Yu 2023 AI Psychology Oxford
Yu 2023 AI Psychology Oxford
Yu 2023 AI Psychology Oxford
Introduction
In 2019, the United States established a national task force for coordinating AI strategies across the federal government, industry, and
academia in order to promote scientific discovery, economic competitiveness, and national security. In addition, the 2023 Organisation for
Economic Co-operation and Development’s report indicates that AI has become a global competition. Hence, AI is here to stay and this
trend is too important to ignore or downplay. From the incubation and development of AI, the relationship between AI and psychology is
symbiotic. As cognitive psychologists and neuroscientists gain more insight into how the brain works, AI has been developed by mimicking
human neural pathways. On the other hand, insights derived from AI research can be applied to a variety of subfields in psychology,
whereas new social issues emerged from AI applications, such as AI bias, echo chambers, and misuse of AI tools (e.g., ChatGPT and
Midjourney), led to new research topics in psychology. Further, several schools of thought of AI, such as the connectionist, symbolic, and
analogist approaches, heavily borrowed ideas from psychological research. For example, in an attempt to build structural intelligence,
connectionists draw an analogy between human neural networks and artificial neural networks. Moreover, AI symbolists subscribe to the
notion that the human mental process is a logical production system. Furthermore, prior psychological research indicates that analogical
thinking is commonly employed for problem-solving, and this notion became the foundation of example-based machine learning. Similarly,
reinforcement learning in AI was inspired by behavioral psychology. Additionally, in line with the findings of developmental psychology that a
child learns best through spontaneous discovery, AI researchers believe that setting the deep learning system free will be the most
promising research direction. A vast number of pioneers of AI research who devoted efforts to the preceding research agendas are
psychologists or received training in psychology, such as Frank Rosenblatt, Allen Newell, John Anderson, David Rumelhart, and Geoffrey
Hinton, whereas some were inspired by cognitive science or neuroscience, such as Fei-Fei Li, Demis Hassabis, and Yann LeCun. In
addition, in their work Artificial Psychology: Psychological Modeling and Testing of AI Systems (cited under Artificial Psychology) Crowder
and his colleagues examined the bidirectional relationship between artificial intelligence and psychology, focusing on cognitive architectures
and artificial emotions. The authors proposed comprehensive mathematical models for cognitive architectures, envisioning an AI system
capable of reasoning about emotions, adapting to humans, and constructing knowledge representations based on experiences.
General Overviews
There are many books that illustrate the history and concepts of artificial intelligence and machine learning (AI/ML). Russell and Norvig
2021 is considered one of the most comprehensive introductions to AI/ML. Another book that explains AI through a historical perspective is
Metz 2021, which is written using a storytelling and accessible approach. The book shows how psychologists and scholars in other
disciplines jointly contributed to the development of AI. For a more concise introduction to the history of AI, Barbet 2020 is recommended.
Besides introducing the internal development of AI schools, Barbet also discussed external factors contributing to the movement. Domingos
2015 compares and contrasts the five schools of AI; namely, symbolism, connectionism, Bayesianism, analogists, and evolutionists. These
five schools of thought can be further explored by checking out the sources listed in the book. Today the terms “data science” (DS) and
“machine learning” (ML) are often used in tandem. Although originally DS and ML were two separate movements, in the early twenty-first
century they merged together as an integrated entity because both methodologies emphasized how to extract patterns and trends from the
data by exploration. To understand the conceptual relationships between DS and ML, consult Yu 2022. Most DSML books are tied to
specific platforms or languages (e.g., Python, R, SAS®, etc.), but it is important not to equate DSML with any particular software tool, or to
confuse computing programming with DSML-based analytics. As an example, LISP was one of the first programming languages that has
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 1/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
traditionally been associated with AI computing. LISP, however, lost popularity over time to other programming languages. It is advised that
readers learn DSML conceptually and keep an open mind when dealing with various AI tools. While in DSML, prediction, recommendation,
and optimizing search for patterns and associations from big data is more central than explanation. Pearl and Mackenzie 2018 drew the
attention of researchers to causal inferences by introducing the Bayesian probabilistic inference network. A number of books about the
relationship between AI and psychology have been published. Schneider 2019 is both psychological and philosophical. Specifically, the
book discusses the meanings of the self, the mind, and consciousness in the context of AI. Brian Christian works closely with cognitive
scientists to study technological impacts on human lives. Christian 2020 discusses the challenge that the goal of AI machines might not fully
align with human values.
Barbet, Jacques. 2020. The maturation of artificial intelligence. In Therapeutic progress in oncology: Towards a revolution in
cancer therapy? Edited by Jacques Barbet, Adrien Foucquier, and Yves Thomas, 79–109, New York: Wiley.
Besides covering different approaches to AI in terms of cognitive models and mathematics, the author also discusses other factors that
facilitate AI development, such as the support from the government and tech giants, the availability of big data for data mining, and the
invention of power computing technologies (e.g., GPU).
Christian, Brian. 2020. The alignment problem: Machine learning and human values. New York: W. W. Norton.
This book discusses the ethical and psychological challenges that occur when the goals of AI systems and human values are misaligned. A
critical analysis of how embedded biases impact decision-making and social impact in AI models is presented in the book. The book uses a
storytelling approach and lay terms.
Domingos, Pedro. 2015. The master algorithm: How the quest for the ultimate learning machine will remake our world. New York:
Basic Books.
This book explores five different schools of AI, including the symbolist, connectionist, Bayesian, evolutionary, and analogical approaches.
Domingos suggests integrating different tribes together in order to develop a more holistic approach. This book is highly accessible.
Metz, Cade. 2021. Genius makers: The mavericks who brought AI to Google, Facebook, and the world. New York: Dutton.
The book chronicles the achievements, struggles, and visions of several AI mavericks, including Geoffrey Hinton, Yann LeCun, and Yoshua
Bengio. It also includes the voice of opposition to deep learning from Gary Marcus. The author uses a narrative approach, and thus the
book is highly readable.
Pearl, Judea, and Dana Mackenzie. 2018. The book of why: The new science of cause and effect. New York: Basic Book.
The first author, Pearl, is a Turing Award Laureate. This book explores the history of causal inference, a fundamental concept of research
that is seldom discussed among machine learning researchers. One of the core messages of the book is that for causal inferences
researchers need the right data, not necessarily big data.
Russell, Stuart, and Peter Norvig. 2021. Artificial intelligence: A modern approach. 4th ed. Harlow, UK: Pearson.
This book presents a comprehensive overview of AI, including its history, concepts, techniques, and applications. It covers the historical
development of AI from the symbolic approach to more recent statistical and connectionist methods.
Schneider, Susan. 2019. Artificial you: AI and the future of your mind. Princeton, NJ: Princeton Univ. Press.
The purpose of this book is to examine the potential impact of AI on our understanding of consciousness, identity, and personhood from a
psychological standpoint. There are thought-provoking questions raised about the relationship between human psychology and AI, but
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 2/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
some issues are hypothetical and might not be strongly relevant to AI users.
Yu, Chong Ho. 2022. Data mining and exploration: From traditional statistics to modern data science. Boca Raton, FL: CRC Press.
This book is a user-friendly introduction to both conceptual and procedural aspects of cutting-edge data science and machine learning
methods, such as neural networks, bagging, and boosting. Unlike other books that are tied to a particular software application, this book
covers a variety of tools, including SAS®, JMP®, SPSS®, and Python.
In this section, fundamental concepts of the field, such as AI, machine learning, narrow AI, strong AI, and AGI, are elucidated. A wide variety
of applications, such as chatbots, generative art, and facial recognition, have proven the success of machine learning. Although machines
are capable of learning from data and improving their performance on specific tasks, they don’t have the same type of learning as humans.
