Advance Davos Draft - Not for Public Circulation 1

Generally Faster
The Economic Impact of Generative AI
April 25th, 2024

Andrew McAfee
2023 - 2024 Visiting Fellow / Technology & Society at Google
Table of contents

Executive Summary 1
Introduction 2
Generative AI is a general-purpose technology 4
Faster This Time 6
Implications 9
Open Questions 17
Acknowledgements 21
References 22

Andrew McAfee is the Co-Founder and Co-Director of MIT’s Initiative on the Digital Economy and a
Principal Research Scientist at the MIT Sloan School of Management.

McAfee currently holds the inaugural Technology & Society Visiting Fellowship at Google. The
fellowship invites renowned experts to produce original research or perspectives on technology’s
impact to cross-cutting societal themes like economic outcomes, sustainability, or the future of work
and education. Throughout their fellowship, researchers engage with experts at Google across a
variety of fields, including artificial intelligence, economics, social sciences, and public policy.

The opinions expressed in this report are solely those of the author.
Executive Summary
Generative AI is one of the rare technologies powerful enough to accelerate overall economic
growth---what economists call a “general-purpose technology.” These innovations have the
potential to positively transform economies and societies.

By one estimate, close to 80% of the jobs in the U.S. economy could see at least 10% of their tasks
done twice as quickly (with no loss in quality) via the use of generative AI.

Previous general-purpose technologies like the steam engine and electrification have brought
their changes over decades. However, we anticipate that generative AI’s effects will be felt more
quickly due to its ease of diffusion.

This technology is already delivering large productivity gains, which will increase and spread as
people and organizations come up with complementary innovations that leverage generative AI’s
capabilities. As a result, economic growth will speed up.

In addition to faster growth, generative AI will bring other changes. It will reduce demand for some
skills, increase demand for others, and create demand for entirely new ones.

Fears of large-scale technological unemployment are probably overblown. The history of

general-purpose technologies shows that the growth they bring is accompanied by strong
demand for labor.

However, this increased demand is often in new occupations. For example, more than 85% of total
U.S. employment growth since 1940 has come in entirely new occupations.

The rapid changes brought by the spread of generative AI will require prompt and effective
reskilling efforts. These efforts will be able to draw on generative AI itself, a tool with the unique
ability to help people learn how to use it better. And because Generative AI accumulates
knowledge and makes it available on demand, it’s particularly effective at improving the
performance of entry-level employees, helping with wage inequality…

Previous general-purpose technologies have resulted in changes to the companies and countries
leading the way in different industries. We believe that generative AI will be similarly powerful.

1
Introduction
Since the Industrial Revolution a handful of technologies have been powerful enough to
accelerate the overall course of economic progress. These “general-purpose technologies”
include the steam engine, the internal combustion engine, electrification, and the computer.

In this report we make the case that generative AI is a new general-purpose technology, and
one that may spread more quickly than its predecessors. Generative AI is a recently developed
type of artificial intelligence capable of producing new and original content such as text,
images, videos, and audio. It accomplishes this by learning statistical patterns from existing
data, then using these patterns to generate new and novel outputs upon request.

The arrival of generative AI is a significant development because, as the physicist Freeman

Dyson put it, technology “is the mother of civilizations, of arts and sciences.”
General-purpose technologies not only boost productivity and economic growth, but also
contribute to many other kinds of advancement. We anticipate that generative AI will speed up
scientific discovery, help innovators and engineers build better, and give creative people new
ways to express themselves and move their audiences.

Of course we recognize that new technologies bring challenges as well as benefits. The
internal combustion engine, for example, polluted the atmosphere, made conflict more lethal,
and led to massive shifts in employment as workers moved from farms to factories.

But the long-run overall effect of tech progress has been hugely positive. In recent decades
working hours have dropped while standards of living have improved around the world.
Lifespans have increased and the burden of disease has eased. And key environmental
indicators have started to move in the right direction. Technological progress is a fundamental
reason for these and many other advances.

2
Exhibit 1

Health and Working hours

Average annual working hours across OECD countries
wealth have declined more than 10% since 1970

increased Standards of living

Global median income per day has increased by about
alongside 150% since 1990

technological Lifespan
progress In 1900 the global average life expectancy of a newborn
baby was 32 years. By 2021 this more than doubled to 71
years

Reduced global disease burden

Lifespan years lost to premature death and disability
decreased by about 35% since 1990. In 1990, the disease
burden in Bangladesh was 150% higher than in Belgium.
By 2019, it was only 40% higher.

Positive trends in environmental

indicators:
Air pollution deaths have declined globally by 45% since
1990. In 2000 61% of people around the world had
access to clean water. By 2022 this had increased to 72%.

Previous general-purpose technologies have diffused relatively slowly. Decades have elapsed
between their introduction and their economy-wide impact on productivity and other
important outcomes. We believe, though, that generative AI’s impact will be felt more quickly;
its transformative effects will start to be felt this decade. This report explains our reasoning.

We begin by making the case that generative AI is in fact a general-purpose technology. We

then explain why we believe that it will spread and boost economic growth more quickly than
its predecessors did. Next, we explore implications of this fast-moving technology in four key
areas: economic growth; skills, jobs, and wages; business transformation; and novel risks and
harms. We conclude by highlighting important open questions about unlocking generative AI’s
potential; technological unemployment; social support for a fast workforce transition; and
national competitiveness.

We caution that our findings and conclusions here are preliminary. Generative AI’s first
deployments are no more than a few years old. It is far too soon to say with confidence how it
will affect entire economies and societies. But the long history of previous general-purpose
technologies, combined with generative AI’s short but impressive history, gives us a base from
which to look ahead.

3
Generative AI is a general-purpose
technology
How can we know whether generative AI is in fact a general-purpose technology? It might
seem as though that question can be answered only in retrospect, by waiting to see if
productivity increases significantly as it spreads. But economists believe that such
technologies can be recognized in advance.

If a technology possesses three key characteristics — rapid improvement, pervasiveness, and

complementary innovation — it’s likely to have a large, positive, economy-wide impact on
productivity growth. In other words, it’s likely to be a general-purpose technology. Together,
the trio ensure that a technology will not only be used in many sectors, but also transform
them. James Watt’s steam engine, for example, was originally used to pump water out of coal
mines. But as it improved, it replaced the water wheels that powered factories and the sails
that moved ships around the world. It also sparked the complementary innovation of the
locomotive, thereby transforming land transportation. The steam engine, in short, kicked off
the Industrial Revolution.

Even though generative AI is a recent development, there’s already strong evidence that, like
the steam engine, it possesses all three characteristics of a general-purpose technology.

Rapid Improvement
Generative AI has improved with remarkable speed at its core task of generating relevant and
accurate content in response to prompts from users. As recently as 2019, the response from a
state-of-the-art system to the prompt: “the best thing about AI is its ability to…?” was
grammatical, but nonsensical:

“The best thing about AI is its ability to see through, and make sense of, the world around us
rather than panicking and ignoring. This is known as AI “doing its job” or AI “run of the mill”...

Just a year later, though, the next release of the same system responded:

“The best thing about AI is its ability to learn and develop over time, allowing it to continually
improve its performance and be more efficient at tasks. AI can also be used to automate
mundane tasks, allowing humans to focus on more important tasks.”

Generative AI has also improved quickly at real-world tasks. OpenAI’s GPT 3.5 system,
released in late 2022, performed better on a version of the U.S. bar exam than approximately
10% of human test takers. GPT 4, released in March of 2023, performed better than 90%.

