visit about.bgov.com.
Methodology
Contract Spending: Bloomberg Government’s Contracts
Intelligence Tool structures all contracts data from
www.fpds.gov. The CIT includes a model of government
spending on artificial intelligence-related contracts that is
based on a combination of government-defined product
service codes and more than 100 AI-related keywords.
For the section “U.S. Government Contract Spending,”
Bloomberg Government analysts used contract spending
data from fiscal year 2000 through fiscal year 2021.
Defense RDT&E Budget: Bloomberg Government

organized all the RDT&E budget request line items
available from the Defense Department Comptroller. For
the section “U.S. Department of Defense (DOD) Budget,”
Bloomberg Government used a set of AI-specific keywords
to identify 500 unique budget activities related to artificial
intelligence and machine learning worth a combined $5.9
billion in FY 2021.

G LO BA L L E G I S L AT I O N R E C O R D S O N A I
For AI-related bills passed into laws, the AI Index performed searches of the keyword “artificial intelligence,” in respective
languages, on the websites of 25 countries’ congresses or parliaments, in full-text of bills. Note that only laws passed
by state-level legislative bodies and signed into law (i.e., by presidents or received royal assent) from 2015 to 2021 are
included. Future AI Index reports hope to include analysis on other types of legal documents, such as regulations and
standards, adopted by state- or supranational-level legislative bodies, government agencies, etc.
Australia Denmark
Website: www.legislation.gov.au Website: https://www.retsinformation.dk/
Keyword: artificial Intelligence Keyword: kunstig intelligen
Filters: Filter:
• Legislation types: Acts • Document Type: Laws
• Portfolios: Department of House of Representatives,
Department of Senate Finland
Note: Texts in explanatory memorandum are not counted. Website: https://www.finlex.fi/
Keyword: tekoäly
Belgium Noting under the Current Legislation section
Website: http://www.ejustice.just.fgov.be/loi/loi.htm
Keyword: intelligence artificielle France
Website: https://www.legifrance.gouv.fr/
Brazil Keyword: intelligence artificielle
Website: https://www.camara.leg.br/legislacao Filter:
Keyword: inteligência artificial • texte consolidé
Filter: • Document Type: Law
• Federal legislation
• Type: Law Germany
Website: http://www.gesetze-im-internet.de/index.html
Canada Keyword: künstliche Intelligenz
Website: https://www.parl.ca/legisinfo/ Filter:
Keyword: artificial Intelligence • All federal codes, statutes, and ordinances that are
Note: Results were investigated to determine how many of currently in force
the bills introduced were eventually passed (i.e., received • Volltextsuche (full text)
royal assent) and bill status was recorded. • Und-Verknüpfung der Wörter (entire word)
China India
Website: https://flk.npc.gov.cn/ Website: https://www.indiacode.nic.in
Keyword: 人工智能 Keyword: artificial intelligence
Filters: Note: The website used allows for a search of keywords
• Legislative body: Standing Committee of the in legalization title but not in the full text, as such it is not
National People’s Congress useful for this particular research. Therefore, a Google
search using the “site” function to search the site with the
keyword of “artificial intelligence” is conducted.

Ireland Singapore
Website: www.irishstatutebook.ie Website: https://sso.agc.gov.sg/
Keyword: artificial intelligence Keyword: artificial intelligence
Filter:
Italy • Document Type: Current acts and subsidiary
Website: https://www.normattiva.it/ legislation
Keyword: intelligenza artificiale
Filter: South Africa
• Document Type: law Website: www.gov.za
Keyword: artificial intelligence
Japan Filter:
Website: https://elaws.e-gov.go.jp/ • Document: acts
Keyword: 人工知能 Note: This search function seemingly does not search
Filter: within the context of the full text and so no results were
• Full text returned. Therefore, a Google search using the “site”
• Law function to search the site with the keyword of “artificial
intelligence” is conducted.
Netherlands
Website: https://www.overheid.nl/ South Korea
Keyword: kunstmatige intelligentie Website: https://law.go.kr/eng/; https://elaw.klri.re.kr/
Filter: Keyword: artificial Intelligence or 인공 지능
• Document Type: Wetten Filter:
• Type: Act
New Zealand Note: Cannot search combined words, so individual
Website: www.legislation.govt.nz analysis is conducted.
Keyword: Artificial intelligence
Filter: Spain
• Document type: acts Website: https://www.boe.es/
• Status option: For the status option (example: acts in Keyword: inteligencia artificial
force, current bills, etc.) Filter:
• Type: law
Norway • Head of state (for passed laws)
Website: https://lovdata.no/
Keyword: kunstig intelligens Sweden
Website: https://www.riksdagen.se/
Russia Keyword: artificiell intelligens
Website: http://graph.garant.ru:8080/SESSION/PILOT/ Filter: Swedish Code of Statutes
main.htm (Database “The Federal Laws” in the official
website of the Federation Council of the Federal Assembly
of the Russian Federation.)
Keyword: искусственный интеллект
Filter:
• Words in text