Human learning involves complex cognitive processes, understanding, reasoning, and consciousness, whereas machines learn by
mathematical optimizations and pattern recognition. One of the major concerns about AI is whether such a system can truly understand and
comprehend the information it processes, or if it simply behaves like an intelligent being. The notion of AI attaining consciousness is another
contentious and philosophical debate. Many experts, including Hubert Dreyfus, John Searle, and Gary Marcus, have argued that current AI
technologies are far from achieving genuine consciousness, leaving uncertainty about the possibility of machines ever possessing such a
quality.
Artificial intelligence, which is intended to simulate human intelligence through computers, is a multidisciplinary field that necessitates the
integration between computer science, engineering, mathematics, psychology, and many other disciplines. The term “artificial intelligence”
was coined by John McCarthy in 1955. Currently, many AI programs specialize in specific tasks, and thus these are considered “narrow AI.”
For example, even though a Go-playing AI program can play the game so well that it can defeat the world champion, this skill cannot be
transferred to other domains, such as facial recognition. If an AI system is capable of performing a variety of intellectual tasks, not just one
or two specific ones, then this type of AI is known as “artificial general intelligence” (AGI). If the AI system can perform any intellectual task
that a human can do, or can reason like a human mind, then it is called “strong AI.” In most cases AGI and strong AI are treated as
synonymous. However, different types of AI can be defined differently, depending on their fields. According to computer scientists,
intelligence refers to our ability to achieve goals, and thus AGI is defined in terms of the variety of tasks accomplished. In contrast,
psychologists have defined AGI as a function of adaptability for different contexts (see, e.g., Kai and Fu 2021), whereas Cichocki and
Kuleshov 2021 argues that strong AI consists of multiple intelligences, as proposed by Gardner 2000; namely, logical-mathematical, visual-
spatial, musical-rhythmic, verbal-linguistic, bodily-kinesthetic, naturalistic, intrapersonal, and interpersonal intelligences. Philosopher John
Searle brought consciousness into the equation (see Searle 1980). An AI system is said to be “weak AI” if it is just a simulated model of the
mind without self-awareness. On the other hand, a strong AI system is a conscious mind that is indistinguishable from humans, such as
HAL 9000 in the sci-fi-movie 2001: A Space Odyssey, or Commander Data in the sci-fi TV show Star Trek: The Next Generation. Following
this line of reasoning, it is possible to have AGI without strong AI. Specifically, even if an AGI system knows virtually anything and can
accomplish nearly any task, it would not be qualified as strong AI due to lacking consciousness. For a discussion of controversial topics
regarding AI, consult Ford 2018. For a brief introduction to the role of consciousness in AI, consult Hildt 2019.
Cichocki, Andrzej, and Alexander Kuleshov. 2021. Future trends for human-AI collaboration: A comprehensive taxonomy of AI/AGI
using multiple intelligences and learning styles. Computational Intelligence and Neuroscience 2021: Article 8893795.
This article discusses how different theories of intelligence in psychology, such as multiple intelligences, emotional intelligence, and social
intelligences, can inform AI researchers to develop strong AI. The paper contains many graphs, making it easy to follow the illustrations.
Ford, Martin. 2018. Architects of intelligence: The truth about AI from the people building it. Birmingham, UK: Packt.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 3/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
This book is a collection of interviews with top experts on AI, including Geoffrey Hinton, Stuart Russell, Rodney Brooks, Demis Hassabis,
and Yoshua Bengio. It covers a discussion on the possibility of AGI, its impact on the labor market, and the possibility of AI’s threat against
humanity.
Gardner, Howard. 2000. Intelligence reframed: Multiple intelligences for the 21st century. London: Hachette.
In this book, Gardner challenged the conventional view of intelligence as a single, unitary trait measured by standardized IQ tests. Gardner
proposes that there are multiple types of intelligence, each representing a different form of cognitive ability. His theory has important
implications for AGI.
Hildt, Elisabeth. 2019. Artificial intelligence: Does consciousness matter?. Frontiers in Psychology 10.
Hildt introduces several contentious issues related to consciousness, such as the “hard problem of consciousness” and
“anthropomorphizing.” She argues that the concept of personhood can be transferred to AI robots as “robothood.” The discussion is both
psychological and philosophical.
Kai, Zhizhong, and Fu Xiaolan. 2021. Psychological model of representation, generation, and adjustment of brief for artificial
general intelligence. Human Behaviors and Emerging Technology 3:865–875.
The goal of this article is to develop a psychological model that will provide a theoretical framework for the implementation of operational
representations of intelligence and belief in AI. Additionally, the authors compared and contrasted psychological and computing approaches
to intelligence.
McCarthy, John. 1955. A proposal for the Dartmouth Summer Research Project on Artificial Intelligence. Hanover, NH: Association
for the Advancement of Artificial Intelligence.
In 1955, McCarthy coined the term “artificial intelligence” in a conference proposal for the Dartmouth Conference, which took place during
the summer of 1956 at Dartmouth College in New Hampshire, USA. In the proposal, McCarthy suggested every aspect of learning and any
other feature of intelligence should be precisely described so that a machine can simulate it.
Searle, John. 1980. Minds, brains, and programs. Behavioral and Brain Sciences 3:417–457
Searle presented his famous “Chinese room” thought experiment in this article, which challenges the notion that computers are capable of
understanding language or possessing human-like mental states. The article raised important questions about the nature of human
cognition.
Although Alan Turing, the inventor of the concept of “Turing machine,” did not explicitly use the term “machine learning” (ML), he envisioned
that humans need a machine that can learn from experience. The formal term “machine learning” was coined by Arthur Samuel in 1959.
However, there was a period of “AI winters” (1970s–1990s), during which manuscripts mentioning AI or neural networks were unwelcome in
journals. In order to evade rejection, many AI researchers used the term “machine learning.” However, these two terms should not be used
interchangeably, because there are subtle differences between them. AI is a broader technology that enables a computer to mimic human
intelligence in various tasks, such as image generation and enhancement, computer vision, speech recognition, decision support, and
natural language processing. On the other hand, ML is a subset of AI that equips a computer to progressively improve its performance in a
specific task based on past examples in spite of lacking explicit programming. Generally speaking, there are three types of machine
learning: supervised learning that trains the algorithm by giving specific labeled targets, unsupervised learning algorithms that aim to find
the hidden structure or process without labeled targets, and reinforcement learning that enables the machine to make smart decisions by
interacting with the environment. Typical examples of supervised learning algorithms are the support vector machine, decision tree, random
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 4/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
forest, and gradient boosting. Examples of unsupervised machine learning methods include principal component analysis, cluster analysis,
association rules, and social network analysis. In machine learning, big data are indispensable because every AI system depends on past
examples for improvement. The more data the system contains, the better the result. For example, chatbots, such as ChatGPT and Claude,
evaluate billions of web pages in order to provide valid answers to users. The same can be said for AI-enabled art generators like
Midjourney and Stable Diffusion, which aggregate billions of images from the Internet to create art. For more information about machine
learning, consult Alpaydin 2020, Bishop 2018, and Murphy 2012. The neural network, which is composed of the input, hidden, and output
layers, is a typical approach to machine learning. The hidden layer is responsible for data transformation using a set of weights and biases.
When a neural network entails multiple hidden layers, this type of machine learning is called “deep learning.” For the details of deep
learning, consult Goodfellow, et al. 2016.
Alpaydin, Ethem. 2020. Introduction to machine learning. Cambridge, MA: MIT Press.