In September of 2020 a team of researchers proposed the Massive Multitask Language

Understanding (MMLU) benchmark, a test designed to assess generative AI systems’
problem-solving abilities and knowledge of the world. It “cover[ed] 57 tasks including

4
elementary mathematics, U.S. history, computer science, law, and more.” The answers
provided by most available systems were only about as accurate as guesses when the MMLU
was published. In December of 2023, however, Google Gemini Ultra became the first
generative AI system to out-perform expert humans on the MMLU, scoring 90% across the
tests’ subject areas.

A final way to see the rapid improvement in generative AI systems is to look at the size of their
“context windows,” or how much information they can accept from users. If, for example, a
user wanted generative AI to summarize or rewrite a report, the report would need to be
included in the context window along with the request. In 2020 state-of-the-art systems had a
context window that could accommodate about 7 and a half pages of text. By late 2023 the
window was 40 times larger and was able to accept about 300 pages of text.

Exhibit 2

Generative AI U.S. bar exam

In less than a year, OpenAI’s GPT models went from
has improved outperforming 10% of people taking the bar exam to
outperforming 90%
quickly in
Problem-solving and knowledge
recent years Three years after the creation of the multi-task MMLU
benchmark, Google’s Gemini Ultra became the first
generative AI system to outperform expert humans,
scoring 90% across test subject areas

Context windows
In three years, state-of-the-art generative AI systems
grew their ability to accept text inputs by 40X, moving
from the equivalent of 7.5 pages to 300 pages

Pervasiveness
A technology that can generate many different kinds of language is likely to become
pervasive. Every industry and profession, after all, relies heavily on communication.
Business leaders recognize this fact, and believe that generative AI will diffuse widely. In a
2023 survey of U.S. executives, almost two-thirds of respondents felt that the technology
would have a “high” or “extremely high” impact on their organizations.

Recent research agrees. A 2023 study examined all the tasks done by workers throughout
the American economy to see which of them could be done at least at least twice as rapidly
with no loss in quality via the use of generative AI. The research concluded that for about
20% of all workers, half or more of their tasks fell into this category. When the threshold
was reduced to 10% of tasks, 80% of workers qualified. For example, interpreters and
translators, survey researchers, and public relations specialists all had at least two-thirds of
their tasks eligible for significant productivity improvement via generative AI. At the other

5
end of the spectrum, workers including short-order cooks, athletes, and oil and gas derrick
operators had no tasks in this category.

Complementary Innovations
Generative AI is already being applied to do more than generate text, pictures, and sound.
A team at Google DeepMind, for example, has combined the technology with a robot and a
machine learning system trained to recognize objects in images. The project’s goal is to
allow a person to control a robot not by specifying a sequence of steps for it to perform,
but instead by simply telling it the task to be accomplished, like “pick up the bag about to
fall off the table.” In addition to successfully accomplishing such tasks, early results
indicate that this system can also do basic reasoning. When given the prompt “I need to
hammer a nail, what object from the scene might be useful?” the robotic arm picked up a
rock instead of a piece of paper.

Generative AI is also being used not just to improve individual tasks but also to streamline
and improve entire processes. A team of business school professors building a web-based
game about entrepreneurship used the technology to suggest and incorporate
improvements to the game. Instead of collecting feedback from human users, the team
instead asked a generative AI system to assume the role of an MBA student playing the
game and recommend changes. The same system then created prototype web pages that
incorporated these changes. During the meeting to discuss these prototypes, team
members asked for refinements. These were created on the fly by the technology, which
also recorded, transcribed, and summarized the meeting and gave all participants their
action items. A process that used to take 1-2 weeks was reduced to one day.

Early efforts give us confidence that innovations making use of generative AI’s capabilities
will accelerate science and engineering. For example, the technology is already being used
to identify new materials with desired properties (like high magnetism and electrical
conductivity or low cost and supply-chain risk), which complements alternative AI efforts to
develop new materials. This work is obviously much more complicated than asking a
generative AI system for a picture of a cat, then refining the result by asking for tabby
stripes or a longer tail, but it relies on many of the same principles.

In summary, because of generative AI’s rapid improvement, pervasiveness, and clear

potential for complementary innovation, we are confident that it merits the label of
“general-purpose technology.” We now turn to the question of how quickly its impact will
be felt.

Faster This Time

Previous general-purpose technologies have taken decades to boost entire economies. The
electric motor, for example, was patented in the U.S. in 1837 and Edison’s first power station
began operating in 1881. However, electrification only began to have a significant positive
effect on American factories’ productivity in the 1920s.

Change was slow because these technologies diffused slowly, often requiring new
infrastructure such as roads and electrical transmission networks. In addition, their biggest

6
benefits only appeared once users had time to come up with complementary innovations
and put them in place. For example, it took many years for manufacturers to realize that
electrification enabled conveyor belts, overhead cranes, and assembly lines, and that these
innovations — especially when combined — would boost productivity so much that they
justified replacing existing equipment and rethinking the factory.

There are grounds to believe, however, that generative AI’s effects will appear more quickly
than was the case with earlier general-purpose technologies. For one thing, much of the
required infrastructure is already in place. Once new generative AI systems are developed
they can be deployed around the world as quickly as web pages and apps can. A large and
growing number of powerful applications using this technology are immediately available at
no cost to anyone with an Internet-connected device. Others are available by subscription.
This wide availability applies to both end users of the technology and developers who want
to build new tools with it. Furthermore, as generative AI continues its rapid improvement,
many of these improvements will propagate globally as soon as they’re released.

The work of building complementary digital innovations for this general-purpose

technology will probably also be fast. For most of the computer era systems, integration
projects were long and expensive, but this is starting to change. Much modern software is
modular; it has standard interfaces that allow components to be easily combined and
recombined, as Lego bricks are. This means that generative AI will quickly be combined
with other types of software, deployed over the internet, and available on demand.

Another reason to expect this technology to spread quickly is that it is an easy one for
people to start working with. They just talk to it. Most of generative AI’s users don’t have to
master a new user interface or programming language; they instead use natural human
language. It requires time and practice to become proficient at interacting with generative
AI, but it doesn’t require many “computer skills.” The technology, in short, is immediately
available to people and quickly useful to them.

This easy access contributes to pervasive use. A 2023 survey of 14,000 users across a
range of professions in 14 countries found 28% of respondents already using generative AI
at work, even without much support from their organizations; of generative AI’s users, 55%
had worked with unapproved tools and 40% with banned tools. 71% of users said the
technology made them more productive, and 58% reported feeling more engaged because
of it. In some professions, generative AI use is already almost universal: by June of 2023,
92% of programmers surveyed were using it at work.

But perhaps the best reason to be confident that generative AI will bring its economic
benefits quickly is that it’s already doing so. Studies have been conducted on the
technology’s performance impact in a range of professional settings. Across this research,
the conclusion is clear and consistent: generative AI brings large and rapid productivity
boosts.

7
Exhibit 3

AI brings large Call center employees

Using a generative AI assistant resolved 14%
and fast more issues per hour, in addition to having lower
employee turnover and higher customer
productivity satisfaction. These benefits began appearing
within a month of the technology’s deployment
boosts across
Management consultants
occupations Using a generative AI system completed 18 tasks
25% faster with, on average, 40% higher quality

Coders
Using an AI pair programmer completed
programming tasks 56% faster

Professional writers
Using a generative AI chatbot completed
occupation-specific writing tasks 37% faster with
significantly higher quality

Physicians
Completed daily patient notes 83% faster thanks
to automated transcription and summarization

In all of these cases, generative AI improved productivity on individual tasks. We stress,

though, that this technology’s greatest benefits will come not at the task level, but instead as
innovators reimagine processes and entire organizations to take advantage of the
technology’s benefits. This reimagining is just starting. We anticipate that it will be relatively
quick by historical standards but no less transformative than was the case with previous
general-purpose technologies.