Switzerland
Website: https://www.fedlex.admin.ch/
Keyword: intelligence artificielle
Filter:
• Text category: federal constitution, federal acts, and
federal decrees, miscellaneous texts, orders, and
other forms of legislation.
• Publication period for legislation was limited to 2015-
2021.
United Kingdom
Website: https://www.legislation.gov.uk/
Filter:
• Legislation Type: U.K. Public General Acts & U.K.
Statutory Instruments
United States
Website: https://www.congress.gov/
Filter:
• Source: Legislation
Status of legislation: Became law

M E N T I O N S O F A I I N A I - R E L AT E D L E G I S L AT I O N P R O C E E D I N G S
For mentions of AI in AI-related legislative proceedings around the world, the AI Index performed searches of the keyword
“artificial intelligence,” in respective languages, on the websites of 25 countries’ congresses or parliaments, usually under
sections named “minutes,” “hansard,” etc.
Australia Denmark
Website: https://www.aph.gov.au/Parliamentary_ Website: https://www.retsinformation.dk/
Business/Hansard Keyword: kunstig intelligens
Keyword: artificial intelligence Filter:
• Minutes
Belgium
Website: http://www.parlement.brussels/search_form_fr/ Finland
Keyword: intelligence artificielle Website: https://www.eduskunta.fi/
Filter Keyword: tiedot
• Document Type: all Filter:
• Parliamentary Affairs and Documents
Brazil • Public document: Minutes
Website: https://www2.camara.leg.br/atividade- • Actor: Plenary sessions
legislativa/discursos-e-notas-taquigraficas
Keyword: inteligência artificial France
Filter: Website: https://www.assemblee-nationale.fr/
• Federal legislation Keyword: intelligence artificielle
• Type: Law Filter:
• Reports of the debates in session
Canada Note: Such documents were only prepared starting in
Website: https://www.ourcommons.ca/PublicationSearch/ 2017.
en/?PubType=37
Keyword: artificial Intelligence Germany
Website: https://dip.bundestag.de/
China Keyword: künstliche Intelligenz
Website: Various reports on the work of the government Filter:
Keyword: 人工智能 • Speeches, requests to speak in the plenum
Note: The National People’s Congress is held once per
year and does not provide full legislative proceedings. India
Hence, the counts included in the analysis only searched Website: http://loksabhaph.nic.in/
the mentions of artificial intelligence in the only public Keyword: artificial intelligence
document released from the Congress meetings, the Filter:
Report on the Work of the Government, delivered by the • Exact word/phrase
Premier.
Ireland
Website: https://www.oireachtas.ie/
Filter: Content of parliamentary debates

Italy Singapore
Website: https://aic.camera.it/aic/search.html Website: https://sprs.parl.gov.sg/search/home
Keyword: intelligenza artificiale Keyword: artificial intelligence
Filter:
• Type: All South Africa
• Search by exact phrase Website: https://www.parliament.gov.za/hansard
Japan Note: This search function does not search within the
Website: https://kokkai.ndl.go.jp/#/ context of the full text and so no results were returned.
Keyword: 人工知能 Therefore, a Google search using the “site” function
Filter: to search https://www.parliament.gov.za/storage/
• Full text app/media/Docs/hansard/ with the keyword “artificial
• Law intelligence” is conducted.
Netherlands South Korea