The book is a user-friendly introduction to both basic and advanced machine learning topics, including supervised learning, Bayesian
decision theory, parametric methods, semiparametric methods, nonparametric methods, multivariate analysis, hidden Markov models,
reinforcement learning, kernel machines, graphical models, and Bayesian estimation.
Bishop, Christopher. 2018. Pattern recognition and machine learning. New York: Springer.
Unlike other books on ML, this one takes a Bayesian approach as its cornerstone. Readers who are trained in the frequency school might
find it refreshing. The book is aimed at doctoral students and researchers, not beginners.
Goodfellow, Ian, Bengio Yoshua, and Courville Aaron. 2016. Deep learning. Cambridge, MA: MIT Press.
The book introduces both the conceptual and computational aspects of deep learning, including regularization, optimization, and
autoencoders. The book is highly mathematical and computational.
Murphy, Kevin. 2012. Machine learning: A probabilistic perspective. Cambridge, MA: MIT Press.
The book is a comprehensive introduction to machine learning from a probabilistic perspective, covering topics such as supervised and
unsupervised learning, graphical models, reinforcement learning, and more. Written in an informal, approachable style, this book is packed
with pseudo codes for learning algorithms.
Dreyfus 1972, Dreyfus 1979, and Searle 1980 were all critical of strong AI. Dreyfus asserted that human intelligence is primarily driven by
unconscious thoughts and disentanglement of contextual ambiguities, rather than explicit symbolic manipulation. There are many things
that can easily be done by humans but computers fail to do. In a similar vein, Searle used the thought experiment of a “Chinese room” to
argue against symbolism for strong AI. Searle imagined himself in a room processing inquiries written in Chinese slipped under the door.
Even though he does not understand any Chinese, through syntactic processing in the computer, he can still produce seemingly sensible
answers in Chinese, but indeed no deep comprehension is involved in the process. Cognitive psychologist Gary Marcus, a leading critic of
deep learning for strong AI and AGI, echoes partial views of Dreyfus and Searle. While both Dreyfus and Searle dismiss symbolism, Marcus
is not opposed to it. Similar to Dreyfus, Marcus and Davis 2020 argued that AI works well in a clean and closed system only, but not a
messy and open system. For instance, a computer vision system once misidentified a parking sign with stickers on it as a refrigerator filled
with lots of food and drinks, a task that is supposed to be simple to humans. In alignment to Searle’s criticism, Marcus pointed out that
humans can be fooled by computers, because we equate apparently intelligent behaviors as genuine intelligence. Marcus explained this
phenomenon using the concept of the fundamental over-attribution error in social psychology. Fundamental over-attribution error is a
cognitive bias by which people tend to overemphasize internal factors to explain someone else’s behavior. In this case the factor is the
“intelligence” of the computer. One of the well-known examples is a chatbot called Eliza, which was misinterpreted by users as
understanding what they said when it simply output compassionate words. Further, inspired by Pinker 2003 and Spelke 2000, by two
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 5/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
nativist psychologists who believe that human knowledge is hardwired into the brain, Marcus 2003, Marcus and LeCun 2016, and Marcus
and Davis 2020 argue that empirical-based deep learning is not capable of achieving AGI or strong AI because current approaches lack
certain salient features, such as compositional structure, common-sense reasoning, knowledge representation, and innate machinery,
which are essential for human-like reasoning. As a remedy, Marcus and colleagues advocate a more interdisciplinary approach to AI
research, which draws on insights from cognitive science, linguistics, and philosophy.
Dreyfus, Hubert. 1972. What computers can’t do: A critique of artificial reason. Cambridge, MA: MIT Press.
Dreyfus argued that the symbolic approach to AI, which relies on rule-based systems, is limited in its ability to mimic human intelligence.
This is because tacit knowledge, intuition, and embodied understanding, which are involved in human intelligence, cannot be adequately
captured by formal rules or algorithms. The book is thought-provoking and controversial.
Dreyfus, Hubert. 1979. What computers still can’t do. Cambridge, MA: MIT Press.
In this sequel, Dreyfus offered a list of what computers cannot do and what humans can do. For example, humans can distinguish between
relevant and irrelevant factors, ignore irrelevant ambiguous words, and recognize categories of things without lists of features. But AI cannot
do any of these things.
Marcus, Gary. 2003. The algebraic mind: Integrating connectionism and cognitive science. Denver, CO: Bradford Books.
In this book Marcus attempted to synthesize symbolism and connectionism. The book introduces the concept of the “algebraic mind,” in
which the human mind utilizes a combination of connectionist and symbolic processing, akin to algebraic operations, to perform complex
cognitive tasks. This book is conceptual, not computational or mathematical.
Marcus, Gary, and Ernest Davis. 2020. Rebooting AI: Building artificial intelligence we can trust. New York: Vintage.
The book is so named because the authors asserted that current AI research is wrong-headed. By proposing an integration of deep learning
and symbolic reasoning, this book makes the case for a new approach to AI that helps systems understand context, deal with uncertainty,
and make ethical decisions.
Marcus, Gary, and Yann LeCun. 2016. Debate: “Does AI need more innate machinery?” YouTube video. New York: NYU Center for
Mind, Brain and Consciousness.
A key question of this debate was whether AI systems should rely primarily on data-driven machine learning algorithms or more on innate
machinery. This debate is intended for the general public.
Pinker, Steven. 2003. The blank slate: The modern denial of human nature. New York: Penguin Books.
The book argued that the human mind is not a blank slate and claimed that human nature is innate. It examined controversial topics like
intelligence, gender differences, violence, and morality, and made a compelling case for evolution and genetics in shaping human behavior.
Searle, John. 1980. Minds, brains and programs. Behavioral and Brain Sciences 3:417–457.
In the paper Searle argued against strong AI, which advocates that AI can “understand” natural language. His “Chinese room” argument is
often cited to challenge symbolism or any AI system that is claimed to be intelligent.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 6/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
Spelke suggested that the emergence of a set of building block systems early in human ontogeny and phylogeny may underlie the
development of complex cognitive skills like reading and mathematics.
Psychology has significantly influenced the development of artificial intelligence, leading to the exploration of various overarching
approaches and specific models. One influential contribution was Rosenblatt’s perceptron algorithm, which was inspired by the functioning
of neurons in the brain. The perceptron laid the groundwork for neural networks, which are computational models based on the
interconnected structure of neurons in the brain. However, objections were raised regarding the perceptron’s limitations in solving complex
problems, leading to initial skepticism about neural networks in early AI research. When neural networks fell out of favor, the symbolic
approach to AI emerged, drawing inspiration from cognitive psychology. Symbolic AI represents knowledge using symbols and manipulates
them through rules and logic. However, symbolic AI faced challenges in handling uncertainty and complexity in real-world tasks. With
advancements in computational power and data availability, the connectionist approach regained prominence. Connectionist models, such
as convolutional neural networks (CNNs), revolutionized computer vision and pattern recognition. CNNs mimic the visual processing
capabilities of the human brain and have achieved remarkable success on tasks like image classification and object detection. The
analogist approach to AI is influenced by analogy-making processes in psychology, involving the mapping of knowledge and problem-
solving techniques from one domain to another. Drawing analogies helps AI systems reason and make inferences, bridging gaps between
different domains. Furthermore, inspired by operant conditioning principles from behaviorist psychology, AI models based on reinforcement
learning learn through dynamic interactions within an environment, receiving feedback in the form of rewards and punishments. This
approach has been successfully applied in areas like game-playing agents and autonomous systems.