Investors also believe that this will be the case. Since late 2022, U.S. public companies with
workforces that are more exposed to generative AI have seen their stock market valuations
increase relative to less-exposed firms. More exposed firms have more opportunities to
improve via the technology and are being valued more highly as a result.

Perhaps the boldest prediction about the economic impact of generative AI: it will soon boost
productivity even for companies and industries that have lagged in realizing the benefits of
the digital era. In recent years, smaller companies have not kept pace with the technology
investments of larger ones. They’ve also fallen behind in both innovation and productivity. We
expect that because of its low cost and ease of adoption, generative AI will help level the
playing field between small and large firms.

8
For similar reasons, we expect that this technology will also speed things up in industries that
have been slowing down. In the US, for example, important sectors like durable goods
manufacturing, wholesale and retail trade, and transportation have seen their rate of
productivity growth slow down during the 21st century. This has been the case even as the
smartphone, machine learning, and other potent technologies have diffused. It is optimistic,
then, to predict that generative AI could help reverse this slowdown. But that is what we
foresee based on generative AI’s unique combination of versatility, utility, ease of use, and
diffusion.

Implications
What are the economic implications of a powerful technology that quickly diffuses and
becomes pervasive throughout an economy? The most obvious one is faster economic
growth. Along with that growth will come shifts in both the workforce and the competitive
landscape. Finally, generative AI may bring novel risks and harms as it spreads.

A New Era of Economic Growth

General-purpose technologies are powerful engines of growth, and generative AI is no
exception. Goldman Sachs estimates that over the next decade the technology will be
responsible for a 0.4 percentage point increase in GDP growth in the U.S., 0.3 points in other
developed markets, and 0.2 points in emerging markets.

Discussions about economic growth are often dry, but economic growth itself is vital. It’s a
wellspring of wellbeing. Greater per capita GDP is strongly associated with better health,
higher life satisfaction, and other positive outcomes at the national level. It’s also good news
for our planet. Low-income countries are relatively heavy polluters; they can’t afford
cutting-edge green technologies or expensive waste treatment systems. More affluent
countries can, and these deeper pockets enable them to spend on the environment. As Indian
Prime Minister Indira Gandhi put it in 1972, “are not poverty and need the greatest polluters?”
As low-income countries see continued economic growth, then, we should expect their
pollution to decline.

9
Exhibit 4

Both life satisfaction and life expectancy trend upward with a nation’s wealth
Self-reported life satisfaction vs. GDP per capita, 2022
Self-reported life satisfaction is measured on a scale ranging from 0-10, where 10 is the highest possible life satisfaction. GDP per
capita is adjusted for inflation and differences in the cost of living between countries. (OurWorldInData.org)

Life expectancy vs. GDP per capita, 2018

The period life expectancy at birth, in a given year. GDP per capita is measured in 2017 international dollars, which adjusts for
inflation and cross-country price differences.(OurWorldInData.org)

10
Exhibit 5

Pollution severity tends to decrease as countries become wealthier

Death rate from outdoor air pollution vs. GDP per capita, 2019
Death rates are measured as the number of premature deaths attributed to outdoor particulate matter air pollution per 100,000
individuals. Gross domestic product (GDP) per capita is measured in constant international-$. (OurWorldInData.org)

Furthermore, generative AI brings benefits that aren’t captured in growth statistics. As it

automates tasks, it frees up people to do work that’s more valuable in every sense. Within
health care, for example, it’s enabling physicians to spend less time on lower-value
administrative tasks and more time using their advanced training and caring for patients.
Already, generative AI deployments are putting hours back in physicians’ days by taking over
some of their paperwork, allowing them to “be 100% present for [their] patients.”

Examining generative AI’s impact on healthcare reveals that the technology does more than
improve productivity; it also improves quality. A 2023 study found that a generative AI system
optimized for medical diagnosis was more accurate than unassisted physicians at diagnosing
difficult cases. Physicians’ accuracy improved significantly once they could consult the
system; generative AI improved their diagnostic abilities more than did access to search
engines and other resources.

A final reason to be optimistic about generative AI as an engine of growth and betterment is its
ability to help people acquire new skills. One of the most common findings from research on
this technology is that it most improves the performance of the newest and least-skilled
workers. These are the people who need to tap into prior accumulated knowledge and apply it.
Generative AI is well suited for this purpose. It generates helpful answers to many questions
people have as they’re going about their work. As such, it’s a powerful and broadly useful
upskilling technology, particularly for those with the most questions.

11
 Exhibit 6

Generative AI Management consultants

performing in the bottom half of their cohort saw a
seems to be a 43% improvement when using a generative AI tool;
those in the top half only improved by 17%
powerful
Professional writers
upskilling initial performance inequalities were more than 50%
erased once low- and high-performing cohorts used
technology a generative AI chatbot

Call center employees

with the lowest historical performance increased
their call resolution rate by 35% when using a
generative AI assistant; the highest performing
cohort saw no improvements

Law students
with the lowest performance improved their final
exam scores by 45 percentiles when using a
generative AI system

Shifts In Skills, Jobs, and Income

General-purpose technologies change the world of work. They reduce demand for some
skills, increase demand for others, and create demand for entirely new ones. Across
general-purpose technologies, a common pattern has been “skill-biased technological
change;” over time, workers need more of the kinds of skills delivered by education. In the
late 19th century, for example, dockworkers needed strength and endurance but not much
schooling. A century later they needed computer skills, not muscle power, to unload ships full
of containers at highly automated ports.

In the era of generative AI, the landscape of valuable skills will once again change. Current
technologies have demonstrated expert-level performance at tasks requiring creativity,
analysis, problem solving, persuasion, summarization, and other advanced skills. Some highly
educated and well paid workers who make a living via these skills today will have to shift what
they do in order to maintain or increase their value in the labor market.

But generative AI won’t necessarily automate all of the tasks where it performs well. Initial
research indicates that top performance comes as people continually interact with the
technology. Mastering this interaction will be a valuable skill for people in many professions,
even those at the highest levels of skill and education. Translators, for example, often no
longer begin their work by translating an entire text themselves. They instead start with a
machine-generated translation, then audit, refine, and improve it.

12
As general-purpose technologies can do more and more, one might expect that human labor
will be needed to do less and less. And sure enough, in several advanced economies the
number of people employed in manufacturing and agriculture has declined in recent decades
even as the output of those industries has increased.

Yet those economies have not experienced growing joblessness. In fact, the opposite has
happened: growing adoption of technology has resulted in unemployment at or near historic
lows. Fewer workers are needed on farms and factories, but more are needed in restaurants,
hotels, gyms, insurance companies, warehouses, airlines, and many other types of
enterprises that make up a modern economy. Economic growth, in short, creates more jobs in
large part by creating new jobs. A 2023 analysis concluded that more than 85% of total U.S.
employment growth since 1940 has occurred within occupations that did not exist in that
year.

This job creation engine has been powerful in recent years, to the point that demand for
workers often exceeds supply. Labor markets remain tight in most OECD countries, including
the U.S.. In late 2023, if every unemployed American worker were placed, the country would
still have more than 3 million open jobs.

In Europe, recent analyses have found that hundreds of occupations were experiencing acute
shortages. Many organizations in OECD countries now cite skill gaps and an inability to
attract talent as key barriers to accelerating their businesses. Many of these challenges will
become more acute in the coming years as today’s “baby boomers” age out of the
workforce.