Website: https://www.tweedekamer.nl/kamerstukken?pk_ Website: http://likms.assembly.go.kr/
campaign=breadcrumb Keyword: 인공 지능
Keyword: kunstmatige intelligentie Filter:
Filter: • Meeting Type: All
• Parliamentary papers - Plenary reports
Spain
New Zealand Website: https://www.congreso.es/
Website: https://www.parliament.nz/en/pb/hansard- Keyword: inteligencia artificial
debates/ Filter:
Keyword: artificial intelligence • Official publications of parliamentary proceedings
Norway Switzerland
Website: https://www.stortinget.no/no/Saker-og- Website: https://www.parlament.ch/
publikasjoner/Publikasjoner/Referater/ Keyword: intelligence artificielle
Keyword: kunstig intelligens Filter:
Note: This search function does not directly allow the • Parliamentary proceedings
keyword within minutes. Therefore, a Google search using
the “site” function to search the site with the keyword of Sweden
“artificial intelligence” is conducted. Website: https://www.riksdagen.se/sv/global/
sok/?q=&doktyp=prot
Russia Keyword: artificiell intelligens
Website: http://transcript.duma.gov.ru/ Filter:
Keyword: искусственный интеллект • Minutes
Filter:
• Words in text

United Kingdom Studies, Council on Foreign Relations, Heritage

https://hansard.parliament.uk/ Foundation, Hudson Institute, MacroPolo, National
Keyword: artificial intelligence Security Institute, New America Foundation, RAND
Filter Corporation, Rockefeller Foundation, Stimson
• References Center, Urban Institute, Wilson Center
• University Institutes & Research Programs: AI and
United States Humanity Cornell University; AI Now Institute, New
Website: https://www.congress.gov/ York University; AI Pulse, UCLA Law; Belfer Center for
Keyword: artificial intelligence Science and International Affairs, Harvard University;
Filter: Berkman Klein Center, Harvard University; Center for
• Source: Congressional record Information Technology Policy, Princeton University;
• Congressional record section: Senate, House of Center for Long-Term Cybersecurity, UC Berkeley;
Representatives, and Extensions of Remarks Center for Security and Emerging Technology,
Georgetown University; CITRUS Policy Lab, UC
Berkeley; Hoover Institution; Institute for Human-
U.S. A I P O L I CY PA P E R S Centered Artificial Intelligence, Stanford University;
Organizations Internet Policy Research Initiative, Massachusetts
To develop a more nuanced understanding of the thought Institute of Technology; MIT Lincoln Laboratory;
leadership that motivates AI policy, we tracked policy Princeton School of Public and International Affairs
papers published by 55 organizations in the United States
or with a strong presence in the United States (expanded Methodology
from the list of 36 organizations last year) across four Each broad topic area is based on a collection of
broad categories: underlying keywords that describe the content of the
• Civil Society, Associations & Consortiums: specific paper. We included 17 topics that represented the
Algorithmic Justice League, Alliance for Artificial majority of discourse related to AI between 2018-2021.
Intelligence in Healthcare, Amnesty International, These topic areas and the associated keywords are listed
EFF, Future of Privacy Forum, Human Rights Watch, below:
IJIS Institute, Institute for Electrical and Electronics • Health & Biological Sciences: medicine, healthcare
Engineers, Partnership on AI systems, drug discovery, care, biomedical research,
• Consultancy: Accenture, Bain & Company, Boston insurance, health behaviors, COVID-19, global health
Consulting Group, Deloitte, McKinsey & Company • Physical Sciences: chemistry, physics, astronomy,
• Government Agencies: Congressional Research earth science
Service, Library of Congress, Defense Technical • Energy & Environment: energy costs, climate change,
Information Center, Government Accountability energy markets, pollution, conservation, oil and gas,
Office, Pentagon Library alternative energy
• Private Sector Companies: Google AI, Microsoft AI, • International Affairs & International Security:
Nvidia, OpenAI international relations, international trade,
• Think Tanks & Policy Institutes: American Enterprise developing countries, humanitarian assistance,
Institute, Aspen Institute, Atlantic Council, Brookings warfare, regional security, national security,
Institute, Carnegie Endowment for International autonomous weapons
Peace, Cato Institute, Center for a New American • Justice & Law Enforcement: civil justice, criminal
Security, Center for Strategic and International justice, social justice, police, public safety, courts