Rosenblatt’s Perceptron
One of the prototypical AI systems, perceptron, was developed by a Cornell psychologist named Frank Rosenblatt. Rosenblatt’s project did
not set out to create an intelligent machine; rather, he wanted to better understand human mental processes by studying how a machine
could learn. The notion of Rosenblatt’s perceptron is based on the work of American neurophysiologist Warren McCulloch and his colleague
Walter Pitts (see McCulloch and Pitts 1943), as well as the learning theory proposed by Canadian psychologist Donald Hebb (see Hebb
1949). According to McCulloch and Pitts, the brain is a system of interconnected neurons that can perform logical operations. Being inspired
by Pavlovian conditioning, Hebb proposed the idea of a “cell assembly,” a group of neurons that become connected through repeated
activation and form the basis for learning and memory. Specifically, if a neuron connects to another through a synapse, and its firing is
continuously followed by another firing, then that synapse should strengthen over time. Rosenblatt was also motivated by the advent of
electronic computers and the development of new mathematical techniques, such as linear algebra, for processing data. In 1958,
Rosenblatt modified and implemented Hebb’s law in a computing system named perceptron, which is considered a prototypical neural
network. While a typical modern AI system consists of numerous artificial neurons and layers, Rosenblatt’s perceptron is just a basic one-
neuron and one-layer classifier for one simple function: separating two classes by a linear discriminant function. Specifically, the activation
function in this system produces a binary output (1 or 0), depending on whether the weighted sum of the inputs exceeds a certain threshold.
The machine learning rule is a simple algorithm that updates the weights based on the residual between the predicted and the actual
output, multiplied by the input values. However, the perceptron is not pre-programmed in the sense of regular computing; rather, weight
modification in the perceptron depends on performance evaluation by an external agent. In today’s terminology it is a form of supervised
learning. Utilizing an IBM 704 mainframe computer, the state-of-the-art computing machine in the 1950s, Rosenblatt used the preceding
approach to distinguish the card marked with a square on the right and the other marked on the left. After fifty iterations, the electronic
system learned how to “perceive” the information and make the correct decision. The technical detail of this precursor of this modern AI
system is summarized in Rosenblatt 1962.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 7/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
McCulloch, Warren, and Walter Pitts. 1943. A logical calculus of ideas immanent in nervous activity. Bulletin of Mathematical
Biophysics 5:115–133.
This is a seminal paper in the field of AI. In the paper, the brain is described as an interconnected network of simple neurons connected by
synapses. Based on a simple threshold function, neurons receive input from other neurons and produce an output signal.
Rosenblatt, Frank. 1962. Principles of neurodynamics: Perceptrons and the theory of brain mechanisms. Kansas City, MO:
Spartan Books.
This book introduced the perceptron algorithm that can perform pattern recognition. It also discusses its potential applications, such as
image recognition and natural language processing. Furthermore, Rosenblatt explains how neural networks can better understand how the
brain works, which was his research objective.
The idea of perceptron was criticized by two other pioneers of AI, Marvin Minsky and Seymour Papert. According to Minsky and Papert
1969, the perceptron method is able to converge the cases into the learning outcome in a finite number of steps if and only if the two
classes are linearly separable. If the two classes are not linearly separable, such as the XOR logical function, the algorithm will be trapped
in an infinite loop. Second, the perceptron algorithm can only learn with a single layer of weights. It cannot incorporate multiple layers, which
are needed for solving more complex problems. Last but not least, the perceptron algorithm is sensitive to noisy data and outliers, and
therefore the model may overfit the training data and fails to generalize to new data. Nonetheless, these limitations are not insurmountable.
Olazaran 1993 and Olazaran 1996 argued that the rejection of perceptron was not merely technologically grounded; rather, it involved
psychological and social factors, such as personality conflict and competition for funding resources. As a result of Minsky and Papert’s
harsh criticisms, neural networks lost momentum. Although neural nets were not completely abandoned, the followers were displaced from
the AI community to other disciplines. The negative perception of neural networks had remained the same for several decades in both the
AI and cognitive science communities. For example, cognitive scientists and philosophers Jerry Fodor and Zenon Pylyshyn argued in Fodor
and Pylyshyn 1988 that neural networks are unable to explain even basic features of human cognition, such as compositionality, which
refers to the notion that the meaning of a complex linguistic unit is a function of the meaning of its parts. In their view, neural networks are
based on statistical association without logical inferences, and thus they are not useful for modeling cognitive processes. Consequentially,
other cognitive models and AI methods, such as the symbolic approach, and other machine learning algorithms, such as decision trees and
support vector machines, gained more traction. Only after practical applications of backpropagation came about in the late 1980s and the
early 1990s did neural networks become a major direction of machine learning research. Today the multilayer perceptron (MLP) in artificial
neural networks has overcome the limitations of Rosenblatt’s perceptron and early neural networks. For the detail of how the distinction
between input, hidden, and output layers of a typical artificial neural network compensates for the shortcomings of Rosenblatt’s perceptron,
consult Rumelhart and McClelland 1986.
Fodor, Jerry, and Zenon Pylyshyn. 1988. Connectionism and cognitive architecture: A critical analysis. Cognition 28:3–71.
The authors argue that neural networks cannot explain certain aspects of cognition. For them, neural networks are at most a theory for
implementing symbol-processing algorithms. Thus, they urged the neural network school to re-describe the operation of the networks in
terms of symbols. The paper is highly technical.
Minsky, Marvin, and Seymour Papert. 1969. Perceptrons: An introduction to computational geometry. Cambridge, MA: MIT Press.
In addition to criticizing Rosenblatt’s perceptron, this book also covers topics related to computational geometry, such as convex hulls,
Voronoi diagrams, and Delaunay triangulations. Further, the book promotes the superiority of symbolic reasoning and discusses its potential
applications of AI, such as image recognition and natural language processing.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 8/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
Olazaran, Mikel. 1993. A sociological history of the neural network controversy. Advances in Computers 37:335–425.
The paper is a thorough historical account of the debate regarding the perceptron and their ability to model complex cognitive processes.
The central theme of the article is that science is a social activity and thus we must understand the debate by taking social factors into
account.
Olazaran, Mikel. 1996. A sociological study of the official history of the perceptrons controversy. Social Studies of Science 26.3:
611–659.
Based on the notion of interpretative flexibility, Olazaran challenged the official explanation that the perceptron was rejected due to its
technical flaws. Instead, the institutionalization and justification of symbolism led to rewriting the history of the debate. Readers who are
interested in exploring alternate views are encouraged to read this article.
Rumelhart, David, and James McClelland. 1986. Parallel distributed processing. Cambridge, MA: MIT Press.
In this book, the connectionist model is defended as a more accurate and realistic method of understanding cognitive processes. The book
paved the way for more advanced neural network models beyond perceptrons in the field of artificial neural networks and cognitive science.
Symbolic Approach to AI
Between the 1950s and 1980s, the dominant paradigm of AI research was the symbolic approach, which was pioneered by cognitive
psychologist Allen Newell, in collaboration with Herbert Simon and John Shaw. The core tenet of symbolism is that AI could be created by
encoding knowledge into symbols and logically manipulating them in a formal system. Newell and Shaw 1957 developed one of the first AI
programming languages, IPL (Information Processing Language), and also the first computer chess program, known as Logical Theorist or
NSS (named after the three inventors; see Newell, et al. 1958). Besides using computers to handle specific tasks, such as chess-playing,
Newell and colleagues were also interested in building a general intelligent system (see Newell, et al. 1959). The fruit of this endeavor is the
General Problem Solver (GPS), an AI program that uses symbolic reasoning to solve problems. The most important contribution of Newell
1980 to the symbolic approach to AI was the physical symbol system, hypothesizing that an intelligent system must have the ability to
manipulate and process physical symbolic representations of information in a structural fashion. In recognition of his contributions to AI, in
1975 Newell was awarded the ACM Turing Award. For a concise summary of Newell’s research, consult Laird and Rosenbloom 1992. There
are several variants of the symbolic approach, and one of the prominent versions is the Adaptive Control of Thought-Rational (ACT-R)
developed by cognitive psychologist John Anderson (see Anderson, et al. 2004). As expected, ACT-R was inspired by the work of Newell.