13
 Exhibit 7

Insufficient Tight labor markets

As of October 2023, the unemployment rate in the
labor supply is OECD, the European Union, and the euro area was
at or within 0.1 percentage point of its record lows;
already and at 0.5 percentage points of its low for the U.S.

creating European occupational shortages

40 occupations received ‘severe shortage’
challenges: designations from 11 or more European countries in
2022: a handful of these, like software developers
these may and application programmers, are directly relevant

deepen for generative AI

U.S. labor shortages

In late 2023, if every unemployed American worker
were placed, the country would still have more than
~3 million open jobs

Future labor force challenges

Projections suggest U.S. workforce participation
will be down to about 60% in 2032 driven by exits
of Americans aged 65 and older. In 2002, workforce
participation was nearly 67%

Global workforce aging

By 2050, individuals aged 65+ will be >2X the
number of children under age 5 and the same as
the number of children under age 12, or 16% of the
global population.

Workforce aging by region

By 2050, individuals aged 65+ could be 25% of
Europe and North America; 19% of Latin America
and the Caribbean; and 26% of Eastern and
South-Eastern Asia.

Previous general-purpose technologies did not cause massive technological unemployment,

but there’s good evidence that they did contribute to the “hollowing out” of the labor force:
more low- and high-income jobs created, and fewer middle-income ones. General-purpose
technologies played a role here by taking over much of the routine physical and cognitive work
done by middle-class workers. Many automotive companies, for example, found that as
automation advanced, they needed fewer assembly line workers (as robots took over tasks like
painting car bodies) and payroll clerks (as information systems calculated paychecks and
directly deposited them into workers’ bank accounts).

14
If generative AI contributes to this hollowing out, it might do so among higher-paying jobs.
Research about the suitability of this technology for all the work done in a modern economy
finds that generative AI affects more of the tasks done by college graduates than by
high-school only graduates. Professional writers, for example, are usually college graduates.
On one popular online freelancing platform, both monthly income and number of jobs for
people offering “writing-related services” declined in 2023 as online generative AI became
widely available.

Meanwhile, workers without college educations have seen both job opportunities and wages
rise quickly in recent years — quickly enough to reverse a decades-long trend of growing
income inequality. Between 1980 and 2020, the income gap between the top 10% and bottom
10% of U.S. workers rose sharply. Since 2020, however, almost 40% of that rise has been
reversed because of wage increases at the bottom.

The faster generative AI spreads throughout economies, the faster it will change skills, jobs,
and wages. And if these changes are both deep and fast, they will bring novel challenges.
Societies have had decades to respond to the workforce shifts caused by previous
general-purpose technologies. In 1900, for example, 40% of the U.S. workforce was in
agriculture and 40% of the population lived on farms. Four decades later, both figures had
been cut in half. However large the stresses caused by this societal shift, they would have
been larger if it happened in five years instead of forty. In a later section of this report we pose
the question of what social support for a fast workforce transition might look like.

Business Transformation and Competitiveness

It’s clear that generative AI will transform many if not most businesses and industries, but it’s
far from clear how. This uncertainty stems partly from the fact that the future evolution of the
technology is unclear. It’s also due to the unpredictability of the complementary innovations
that accompany general-purpose technologies.

At the dawn of electrification, few people would have correctly predicted what factories would
eventually look like. Similarly, it’s difficult to foresee all the changes generative AI will bring to
the world of work. We know that it’s already improving the productivity and quality of many
tasks. We also see that the technology is beginning to be used to redesign multi-step,
multi-group processes, making them faster and less labor-intensive.

Task-level and process-level innovation with generative AI will continue and expand. But this
technology’s deepest impact on the world of work will come as it’s used to reimagine entire
organizations. This deep reimagination will be a decentralized and distributed phenomenon,
carried out by innovators and entrepreneurs throughout the economy.

Because of their talent pools, and access to customers, and large amounts of data, today’s
large incumbent companies might seem to be well-positioned to lead the transformation of
work in the era of generative AI. History, however, reveals a different pattern: the companies
on top at the start of previous general-purpose technology-enabled transformations did not
remain on top. Instead, a great deal of disruption occurred, with new entrants appearing and
eventually dominating.

This disruption did not happen because incumbents were unaware of the new technology’s
power. They’re usually keenly aware, in fact, and generative AI is no exception. A recent survey

15
of business executives found that more than half of respondents expected this technology to
be moderately or severely disruptive to their industries. In the past, though, such executives
were often unable to envision how much a new general-purpose technology could improve on
the status quo, or unwilling to make large enough changes to their successful businesses. As a
result, they fell behind bold upstarts.

We see evidence that a similar pattern of disruption is now playing out in the era of AI. As the
digital transformation of the business world has progressed over the course of the 21st
century, “superstar firms” have appeared in many countries and industries. These superstars
are highly productive compared to their peers and have been gaining market share and
profitability. In the U.S., many of them are relatively young companies.

If generative AI further empowers a small set of superstars and allows them to outpace their
rivals, workforce changes will increase. Fading incumbents will conduct layoffs, and the
number of people who need to find new jobs and acquire new skills will increase. Research
conducted over the last decade reveals that companies investing heavily in machine learning
are not the ones conducting layoffs. Workforce reductions come instead from companies that
did not embrace the technology. We predict that a similar pattern will hold with generative AI.

Novel Risks and Harms

For all their benefits, general-purpose technologies can also introduce new risks and amplify
existing ones: Cybercrime, for example, is expected to incur global costs of $10.5 trillion by
2025. Computers make all of us more productive — including criminals, unfortunately.

Google DeepMind researchers and partners have created a taxonomy classifying the potential
harms from generative AI, along with a discussion of mitigation approaches. Some of these
risks, while notable, are outside the economic scope of this report. For example, generative
AI’s ability to increase the volume and spread of disinformation could also make it easier for
malicious actors to conduct personalized scams at scale, manipulate financial markets, and
effectively sway public political opinions and actions.

While the risk of workforce disruption is a focus of this report, there are also other risks with
economic implications. For example, most of today’s generative AI models are trained heavily
in English and may perform less well in other languages. Models that capture the language of
one group with less accuracy – or not at all – may drive inequities by reducing
underrepresented cohorts’ abilities to develop or access generative AI applications. If done
thoughtfully, however, the technology could accomplish the opposite outcome by helping
speakers of “long-tail languages” access knowledge that was previously available only in
English.

In the late 19th century, public fear over electricity’s novel risks – driven by multiple incidents
of electrocution from unsafe overhead wires – led one New Yorker to call electricity “a fearful
source of death, [...] a constant menace to the lives of our fellow-citizens” and urge a return
to gaslights. Fortunately, the publicity brought to electricity’s risks helped drive reform that
enabled safer use of the technology, like the eventual burying of Manhattan’s overhead wires.

As with electricity, collective awareness of generative AI’s risks, plus frameworks for assessing
them in context, can enable us to enjoy the technology’s benefits while triaging potential

16
harms. We next explore the open questions that a few key risks, including job disruption, make
salient in the near term.

Open Questions

What Needs to Go Right?

Generative AI has all three characteristics of a general-purpose technology: rapid
improvement, pervasiveness, and complementary innovations. This fact alone, however, does
not guarantee that this new technology will be as transformative as the steam engine or
electrification.Those innovations delivered their benefits because governments, companies,
schools and universities, and civil society came together in ways that supported the useful
diffusion of the technology, allowed its benefits to spread widely, and dealt with its negative
consequences.

We will surely need a similar coalition with generative AI to find answers to the questions
posed by this powerful new tool. How will education at all levels need to change? What’s the
right way to approach regulation? How can we best assess and minimize the biases of
generative AI systems, especially ones that are used to make important decisions about, for
example, employment, credit, justice, and health care?