• Communications & Media: social media,

disinformation, media markets, deepfakes
• Government & Public Administration: federal
government, state government, local government,
public sector efficiency, public sector effectiveness,
government services, government benefits,
government programs, public works, public
transportation
• Democracy: elections, rights, freedoms, liberties,
personal freedoms
• Industry & Regulation: economy, antitrust, M&A,
competition, finance, management, supply chain,
telecom, economic regulation, technical standards,
autonomous vehicle industry and regulation
• Innovation & Technology: advancements and
improvements in AI technology, R&D, intellectual
property, patents, entrepreneurship, innovation
ecosystems, startups, computer science, engineering
• Education & Skills: early childhood, K-12, higher
education, STEM, schools, classrooms, reskilling
• Workforce & Labor: labor supply and demand, talent,
immigration, migration, personnel economics, future
of work
• Social & Behavioral Sciences: sociology, linguistics,
anthropology, ethnic studies, demography,
geography, psychology, cognitive science
• Humanities: arts, music, literature, language,
performance, theater, classics, history, philosophy,
religion, cultural studies
• Equity & Inclusion: biases, discrimination, gender,
race, socioeconomic inequality, disabilities,
vulnerable populations
• Privacy, Safety & Security: anonymity, GDPR,
consumer protection, physical safety, human
control, cybersecurity, encryption, hacking
• Ethics: transparency, accountability, human
values, human rights, sustainability, explainability,
interpretability, decision-making norms

Artificial
Intelligence
Index Report 2022

2022 AI Index Report Master

Uploaded by

Copyright:

Available Formats

2022 AI Index Report Master

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2022 AI Index Report Master

Uploaded by

Copyright:

Available Formats

Artificial Intelligence

Index Report 2022

Jack Clark and Ray Perrault

U.S. and China dominated cross-country collaborations on AI:

The rise of AI ethics everywhere:

AI becomes more affordable and higher performing:

Data, data, data:

More global legislation on AI than ever:

Robotic arms are becoming cheaper:

Staff and Researchers

Latisha Harry Ellie Sakhaee

How to Cite This Report

The AI Index 2022 Annual Report by Stanford University is licensed under

Public Data and Tools

AI Index and Stanford HAI

We welcome feedback and new ideas for next year.

Analytics and Research Partners

Research and Development

The Economy and Education

AI Policy and Governance

Global AI Vibrancy Tool

Center for Security and Emerging LinkedIn

Women in Machine Learning

CHAPTER 1 Research and Development 13

CHAPTER 2 Technical Performance 47

CHAPTER 3 Technical AI Ethics 100

CHAPTER 4 The Economy and Education 139

CHAPTER 5 AI Policy and Governance 172

ACCESS THE PUBLIC DATA

Overview 24 Conference Attendance 41

By Region 25 Women in Machine Learning (WiML)

ACCESS THE PUBLIC DATA

Table of Contents Chapter 1 Preview 15

Table of Contents Chapter 1 Preview 16

NUMBER of AI PUBLICATIONS in the WORLD, 2010–21

Table of Contents Chapter 1 Preview 17

By Type of Publication Books, book chapters, theses, and unknown document

NUMBER of AI PUBLICATIONS by TYPE, 2010–21

180 172.11, Journal

20.60, Book Chapter

Table of Contents Chapter 1 Preview 18

By Field of Study Other areas strongly influenced by deep learning, such

NUMBER of AI PUBLICATIONS by FIELD of STUDY (EXCLUDING OTHER AI), 2010–21

51.69, Pattern Recognition

39.93, Machine Learning

24.80, Computer Vision

10 9.70, Human–Computer Interaction

Table of Contents Chapter 1 Preview 19

By Sector United Kingdom (Figure 1.1.4d).3 The education sector

AI PUBLICATIONS (% of TOTAL) by SECTOR, 2010–21

60% 59.58%, Education

20% 20.77%, Unknown

AI PUBLICATIONS in UNITED STATES (% of TOTAL) by SECTOR, 2010–21

Table of Contents Chapter 1 Preview 20

AI PUBLICATIONS in CHINA (% of TOTAL) by SECTOR, 2010–21

60% 60.24%, Education

AI PUBLICATIONS in EUROPEAN UNION and UNITED KINGDOM (% of TOTAL) by SECTOR, 2010–21

20% 19.76%, Unknown

Table of Contents Chapter 1 Preview 21

By far, the greatest number of collaborations in the past