Anderson argued that in our mental structure there is a long-term declarative memory in the form of a semantic net, or “chunks,” and a long-
term procedural memory in the form of production rules. Interactions between these two types of knowledge lead to complex cognition.
Knowledge is encoded into symbols, which are stored in associative nodes or links to constitute a schema. Following this mental
framework, declarative knowledge must be semantically meaningful and the procedural knowledge must be logical. Similar to Newell’s
theory, in ACT-R information processing is considered goal-directed. Although Anderson’s ACT-R theory combines symbolic processing with
some connectionist mechanisms, the symbolic approach is still the key component of the theory. Earlier in his career Minsky was one of the
vocal advocates of symbolism, but later, in Minsky 1986, he turned away from symbolism to connectionism, yet he did not fully embrace
neural networks.
Anderson, John, Daniel Bothell, Michael Byrne, Scott Douglas, Christian Lebiere, and Yulin Qin. 2004. An integrated theory of the
mind. Psychological Review 111:1036–1060.
This article introduces the evolved version of ACT, namely ACT-R. The modified theory consists of multiple sub-symbolic processes.
Anderson wrote many books and articles on ACT-R, and this is a concise yet comprehensive introduction to the theory.
Laird, John, and Paul Rosenbloom. 1992. The research of Allen Newell. AI Magazine 13:17–45.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 9/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
This article documents how Newell contributed to the development of AI. Newell and his colleagues established many key concepts of AI
and cognitive psychology, including heuristic search, list processing, problem spaces, production systems, and chunking. Both
psychologists and computer scientists should read this article to understand the multidisciplinary nature of AI.
Minsky, Marvin. 1986. The society of mind. New York: Simon & Schuster.
The tenet of this classic book is that an intelligent system can be composed of a vast society of non-intelligent agents. In contrast to his
earlier position, Minsky argued that symbolic AI isn’t capable of producing truly intelligent behavior, and thus a more bottom-up, distributed
method is needed.
Newell, Allen, and John Shaw. 1957. Programming the logic theory machine. In Proceedings of the Western Joint Computer
Conference. Edited by M. M. Astrahan, 230–240. New York: Association for Computing Machinery.
This seminal paper introduces a computer program that can perform automated reasoning, such as combining, modifying, and
decomposing existing mathematical theorems.
Newell, Allen, John Shaw, and Herbet Simon. 1958. Chess-playing programs and the problem of complexity. IBM Journal of
Research and Development 2:320–325.
This article proposed developing a chess-playing program based on heuristic search that can guide useful search and thus reduce the
number of possible moves in a chess game. The paper is of historical interest because later deep learning has provided better results in
playing chess and other games than previous methods.
Newell, Allen, John Shaw, and Herbet Simon. 1959. Report on a general problem-solving program. In Proceedings of the
International Conference on Information Processing. Edited by V. Veronese, 256–264. Paris, France.
This seminal paper introduced a computer program capable of solving problems by means-ends analysis. The idea that problem solving is
goal-directed substantively influenced both cognitive scientists and AI developers.
Connectionist Approach to AI
The term “connectionism” was coined by American psychologist Edward Thorndike in the early twentieth century. Thorndike conceptualized
learning as a result of repeated stimuli and responses, and his ideas contributed to the development of neural networks in AI. Nonetheless,
the term did not become widely used in AI until the 1980s. Despite the popularity of symbolism and Minsky and Papert’s criticism against
the perceptron approach, in the 1970s some researchers insisted that the neural network approach was still viable. One example is a
research consortium in California called the Parallel Distributed Processing (PDP) group. The PDP team consisted of academicians in the
psychology department at the University of California, San Diego (UCSD), and a neuroscientist named Francis Crick from the Salk Institute.
One of the members of research group, David Rumelhart, who was trained in both psychology and mathematics, attempted to remediate
the limitations of the perceptron system by introducing the backpropagation algorithm into a multilayer network. Using the backpropagation
algorithm, also known as backpropation or backprop, the weights of connections between neurons can be automatically adjusted across
multiple layers in order to minimize the residual between the predicted and actual output. At first backpropagation did not work when the
weights were set to zero. A newcomer to UCSD named Geoffrey Hinton, a cognitive psychologist and computer scientist, came to the
rescue. After PDP adopted Hinton’s suggestion to use random weights instead of zero weights, the algorithm started working. The idea of
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 10/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
backpropagation was published in Rumelhart, et al. 1986. Like Rosenblatt, Hinton was also inspired by the learning theory of Donald Hebb.
According to Hebb’s law, learning is a result of neuron firing and wiring, which is the foundational framework of neural works or the
connectionist approach. Based on Hebb’s theory and connectionism, Hinton and Sejnowski 1983 and Hinton, et al. 2000 made several
other significant contributions to AI development, including the Boltzmann Machine, the restricted Boltzmann Machine, and the Capsule
Network. The Boltzmann Machine is a type of neural network that utilizes stochastic processes to learn from data, while the restricted
Boltzmann Machine is a simplified version for easier training. The Capsule Network is an architecture aimed at handling hierarchical
relationships between objects in images and other data formats. These approaches are widely applied in computer vision, speech
recognition, and natural language processing. Due to his significant contributions to AI research, Hinton won the Turing Award in 2019.
Hinton, Geoffrey, Zoubin Ghahramani, and Yee Whye Teh. 2000. Learning to parse images. In Advances in Neural Information
Processing Systems. Edited by S. A Solla, T. K. Leen, and L. Müller, 463–469. Cambridge, MA: MIT Press.
This article proposes a hierarchical approach to image parsing in which a set of filters is used to break down the image into a set of local
features. These local features are then combined into higher-level features that represent more complex structures, such as edges, corners,
and textures. This paper is highly mathematical.
Hinton, Geoffrey, and Terrence Sejnowski. 1983. Analyzing cooperative computation. In Proceedings of 5th Annual Congress of
the Cognitive Science Society. Edited by Patrick Hayes. 81–185. Seattle, WA: Cognitive Science Society.
This article proposes cooperative computation to overcome the limitations of traditional neural networks, which cannot cope with complex
interactions between neurons that are essential to object recognition. This paper is highly mathematical.
Thorndike, Edward. 1932. The fundamentals of learning. New York: Teachers College Press.
In this classic book, Thorndike explained that all learning was based on the formation of connections between stimuli and responses. The
book covers several fundamental principles of learning, including the law of effect, the law of exercise, and the law of readiness.
Rumelhart, David, Geoffrey Hinton, and Ronald Williams. 1986. Learning representations by back-propagating errors. Nature
323:533–536.
This seminal paper describes how the backpropagation algorithm can repeatedly adjust the weights of the connections in a neural network
so that it can minimize the gap between the actual output vector and the expected output vector. This paper is highly mathematical.
The convolutional neural network (CNN), which was invented by Lecun and colleagues (see LeCun and Bengio 1995, LeCun, et al. 2015),
is a type of deep learning algorithm specific to image classification and other computer vision tasks based on the connectionist paradigm.
CNN is inspired by the finding of neuroscience that the human visual system is capable of recognizing objects quickly and accurately. Thus,
CNN was developed to mimic such human capability. LeCun’s early work on CNNs was derived from the “neo-cognitron,” a form of
hierarchical neural network for recognizing handwritten digits. It is necessary to convert the input image into features before CNN can have
a “vision.” For example, if the input is a color image, then the input carries information of three dimensions: the height, width, and depth.