Overconstraining generative AI will reduce its benefits, while underconstraining it will increase
its harms and empower bad actors. The best points of balance between these two outcomes
are not yet clear, in part because the technology is so young. What is clear at this early stage
is that ignoring generative AI — refusing to confront the issues it raises — is a poor strategy. A
better one is to start learning how to responsibly advance our collective use of this new
general-purpose technology.

Will Generative AI Bring Large-Scale Technological

Unemployment?

Is generative AI powerful enough to significantly change the historical relationship between

technological progress and increased demand for human labor? It might seem so, especially
since it can be applied to so many tasks. But the relationship between technological progress
and labor demand is not a simple one. If generative AI can take over a quarter of a job’s tasks,
for example, we shouldn’t automatically expect employment in that profession to fall by 25%.

To see why not, consider the profession of radiology. Beginning about a decade ago, a type of
AI called deep learning quickly improved at the task of detecting patterns within images. Since
that task is a central part of a radiologist’s work, many observers felt that the profession would
become highly automated, and that overall demand for humans would soon decrease. As a
prominent AI researcher put it in 2016, “People should stop training radiologists now. It’s just
completely obvious within five years deep learning is going to do better than radiologists.” And
sure enough, by 2022 more than 200 AI-based radiology imaging products had been
approved for commercial use in the U.S..

17
Yet demand for human radiologists didn’t decline. Between 2016 and 2021, U.S. radiology
employment increased by 3%, and demand in the profession remained strong. By 2022, in fact,
a global shortage of human radiologists loomed. As one academic department chair put it,
“The demand for imaging is outpacing what we’re doing on the training side. The number of
radiologists in the workforce is not growing as fast as [demand].”

It turned out that while deep learning helped with the interpretation of scanned medical
images, it didn’t completely automate it in most cases. What’s more, the technology didn’t
touch many of a radiologist’s other important tasks, such as speaking with patients and
conferring with colleagues on a course of treatment.

As this example shows, there is no straight line between the diffusion of a powerful technology
and unemployment. Generative AI is unlikely to change this fact, at least in the short run,
because of its limitations. It is not yet able to reliably do multi-step work that involves planning,
reasoning, or memory. A 2023 study created hundreds of questions whose solutions involved
this kind of work, then administered it to both people and a top-performing generative AI
system. Human solvers averaged 92% correct answers; generative AI 15%.

Of course, generative AI is improving quickly and intense research is underway at many

organizations to address the technology’s known weaknesses. It is far from clear how
successful this research will be, and how fast it will progress. Will it succeed to the point that
generative AI can do most, if not all, of the tasks that make up many of today’s jobs? If so, will
innovators and entrepreneurs quickly create enough new jobs that still require human labor? It
is too soon to know the answers to these questions with any confidence, but not too soon to
start asking them.

We conclude this discussion of the possibility of large-scale technological unemployment by

observing that there is no shortage of important work to be done in every society. A great deal
of this work, from installing the infrastructure needed to accomplish the global clean energy
transition; to teaching children; to caring for the sick and elderly, can’t be done by the robots
and AI of today, as powerful as they are. We anticipate that this situation will not soon change
and that much of our most important work will remain heavily in the hands of humans.

What Does Social Support for a Fast Transition Look

Like?
If generative AI does in fact spread rapidly around the world and bring large changes to jobs,
wages, and needed skills, then societies will need to respond quickly. This indicates that the
approaches developed during the sudden COVID-19 pandemic to provide short-term income
and wage support might well be effective in helping workers transition to new roles.

In addition to economic support, it will be important to help workers displaced by

technological progress acquire in-demand skills and find new jobs. There are many
approaches to accomplishing this, including apprenticeships and on-the-job training;
vocational schools and university degree programs; and training that leads to certifications
provided by professional groups and companies. Because generative AI is so new and its
effects on the workforce still so uncertain, it’s difficult to know at present how to respond to
the challenges it could bring. We anticipate that the societies most successful at navigating

18
generative AI’s possibly rapid workforce transitions will be those that encourage a range of
responses, learn rapidly how well they’re working, and concentrate on those that deliver
results.

We also anticipate that many of the most successful reskilling efforts will be those that use
generative AI itself. As discussed above, one of the most exciting aspects of the technology is
its ability to accumulate knowledge and deliver it on demand to people who need it. In short,
it’s a tool that can teach its users how to use it. This is an important innovation. Throughout
the digital era, there have been countless attempts to create technologies for knowledge
management. Most of them have underwhelmed or failed. Generative AI seems to be
succeeding, which makes it a key ingredient of any training and reskilling efforts.

Perhaps the most important element of any effort to successfully transition a workforce in an
era of pervasive generative AI is rapid economic growth. Fast growth means greater demand
for workers, greater upward pressure on wages, and greater willingness by employers to reskill
and retrain people. If generative AI brings about a large and fast workforce transition, it will
probably be because it also causes a boost in economic growth. The faster growth will make
the transition easier to accomplish.

Which Countries Will Lead With Generative AI?

History shows that in addition to investing in training and education, the countries best able to
harness the power of general-purpose technologies have a few things in common. They have
market-oriented economies, high levels of infrastructure, educational systems that provide
needed skills, and a well-functioning legal system that enforces contracts and property rights
and establishes clear rules for where and under what circumstances liability may accrue. And
they have accepted some degree of capitalism’s “gale of creative destruction,” preventing old
industries from using legal or regulatory capture to block the emergence of disruptive new
technologies.

These countries are also successful at converting ideas into useful innovations. In a modern
economy, this work often involves government funding for basic research and
research-oriented universities, clear paths to commercialization, and investors with an
appetite for risk. As a rule, the countries that have been most successful with previous
general-purpose technologies are those that found ways to tap into the talents and
aspirations of their people. We anticipate that generative AI will not be an exception.

History also shows that the stakes are high. Some general-purpose technologies have shifted
the global balance of power. The steam engine, which initiated the industrial revolution, was
instrumental in making England the world’s richest country in the 19th century and ensuring
that “the sun never set” on the British Empire. America’s success with the internal combustion
engine and electrification helped it succeed the United Kingdom as the global superpower of
the 20th century.

The countries most successful at developing general-purpose technologies and harnessing

their power have reaped comparatively high and fast-growing standards of living, better
national security, and a greater ability to pursue their interests and influence events around
the world. They are also able to shape the diffusion of technologies in ways that reflect their
goals and values. Countries far from the frontier with general-purpose technologies,
meanwhile, tend also to be farther from these benefits.

19
We share the view that artificial intelligence will be one of the defining technologies of the 21st
century and that generative AI is an especially important addition to the AI portfolio. This
implies that the countries aspiring to global prominence will need to create an environment
that fosters the development and deployment of this technology.

20
Acknowledgements
I sincerely thank the individuals who supported and advanced this research. Their
contributions, in various forms, have been integral to the completion of this work.

Google Executives
Special thanks to James Manyika for sponsoring the Technology & Society Visiting Fellow
program.

Additional thanks to Ruth Porat, Kent Walker, Jeff Dean, and Hal Varian, who, along with James,
engaged on this report multiple times over the course of its creation.

Google Core Contributors

Travis Beals, Guy Ben-Ishai, Luke Garske, Elana Burton.

Google Interviewees
Adam Cohen, Alice Friend, Anoop Sinha, Anthony House, Ashley Zlatinov, Blaise Aguera y
Arcas, Chris Ludwick, Conor Griffin, Diane Tang, Doug Eck, Greg Corrado, Jay Yagnik, Josh
Woodward, Lisa Gevelber, Marian Croak, Mukund Sundararajan, Nick Fox, Nicklas Lundblad,
Preston McAfee, Robert Wong, Ryan Harms, Alexander Chen, Shiv Venkataraman, Steve Seitz,
Steven Johnson, Tulsee Doshi.