These features are checked by a filter, and this feature-filtering process is called a “convolution.” The convolutional layer is followed by a
pooling layer for subsampling, which allows the information to be reduced in size, simplifying computation. Next, a layer of neurons is used
to perform classification before the output layer is created. LeCun’s early motivation for research on intelligence was sparked by reading the
proceedings of the debate between Noam Chomsky, the linguist who suggested the existence of a universal grammar, and Jean Piaget, the
psychologist who was well known for his work on cognitive development in children. LeCun deemed Chomsky all wrong on AI and tended to
side with Piaget. The debate is documented in Piattelli-Palmarini 1980. For the details of how LeCun was influenced by Piaget, see
O’Connor 2017. Piaget’s theory of cognitive development, which proposed that children actively construct their understanding of the world
through exploring surrounding objects in the environment, influenced LeCun, Bengio, and Hinton’s approach to machine learning. LeCun
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 11/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
and his associates adopted the idea of discovery learning in their work on neural networks. Based on the notion that learning from data is
similar to how infants learn from exploration, they put the machines on the loose in the form of unsupervised meaning. One of the merits of
CNN is that it can do everything automatically without human intervention. Furthermore, LeCun was also influenced by cognitive scientist
Emmanuel Dupoux, particularly in the area of language processing (see Dupoux 2018). Specifically, Dupoux’s work on phonetic categories
and contextual speech recognition inspired LeCun’s research on deep learning models for speech recognition. For an accessible
introduction to CNN, see Taye 2023.
Dupoux, Emmanuel. 2018. Cognitive science in the era of artificial intelligence: A roadmap for reverse-engineering the infant
language-learner. Cognition 173:34–59.
Due to the fact that AI offers powerful learning algorithms that can achieve human-like performance on many linguistic tasks, Dupoux
proposed developing a computational model to mimic how infants acquire linguistic skills by studying raw sensory signals available to
infants.
LeCun, Yann, and Yoshua Bengio, 1995. Convolutional networks for images, speech, and time-series. In The handbook of brain
theory and neural networks. Edited by M. A. Arbib, 225–258. Cambridge, MA: MIT Press.
In this book chapter LeCun and Bengio argue that Convolutional neural networks (CNNs) are well-suited for image recognition. CNNs are
built on several novel ideas, such as local receptive fields, shared weights, and spatial or temporal subsampling. A major advantage of
CNNs is that it is relatively implemented in hardware.
LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. 2015. Deep learning. Nature 521:436–444.
This article introduces how deep convolutional neural nets led to breakthroughs in processing images, video, speech and audio. It also
covers how sequential input can be utilized in recurrent neural networks. The article is highly technical.
O’Connor, William. 2017. Facebook’s robot philosopher king. Daily Beast, 11 July 2017.
This article was published when LeCun had just been hired by Meta (Facebook) to head the AI research lab. This article is a biography of
LeCun, covering how he was inspired by the debate between Piaget and Chomsky.
Piattelli-Palmarini, Massimo, eds. 1980. Language and learning: The debate between Jean Piaget and Noam Chomsky. Cambridge,
MA: Harvard Univ. Press.
The central issue of the debate is whether cognitive structures are innate. Chomsky, who asserted the existence of a universal grammatical
structure, argued that certain cognitive skills cannot be acquired by experience alone, whereas Piaget denied the innateness of any
cognitive structures. Both appealed to logic deduction, rather than empirical science.
Taye, Mohammad Mustafa. 2023. Theoretical understanding of convolutional neural network: Concepts, architectures,
applications, future directions. Computation 11.3: Article 52.
This accessible article provides a step-by-step illustration of how CNN works. The authors also discussed how CNN can be applied to
image recognition, optical character recognition, pricing prediction, spam filtering, fraud detection, healthcare, and transportation.
A major difference between symbolism and connectionism is that in the former, concepts and symbols are one-to-one correspondences,
whereas in the latter, a single concept can be represented by numerous neurons and a neuron can carry information of different concepts.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 12/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
As such, symbolism is limited by the fact that notations must be logically connected in a rule-based system. For example, if A1, then do B1;
else if A2, then do B2; else do B3; and so on. Specifically, the symbolic approach is a production system, which consists of a set of rules or
“productions” for deriving new information from existing information, as illustrated by Rychener and Newell 1977. However, it is difficult, if
not impossible, to build a comprehensive production system that includes millions of rules. Indeed, no system can exhaust every existing
concept or event in the world. Let alone new cases. As a consequence, the computer system may not know what to do when encountering
a new scenario that is never encoded (e.g., A100). In contrast, connectionism is an example-based or knowledge-based system that is
capable of handling new information through pattern recognition. Another major difference between the two approaches is that the former is
confined to sequential processing while the latter can utilize parallel processing. In a symbolic system, deduction is implemented in a step-
by-step fashion. For instance, if A then do B; if B then do C; if C then do D; and so on. But using the connectionist approach, all neurons
learn at the same time, following Hebb’s law. Today, serial computing is limited, whereas high-performance computing for AI and big data
analytics must be parallel processing. Nevertheless, some researchers have argued that the two rival approaches can be integrated to build
a holistic AI system. For example, Mao, et al. 2019 proposed the Neuro-Symbolic Concept Learner that incorporates both symbolism and
connectionism. Specifically, data of visual objects are fed into the AI system for pattern recognition, but at the same time the system is
programmed to categorize symbolic concepts. By doing so, when the system faces a new question in a different scenario, it can recognize
the visual object by mapping it to the concept. Along this line, Hersche, et al. 2023 created the neuro-vector-symbolic architecture that
combines connectionist deep learning and vector-based symbolic computations. Rather than learning from data alone, the system is given
background information. For a comparison between symbolism and connectionism, consult Domingos 2015 and Pfeifer and Scheier 1999.
Domingos, Pedro. 2015. The master algorithm: How the quest for the ultimate learning machine will remake our world. New York:
Basic Books.
In this book, Domingos argued that although symbolic AI has made significant contributions, it has limitations in handling complex and noisy
real-world data. While neural networks from the connectiontist school are powerful and versatile, they often lack transparency and
interpretability, making it difficult to understand why they draw certain conclusions.
Hersche, Michael, Mustafa Zeqiri, Luca Benini, Abu Sebastian, and Abbas Rahim. 2023. A neuro-vector-symbolic architecture for
solving Raven’s progressive matrices. Nature Machine Intelligence 5.4: 363–375.
These authors argue that neither deep neural networks nor symbolic AI alone have reached the level of human intelligence. They propose
the neuro-vector-symbolic architecture (NVSA) by developing a common language between neural networks and symbolic AI. The novelty
of this article is impressive.
Mao, Jiayuan, Chuang Gan, Pushmeet Kohli, Joshua B. Tenenbaum, and Jiajun Wu. 2019. The Neuro-Symbolic Concept Learner:
Interpreting scenes, words, and sentences from natural supervision. arXiv:1904.12584. Ithaca, NY: Cornell University arXiv.
These authors attempted to integrate symbolism and connectionism by proposing the Neuro-Symbolic Concept Learner (NS-CL), a model
that learns visual concepts, words, and semantic parsing of sentences without explicit instruction. This type of synthesis may lead to a new
paradigm of AI.
Pfeifer, Rolf, and Christian Scheier. 1999. Understanding intelligence. Cambridge, MA: MIT Press.
This book provides a comprehensive overview of various approaches to AI, including symbolism, connectionism, Bayesian, and analogist
approaches. The authors propose a coherent framework for studying naturally and artificially intelligent systems, or autonomous agents.
This book is highly technical.
Rychener, Michael, and Allen Newell. 1977. An instructable production system: Basic design issues. Association for Computing
Machinery SIGART Bulletin 63:37–38.