Other Interviewees
Daniel Rock, Erik Brynjolfsson, Joseph Briggs, Michael Mandel.

21
References
Executive summary

p. 1 80% of jobs exposed to generative AI: Eloundou, Tyna, et al. “GPTs Are GPTs: An Early Look at the Labor Market Impact Potential of
Large Language Models.” arXiv, March. 2023, arxiv.org/abs/2303.10130

Part 1: Introduction

p. 2 Physicist Freeman Dyson: Dyston, Freeman. Infinite in all directions: Gifford lectures given at Aberdeen, Scotland, April-November 198
Penguin, 1989. Google Books, https://www.google.com/books/edition/Infinite_in_All_Directions/e8TaAAAAMAAJ?hl=en&gbpv=0&bsq=gift

2 Life expectancy growth: Dattani, Saloni, et al. "Life Expectancy.” Our World in Data, 2023, ourworldindata.org/life-expectancy

2 Environmental indicators (CO2): World Bank, Global Carbon Budget, and various sources -- with processing by Our World in Data.
“Change in per Capita CO₂ Emissions and GDP.” Our World in Data, 2023, ourworldindata.org/grapher/co2-emissions-and-gdp-per-cap

2 Working hours decline: Bick et al (2019); Feenstra et al. (2015), Penn World Table (2021); Huberman & Minns (2007; and Organisation fo
Economic Co-operation and Development, Labour Force Statistics – processed by Our World in Data. “Annual Working Hours per Worke
Various Sources.” Our World in Data, ourworldindata.org/grapher/compare-sources-working-hours?country=~OECD+Countries

2 Standard of living growth: World Bank Poverty and Inequality Platform (2022) – with processing by Our World in Data. “Median income
consumption per day, 1990 to 2019.” Our World in Data,
ourworldindata.org/grapher/daily-median-income?tab=chart&country=~OWID_WRL

2 Burden of disease reduction: Institute for Health Metrics and Evaluation, Global Burden of Disease (2019) – processed by Our World in
Data. "Burden of disease, 1990 to 2019." Our World in Data,
https://ourworldindata.org/grapher/dalys-rate-from-all-causes?tab=chart&country=OWID_WRL~BGD~BEL

3 Air pollution deaths: IHME, Global Burden of Disease Study (2019) -- processed by Our World in Data. “Death Rate From Air Pollution.”
Our World in Data, ourworldindata.org/grapher/death-rates-from-air-pollution

3 Access to drinking water: World Health Organization and UNICEF – processed by Our World in Data. “Share of the Population Using
Safely Managed Drinking Water.” Our World in Data,
ourworldindata.org/grapher/proportion-using-safely-managed-drinking-water?tab=chart&country=~OWID_WRL

Part 2: Generative AI is a general-purpose technology

p. 4 General purpose technology criteria: Bresnahan, Timothy F., and Manuel Trajtenberg. “General Purpose Technologies ‘Engines of
Growth’?” Journal of Econometrics, vol. 65, no. 1, Jan. 1995, pp. 83–108. https://doi.org/10.1016/0304-4076(94)01598-t

4 “The best thing about AI...": Wolfram, Stephen. What Is ChatGPT Doing ... And Why Does It Work?, Wolfram Media Inc, 2023

4 GPT-3.5 and GPT-4 U.S. bar exam performance: OpenAI. “GPT-4 Technical Report.” arXiv, March 2023, arxiv.org/abs/2303.08774

4 Massive Multitask Language Understanding (MMLU) benchmark: Hendrycks, Dan, et al. “Measuring Massive Multitask Language
Understanding.” arXiv, Jan. 2021, arxiv.org/abs/2009.03300

4 Gemini Ultra performance: Pichai, Sundar, and Demis Hassabis. “Introducing Gemini: Our Largest and Most Capable AI Model.” Googl
The Keyword Blog, Dec. 2023, blog.google/technology/ai/google-gemini-ai/#performance

5 7.5 page context window (2020): “What Is Context Window for LLMs?” Hopsworks,
www.hopsworks.ai/dictionary/context-window-for-llms#:~:text=For%20example%2C%20in%20GPT%2D3,their%20usefulness%20acros
%20various%20applications

22
5 300 page context window (2023): "New Models and Developer Products Announced at DevDay." OpenAI Blog, November 2023,
openai.com/blog/new-models-and-developer-products-announced-at-devday

5 U.S. executive survey: “KPMG Generative AI Survey.” KPMG, July 2023,

info.kpmg.us/news-perspectives/technology-innovation/kpmg-generative-ai-2023.html

5 Exposure of tasks in the American economy: Eloundou, Tyna, et al. “GPTs Are GPTs: An Early Look at the Labor Market Impact Poten
of Large Language Models.” arXiv, March. 2023, arxiv.org/abs/2303.10130

5-6 Robotic actions as a complementary innovation: “RT-2: New Model Translates Vision and Language Into Action.” Google DeepMind
Blog, July 2023, deepmind.google/discover/blog/rt-2-new-model-translates-vision-and-language-into-action

6 MBA entrepreneurship game: Mollick, Ethan. “Reshaping the Tree: Rebuilding Organizations for AI.” One Useful Thing, November 2023
www.oneusefulthing.org/p/reshaping-the-tree-rebuilding-organizations

6 New material design (Microsoft): Fowler, Andrew, et al. "MatterGen: Property-guided materials design." Microsoft Research Blog,
December 2023. https://www.microsoft.com/en-us/research/blog/mattergen-property-guided-materials-design/

6 New material design (Google DeepMind): Merchant, Amil and Ekin Dogus Cubuk. "Millions of new materials discovered with deep
learning." Google DeepMind Blog, November 2023,
https://deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/

Part 3: Faster This Time

p. 6 U.S. electric motor patent: “U.S. Patent: 132 - Improvement in Propelling Machinery by Magnetism and Electro-magnetism.” Directory
American Tool and Machinery Patents, www.datamp.org/patents/displayPatent.php?pn=132&id=22442

6 American electrification (1880s - 19020s): David, Paul A. “The Dynamo and the Computer: An Historical Perspective on the Modern
Productivity Paradox.” The American Economic Review, vol. 80, no. 2, May 1990. JSTOR, www.jstor.org/stable/2006600

7 Survey of 14,000 professionals: “More Than Half of Generative AI Adopters Use Unapproved Tools at Work.” Salesforce News & Insigh
Nov. 2023,
www.salesforce.com/news/stories/ai-at-work-research/?utm_campaign=amer_cbaw&utm_content=Salesforce_World+Tour&utm_mediu
organic_social&utm_source=linkedin

7 Programmer survey: Shani, Inbal. “Survey Reveals AI’s Impact on the Developer Experience - the GitHub Blog.” The GitHub Blog, June
2023, github.blog/2023-06-13-survey-reveals-ais-impact-on-the-developer-experience

8 Call center employees resolve more calls: Brynjolfsson, Erik, et al. "Generative AI at Work." National Bureau of Economic Research, N
2023, www.nber.org/papers/w31161. Working paper

8 Management consultants perform better: Dell’Acqua, Fabrizio, et al. “Navigating the Jagged Technological Frontier: Field Experiment
Evidence of the Effects of AI on Knowledge Worker Productivity and Quality.” Social Science Research Network, Jan. 2023,
https://doi.org/10.2139/ssrn.4573321. Harvard Business School Technology & Operations Mgt. Unit Working Paper No. 24-013