This paper discusses the basics of an instructable production system, such as handling exceptions, generating instructions that are easy to
follow, and evaluating its performance. The authors propose a problem-solving strategy that is based on means-ends analysis and goal-
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 13/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
subgoal search. This paper is highly technical.
Analogist Approach to AI
The psychological root of the analogist approach to AI can be traced back to the nineteenth century. In his research regarding the role of
analogy in human reasoning, William James argued that humans make sense of the world by detecting sameness between different events
and objects (see James 1890). According to Goldstein 2015, modern psychological research suggests that people construct mental models
of the world and use these analogical models to make inferences about new situations. The structure mapping theory developed by
cognitive psychologist Dedre Gentner is a good example (see Gentner 1983). By distinguishing between literal similarity and analogy of
structural properties, this theory aims to improve on previous theories of analogy. According to Gentner, a meaningful analogy maps
relationships or structures, rather than superficial attributes of objects. For example, the solar system and the atom share a structural
commonality, and therefore we can map planets to electrons and the sun to the nucleus. However, the fact that the size of an atom and that
of celestial objects is incomparable doesn’t matter. Cognitive scientist Douglas Hofstadter and psychologist Emmanuel Sander went even
further to assert that all intelligent ideas, varying from common sense to revolutionary theories, can be reduced to analogy (see Hofstadter
and Sander 2013). As such, psychology contributes to the analogist approach to AI by providing insights into how humans solve problems
using analogies and mental models. Nearest-neighbor learning, support vector machines (SVM), recommenders (also known as
recommendation systems), and meta-learning in language translation systems are good examples of how AI researchers developed
algorithms that operate by analogical inferences. Specifically, both the nearest-neighbor algorithm and SVM aim to detect similarities
between different cases. In addition, the recommendation system can guess what the customer would like to buy next based on the
products or services purchased by the same customer or similar customers in the past. In an AI-enabled language translation system,
rather than programming grammatical rules into the system, the system is exposed to a few examples of English-Spanish translation. Next
time when the English sentence is presented to the machine-learning system, it can correctly translate the English sentence into Spanish.
This efficient approach relies on meta-learning, in which a model is trained to use similar examples to perform new tasks through analogical
reasoning. Upon mapping more structures, the meta-learning system becomes proficient at learning how to learn. For the details of meta-
learning, consult Hospedales, et al. 2022.
Gentner, Dedre. 1983. Structure-mapping: A theoretical framework for analogy. Cognitive Science 7:155–170.
This seminal paper explains the mapping principles based on analogy, such as mapping the relations between objects, not the attributes,
and also focusing on high-order relations. The important idea paved the way for the analogist approach to AI.
Goldstein, Bruce. 2015. Cognitive psychology: Connecting mind, research, and everyday experience. 4th ed. Belmont, CA:
Wadsworth.
This accessible textbook of cognitive psychology provides an overview of how analogical reasoning is employed in different kinds of
problem-solving. It also briefly explains how the human neural network functions, though it does not directly compare it to the artificial neural
networks.
Hofstadter, Douglas, and Emmanuel Sander. 2013. Surfaces and essences: Analogy as the fuel and fire of thinking. New York:
Basic Books.
According to Hofstadter and Sander, analogical reasoning can help us solve problems that are ill-defined or unpredictable. The book is
accessible because they use examples from art, literature, music, mathematics, and science, as well as their own research findings, to
illustrate the power of analogy.
Hospedales, Timothy, Antreas Antoniou, Paul Micaelli, and Amos Storkey. 2022. Meta-learning in neural networks: A survey. IEEE
Transactions on Pattern Analysis and Machine Intelligence 44:5149–5169.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 14/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
This article is a general survey of meta-learning, or learning to learn. Deep learning requires extremely big data, thus facing data and
computation bottlenecks. To rectify the situation, the authors propose a new taxonomy of meta-learning. The direction is promising and
awaits further research and development.
Reinforcement learning (RL) in AI was inspired by behaviorism in psychology. Between the 1920s and the mid-1950s, behaviorism was the
dominant school of thought in psychology. During the 1960s its popularity faded with the emergence of the “cognitive revolution.” As the
name implies, behaviorism focuses on observable, external behaviors instead of abstract, internal mental processes. Although cognitive
psychologists disapprove of this idea, AI developers have found it useful. According to Grim 2017, the Turing test, an assessment of a
machine’s ability to exhibit intelligent behavior that is indistinguishable from that of a human, is behavioral in essence. For Turing, whether
the computer truly comprehends the information doesn’t matter. As long as the computer produces seemingly “intelligent” behaviors that
can fool the human, it passes the Turing test. One of the tenets of behaviorism is reinforcement learning, which was conceptualized by
Pavlov in 1903 to illustrate how the response of a dog was strengthened when an unconditioned and a conditioned stimulus were paired
together. Based on Pavlov’s research, John Watson coined the term “behaviorism” in 1913 and titled his 1924 book using the term (see
Watson 1924). The stimulus and response (S-R) model was further developed in Skinner 1971, which emphasized the role of environmental
factors in shaping behavior. In the view of behaviorism, a learning process is inevitably mechanistic. Specifically, if the feedback is
rewarding, it reinforces good behaviors. If not, we avoid detrimental behaviors. Critics charge that this overly simplistic approach ignores
complex human cognition (e.g., bias, emotion, free will, etc.) and individual differences. Later, behaviorism fell out of favor among
psychologists, but computer scientists found it useful, since computers are not influenced by bias, emotion, or free will. A number of RL
algorithms were developed in the 1970s and 1980s for solving simple problems, such as controlling robots and playing games. A key
breakthrough is the Q-learning algorithm developed by Watkins 1989, which involves teaching an agent to make decisions based on the
maximum expected value of an action given a particular state. In the 1990s and the early twenty-first century, more sophisticated algorithms
were developed for robotics, game playing, and control systems. A notable example is the TD-Gammon algorithm, created by Tesauro and
his associates (see Tesauro 1992), which learned to play backgammon by competing against itself and evaluating the results.
Skinner, Burrhus. 1971. Beyond freedom and dignity. Indianapolis, IN: Hackett.
This book is highly controversial. According to Skinner, free will is illusory because human behavior is mainly influenced by environmental
factors and being modified by reinforcements. Moreover, he advocated using scientific methods to shape desirable humans, thus putting
aside dignity and individual differences.
Tesauro, Gerald. 1992. Practical issues in temporal difference learning. Machine Learning 8:257–277.
Tesauro explained how temporal difference (TD) entails adjusting the value of each possible move to better predict future rewards. He
argued that the TD method could be successfully applied to complex real-world problems, and could outperform many other conventional
commercial programs. The paper is highly mathematical.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 15/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
Watkins, Christopher. Learning from delayed rewards. PhD diss., Univ. of Cambridge, England, 1989.
This dissertation details how RL-based Q-learning utilizes an incremental method for dynamic programming that requires limited computing
resources only. The dissertation has many details on computation.