8 Programmers code faster: Peng, Sida, et al. “The Impact of AI on Developer Productivity: Evidence From GitHub Copilot.” arXiv, Feb.
2023, https://doi.org/10.48550/arxiv.2302.06590

8 Writers improve speed and quality: Noy, Shakked, and Whitney Zhang. “Experimental Evidence on the Productivity Effects of Generat
Artificial Intelligence.” Social Science Research Network, Apr. 2023, https://doi.org/10.2139/ssrn.4375283. Working paper

8 Physicians complete notes faster: Lohr, Steve. "A.I. May Someday Work Medical Miracles. For Now, It Helps Do Paperwork." The New
York Times, June 2023, https://www.nytimes.com/2023/06/26/technology/ai-health-care-documentation.html

8 U.S. public company valuations: Eisfeldt, Andrea L., et al. “Generative AI and Firm Values.” National Bureau of Economic Research, Ma
2023, https://doi.org/10.3386/w31222 . Working paper

8 Smaller companies lag in technology investments: Govindarajan, Vijay, et al. “The Gap Between Large and Small Companies is

23
 Growing. Why?” Harvard Business Review, August 2019,
https://hbr.org/2019/08/the-gap-between-large-and-small-companies-is-growing-why

8-9 Sector productivity slowdown: Sprague, Shawn. “The U.S. productivity slowdown: an economy-wide and industry-level analysis.” U.S
Bureau of Labor Statistics, April 2021,
https://www.bls.gov/opub/mlr/2021/article/the-us-productivity-slowdown-the-economy-wide-and-industry-level-analysis.htm

Part 4: Implications

p. 9 Generative AI’s GDP impacts: “Upgrading Our Longer-Run Global Growth Forecasts to Reflect the Impact of Generative AI.” Goldman
Sachs Research, Oct. 2023,
www.gspublishing.com/content/research/en/reports/2023/10/30/2d567ebf-0e7d-4769-8f01-7c62e894a779.html

9 Better health from economic growth: UN WPP (2022); HMD (2023); Zijdeman et al. (2015); Riley (2005); Maddison Project Database
2020 (Bolt and van Zanden, 2020); Gapminder - Population v7 (2022) and other sources; Our World in Data – with processing by Our
World in Data. “Life Expectancy Vs. GDP per Capita.” Our World in Data, ourworldindata.org/grapher/life-expectancy-vs-gdp-per-capit

9 Indira Gandhi on poverty and pollution: United Nations, Environment Programme. Effective, inclusive and sustainable multilateral
actions to tackle climate change, biodiversity loss, and pollution.
https://wedocs.unep.org/bitstream/handle/20.500.11822/43634/ED-Report-UNEA6-AdvanceUnEditedVersion.pdf?sequence=3&isAllowe
y

10 More life satisfaction with economic growth: World Happiness Report (2023); World Bank (2023); Gapminder - Population v7 (2022)
other sources; Our World In Data -- processed by Our World in Data. “Self-reported Life Satisfaction Vs. GDP per Capita.” Our World in
Data, ourworldindata.org/grapher/gdp-vs-happiness

11 Pollution and economic growth: IHME, Global Burden of Disease Study (2019); World Bank (2023); Gapminder - Population v7 (2022) a
other sources; Our World In Data – processed by Our World in Data. "Death rate from outdoor air pollution vs. GDP per capita." Our Wo
in Data, https://ourworldindata.org/grapher/outdoor-pollution-rate-vs-gdp

11 Hours back in physicians' days: Lohr, Steve. "A.I. May Someday Work Medical Miracles. For Now, It Helps Do Paperwork." The New Yor
Times, June 2023. https://www.nytimes.com/2023/06/26/technology/ai-health-care-documentation.html

11 Accuracy of medical diagnoses: McDuff, Daniel, et al. “Towards Accurate Differential Diagnosis With Large Language Models.” arXiv,
Nov. 2023, arxiv.org/abs/2312.00164

12 Management consultant upskilling: Dell’Acqua, Fabrizio, et al. “Navigating the Jagged Technological Frontier: Field Experimental
Evidence of the Effects of AI on Knowledge Worker Productivity and Quality.” Social Science Research Network, Jan. 2023,
https://doi.org/10.2139/ssrn.4573321. Harvard Business School Technology & Operations Mgt. Unit Working Paper No. 24-013

12 Writer upskilling: Noy, Shakked, and Whitney Zhang. “Experimental Evidence on the Productivity Effects of Generative Artificial
Intelligence.” Social Science Research Network, Apr. 2023, https://doi.org/10.2139/ssrn.4375283. Working paper

12 Call center employee upskilling: Brynjolfsson, Erik, et al. "Generative AI at Work." National Bureau of Economic Research, Nov. 2023,
www.nber.org/papers/w31161. Working paper

12 Law student upskilling: Choi, Jonathan H., and Daniel Schwarcz. “AI Assistance in Legal Analysis: An Empirical Study.” Social Science
Research Network, Jan. 2023, https://doi.org/10.2139/ssrn.4539836

12 Skill biased technological change: Autor, David, et al. “Extending the Race Between Education and Technology.” AEA Papers and
Proceedings, May 2020, scholar.harvard.edu/lkatz/publications/extending-race-between-education-and-technology

12 Automated unloading of container ships: Kay, Grace. “Meet the Port of Los Angeles’ Dock Workers Who Make Over $100,000 Workin
With Some of the World’s Largest Robots.” Business Insider, Sept. 2021,
www.businessinsider.com/port-la-dock-workers-longshoremen-make-100k-salary-overtime-2021-9

12 Interacting with generative AI enables top performance: Mollick, Ethan. “Centaurs and Cyborgs on the Jagged Frontier.” One Usefu
Thing, Sept. 2023, www.oneusefulthing.org/p/centaurs-and-cyborgs-on-the-jagged

24
12 Manufacturing employment trends: “U.S. Bureau of Labor Statistics, All Employees, Manufacturing.” FRED, Federal Reserve Bank of S
Louis, Dec. 2023, kfred.stlouisfed.org/series/MANEMP

12 Agriculture employment trends: “Organization for Economic Co-operation and Development, Employment by Economic Activity:
Agriculture: All Persons for United States.” FRED, Federal Reserve Bank of St. Louis, Nov. 2023,
kfred.stlouisfed.org/series/LFEAAGTTUSM647S

13 85% of U.S. employment growth since 1940: “Upgrading Our Longer-Run Global Growth Forecasts to Reflect the Impact of Generativ
AI.” Goldman Sachs Research, Oct. 2023,
www.gspublishing.com/content/research/en/reports/2023/10/30/2d567ebf-0e7d-4769-8f01-7c62e894a779.html.

13 Tight labor markets in OECD countries: “OECD Employment Outlook 2023: AI and Jobs, an Urgent Need to Act.” OECD, 2023,
oecd.org/employment-outlook/2023

13 Tight labor markets in the U.S.: Jefferson, Nathan and Jack Fuller. "A State-Level Look at U.S. Labor Market Supply and Demand." FRE
Federal Reserve Bank of St. Louis, April 2023,
https://www.stlouisfed.org/publications/regional-economist/2023/apr/state-level-us-labor-market-supply-demand

13 American labor shortage: Ferguson, Stephanie. “America Works Data Center.” U.S. Chamber of Commerce, Dec. 2023,
www.uschamber.com/workforce/america-works-data-center

13 European occupational shortages: "EURES Report on labor shortages and surpluses 2022." European Labor Authority, March 2023,
https://www.ela.europa.eu/sites/default/files/2023-09/ELA-eures-shortages-surpluses-report-2022.pdf

13 Skill gaps and an inability to attract talent: "The Future of Jobs Report 2023." World Economic Forum, April 2023,
https://www.weforum.org/publications/the-future-of-jobs-report-2023/

14 OECD unemployment rates: "Unemployment Rates, OECD - Updated: Dec. 2023." OECD, Better Policies for Better Lives, Dec. 2023,
https://www.oecd.org/employment/unemployment-rates-oecd-updated-december-2023.htm

14 Future U.S. workforce participation: "Civilian labor force participation rate by age, sex, race, and ethnicity." U.S. Bureau of Labor
Statistics, September 2023, https://www.bls.gov/emp/tables/civilian-labor-force-participation-rate.htm

14 Workforce exits of Americans aged 65 and older: "Economic News Release, Employment Projections 2022-2032 Summary." U.S. Bure
of Labor Statistics, September 2023, https://www.bls.gov/news.release/ecopro.nr0.htm

14 Workforce aging: "World Population Prospects 2022." United Nations Department of Economic and Social Affairs, 2022,
https://www.un.org/development/desa/pd/sites/www.un.org.development.desa.pd/files/wpp2022_summary_of_results.pdf

14 Hollowing out of labor force: Canon, Maria and Elise Marifan. "Job Polarization Leaves Middle-Skilled Workers Out in the Cold." Fred,
Federal Reserve Bank of St. Louis, January 2013,
https://www.stlouisfed.org/publications/regional-economist/january-2013/job-polarization-leaves-middleskilled-workers-out-in-the-co

14-15 College graduate tasks more exposed to generative AI: Eloundou, Tyna, et al. “GPTs Are GPTs: An Early Look at the Labor Market
Impact Potential of Large Language Models.” arXiv, March. 2023, arxiv.org/abs/2303.10130

14-15 Freelance writing pay and jobs: Hui, Xiang, et al. “The Short-Term Effects of Generative Artificial Intelligence on Employment: Evidenc
From an Online Labor Market.” Social Science Research Network, Aug. 2023, https://doi.org/10.2139/ssrn.4527336

15 Wage trends: Autor, David H., et al. “The Unexpected Compression: Competition at Work in the Low Wage Labor Market.” National Bure
of Economic Research, Nov. 2023, https://doi.org/10.3386/w31010. Working paper

15 40% of U.S. workforce in agriculture: Klein, Ezra, and Susannah Locke. “40 Maps That Explain Food in America.” Vox, June 2014,
www.vox.com/a/explain-food-america

15 40% of U.S. population living on farms: Lusk, Jason. “The Evolving Role of the USDA in the Food and Agricultural Economy.” Mercatu
Center, June 2016, www.mercatus.org/research/research-papers/evolving-role-usda-food-and-agricultural-economy.

25
15 Companies disrupted by general purpose technologies: Caves, Richard. “The Decline of Dominant Firms, 1905-1929.” Quarterly
Journal of Economics, vol. 99, no. 3. JSTOR, www.jstor.org/stable/1885963

15 U.S. executive survey: “KPMG Generative AI Survey.” KPMG, July 2023,

info.kpmg.us/news-perspectives/technology-innovation/kpmg-generative-ai-2023.html

16 Superstar firms in the 21st century: Autor, David H., et al. “The Fall of the Labor Share and the Rise of Superstar Firms.” The Quarterly
Journal of Economics, vol. 135, no. 2, May 2020, pp. 645–709. https://doi.org/10.1093/qje/qjaa004

16 Young U.S. superstar firms: McAfee, Andrew. The Geek Way: The Radical Mindset That Drives Extraordinary Results. Little, Brown and
Company, November 2023

16 Companies conducting layoffs: Babina, Tania, et al. “Artificial Intelligence, Firm Growth, and Product Innovation.” Journal of Financial
Economics, vol. 151, Jan. 2024, p. 103745. https://doi.org/10.1016/j.jfineco.2023.103745

16 Global cost of cybercrime: Muggah, Robert, and Mac Margolis. “Why We Need Global Rules to Crack Down on Cybercrime.” World
Economic Forum, Jan. 2023, www.weforum.org/agenda/2023/01/global-rules-crack-down-cybercrime

16 Google DeepMind taxonomy of generative AI risks: Weidinger, Laura, et al. "Ethical and social risks of harm from Language Models."
arXiv, Dec. 2021, https://doi.org/10.48550/arXiv.2112.04359

16 Electricity's novel risks: Sullivan, J. P. “Fearing Electricity: Overhead Wire Panic in New York City.” IEEE Journals & Magazine | IEEE
Xplore, vol. 14, no. 3, 1995, ieeexplore.ieee.org/document/464629

16 Google DeepMind framework for assessing generative AI risks in context: Weidinger, Laura, et al. "Sociotechnical Safety Evaluatio
of Generative AI Systems." arXiv, Oct. 2023, https://doi.org/10.48550/arXiv.2310.11986

Part 5: Open Questions

p. 17 2016 prediction on future of radiology: Windsor, Matt. “This Radiologist Is Helping Doctors See Through the Hype to an AI Future - the
Reporter.” UAB Reporter, Dec. 2022,
www.uab.edu/reporter/people/achievements/item/9925-this-radiologist-is-helping-doctors-see-through-the-hype-to-an-ai-future

17 200 AI-based radiology imaging products: “Your Resource for Strategic AI Solutions.” American College of Radiology, Data Science
Institute, aicentral.acrdsi.org

17 U.S. radiology employment 2016 - 2021: “Percentage Change in the Number of Active Physicians by Specialty, 2016-2021 | AAMC.”
Association of American Medical Colleges,
www.aamc.org/data-reports/workforce/data/percentage-change-number-active-physicians-specialty-2016-2021

17 Imaging demand: Henderson, Mary. “Radiology Facing a Global Shortage.” Radiology Society of North America, May 2022,
www.rsna.org/news/2022/may/global-radiologist-shortage.

18 Planning, reasoning, and memory performance: Mialon, Grégoire, et al. “GAIA: A Benchmark for General AI Assistants.” arXiv, Nov.
2023, arxiv.org/abs/2311.12983

26
Appendix: Further Reading

● Acemoglu, Daron and Simon Johnson. Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity. PublicAffairs, 2023.

● Agrawal, Ajay, et al. Power and Prediction: The Disruptive Economics of Artificial Intelligence. Harvard Business Review Press, 2022.

● Artigas, Carme, et al. “What Global AI Governance Must Do.” Project Syndicate, Dec. 2023,
https://www.project-syndicate.org/commentary/ai-governance-un-advisory-body-five-principles-by-ian-bremmer-et-al-2023-12?barrier=acc
esspaylog

● Brynjolfsson, et al. “The Productivity J-Curve: How Intangibles Complement general-purpose technologies.” National Bureau of Economic
Research, January 2020, https://www.nber.org/papers/w25148. Working paper.

● Brynjolfsson, Erik. “The Turing Trap: The Promise & Peril of Human-LIke Artificial Intelligence.” Dædalus, Spring 2022,
https://www.amacad.org/publication/turing-trap-promise-peril-human-artificial-intelligence

● Manyika, James and Michael Spence. “The Coming AI Economic Revolution.” Foreign Affairs, November / December 2023.
https://www.foreignaffairs.com/world/coming-ai-economic-revolution

● McAfee, Andy and Erik Brynjolfsson. The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies. W.W. Norton
& Company, January 2016.

● Mckinsey & Company. The Economic Potential of Generative AI: the Next Productivity Frontier. June 2023,
https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier
#introduction.

● Walker, Kent. “An opportunity agenda for AI.” Google: The Keyword Blog, Nov. 2023.
https://blog.google/outreach-initiatives/public-policy/google-ai-opportunity-agenda/

27