In the twenty-first century, major breakthroughs of reinforcement learning (RL) have been accomplished by Google’s Greatmind, NVIDIA,
and many others. AlphaZero, AlphaGo, and AlphaStar, developed in Google’s DeepMind, which defeated human chess experts and video
game players, are all based on RL. Specifically, AlphaZero acquired world-class game-playing skills without prior knowledge of the game
rules. By self-playing, AlphaZero learned the basic rules of the game instead of being programmed with explicit rules. During the 2017
chess tournament, AlphaZero defeated Stockfish, the then world champion chess program. In 2016 and 2017, AlphaGo defeated two top
Go players. In 2019, AlphaStar won a series of matches with professional StarCraft II players. These stunning accomplishments based on
RL are documented in Mnih, et al. 2015; Silver, et al. 2016; and Vinyals, et al. 2019. In addition, NVIDIA, the world’s leader in graphical
processing units (GPU), developed specialized hardware and software for RL, such as the NVIDIA Tensor Cores and the NVIDIA Deep
Learning SDK (see NVIDIA 2023). These technologies can accelerate the computation of matrix operations, which are essential for many
RL algorithms. RL is also used in economics and game theory, in which the optimal outcome is determined by the interaction between
different agents. What distinguishes RL from other learning algorithms is that rather than yielding the lowest short-term loss, RL aims to
maximize long-term gain. Take playing chess or Go as an example. The ultimate goal of playing the game is to win. If the machine learning
algorithm focuses too much on the immediate reward or the minimal loss at each step, at the end the player might win many battles but lose
the war. In contrast, RL is a far-sighted algorithm in the sense that its optimization takes unseen consequences into account. For example,
Move 37 of AlphaGo against the world champion in Game 2 seemed absurd to humans, but after a hundred moves people realized that that
earlier move led to AlphaGo’s ultimate victory. In spite of the preceding merits and successful examples, Botvinick, et al. 2019 argues that
RL is inefficient, because to attain expert human-level performance, RL systems have to acquire much more data than their human
counterparts. For the details of the relationships between RL, psychology, and neuroscience, see Subramanian, et al. 2022.
Botvinick, Matthew, Sam Ritter, Jane X. Wang, Zeb Kurth-Nelson, Charles Blundell, and Demis Hassabis. 2019. Reinforcement
learning, fast and slow. Trends in Cognitive Sciences 23:408–422.
These authors propose several solutions to overcome some weaknesses of RL, including its slowness in learning. They also discuss the
implications of this insight for psychology and neuroscience. This interdisciplinary article is beneficial to both computer scientists and
psychologists.
Mnih, Volodymyr, Koray Kavukcuoglu, David Silver, et al. 2015. Human-level control through deep reinforcement learning. Nature
518:529–533.
Mnih and his team describe how a particular type of RL network—namely, the Q-network—is able to learn by acquiring multidimensional
data. This approach was tested by playing Atari 2600 games. The paper is highly technical.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 16/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
Silver, David, Aja Huang, Chris J. Maddison, et al. 2016. Mastering the game of Go with deep neural networks and tree search.
Nature 529:484–489.
In this paper, Silver and his team document how the RL program AlphaGo outperformed other Go programs by 99.8 percent, and also
defeated an expert human Go player by 5 games to 0. The paper is highly technical.
Subramanian, Ajay, Sharad Chitlangia, and Veeky Baths. 2022. Reinforcement learning and its connections with neuroscience and
psychology. Neural Networks 145:271–287.
These authors argue that the accomplishment of RL is attributed to the findings in neuroscience, psychology, and animal learning. In this
paper, various RL models were mapped to specific findings in neuroscience and behavioral literature. Researchers who are interested in
multidisciplinary research are recommended to read this paper.
Vinyals, Oriol, Igor Babuschkin, Wojciech M. Czarnecki, et al. 2019. Grandmaster level in StarCraft II using multi-agent
reinforcement learning. Nature 575:350–354.
In this paper, Vinyals and his team document how general-purpose learning methods, such as a multi-agent reinforcement learning
algorithm, can be used to solve complex problems. The paper is highly technical.
Artificial intelligence has significantly influenced the research agendas of various branches of psychology by providing new perspectives,
methodologies, and avenues for investigation. For example, social psychologists study topics related to how AI affects human thoughts and
perceptions, whereas industrial-organizational psychologists examine challenges and opportunities presented by human-AI collaboration
and interaction. In quantitative psychology, machine learning methods are employed to overcome limitations of traditional statistical
approaches, such as overfitting the model to a particular sample. In cognitive psychology, AI models and techniques, such as neural
networks and deep learning, allow more accurate simulation of cognitive processes like perception, memory, decision-making, and problem-
solving, improving our understanding of the human mind. For clinical psychologists, AI has begun to assist in diagnosing mental health
disorders, predicting treatment outcomes, and offering personalized interventions. In addition, numerous applications of AI in educational
psychology seek to enhance teaching and learning. Educators can employ AI to develop personalized instruction tailored to individual
learning styles and competence. Further, AI can improve the accuracy and efficacy of forensic investigations via predictive analytics, risk
assessment, textual analysis, and data mining. AI even inspires a new subfield of psychology, artificial psychology, which studies cognitive
architectures and artificial emotions in AI systems by integrating psychology and information science.
Social Psychology
Social psychologists study various topics related to AI, including how AI affects human thoughts and how humans perceive AI. As an
example of the former, recommendation algorithms of social media constantly feed users one-sided information based on what the users
selectively read or watch in the past, and consequently social media users are trapped in a filter bubble or an echo chamber. The term “filter
bubble,” coined by Pariser 2012, refers to a situation in which Internet users only absorb information that confirms their preexisting beliefs.
As a consequence, they live in a self-contained bubble. It is difficult to trace the exact origin of the term “echo chamber.” The term uses an
analogy of the sound of reverberation in a small room to describe how people hear their own voice only. This phenomenon has led
psychologists to study which personalities and ideological orientations are more prone to filter bubbles and echo chambers. Sindermann, et
al. 2020 found that participants with lower conscientiousness and higher Neuroticism, measured by the Big-Five Inventory, and subscribe to
Right-Wing Authoritarianism, tended to consume news via social media only. Luzsa and Mayr 2021 goes further to assert that social media
news feeds generate a false consensus. Participants inferred public support for their own convictions from the level of agreement in online
messages, but were skeptical of “likes,” especially if they were highly interested in the topic. Psychologists have developed remedies to
rectify the situation. For instance, Ognibene, et al. 2022 proposed the “Social Media Virtual Companion” in an attempt to educate users how
to properly navigate social media. The companion will be powered by a recommender aiming to promote collective well-being, instead of
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 17/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
maximizing corporate profit. How humans perceive and react to AI is another important topic in social psychology. Li and Sung 2021
indicates that people tend to anthropomorphize AI assistants positively, even though those AI assistants have no human attributes. Similarly,
Zhang, et al. 2022 found that participants perceived AI as more likely to make utilitarian choices than human beings are. However, Castelo,
et al. 2019 observes that humans only partially trust AI. Specifically, people trust AI algorithms more when they process factual information,
rather than making subjective judgments. The finding of Longoni, et al. 2019 is more extreme. It was found that people tended to rule out AI
for all medical care, ranging from prevention to diagnosis to treatment.
Castelo, Noah, Maarten W. Bos, and Donald R. Lehmann 2019. Task-dependent algorithm aversion. Journal of Marketing Research
56:809–825.
In this study, the authors examined whether users trust algorithms when it comes to decision support in a variety of tasks. The sample size
is 57,400, which supports the generalizability of the study. While classical statistical procedures were used for the analysis (e.g., logistic
regression), data visualization techniques revealed the patterns of the data.
Li, Xinge, and Yongjun Sung. 2021. Anthropomorphism brings us closer: The mediating role of psychological distance in user-AI
assistant interactions. Computers in Human Behavior 118: Article 106680.
This empirical study showed that participants anthropomorphized and evaluated their AI assistants more favorably, and this effect was
mediated by psychological distance, a subjective concept related to closeness to oneself. Using these findings to design user interfaces can
be beneficial.
Longoni, Chiara, Andrea Bonezzi, and Carey Morewedge. 2019. Resistance to medical artificial intelligence. Journal of Consumer
Research 46:629–650.
These researchers documented a reluctance to use medical care delivered by AI providers across a variety of medical decisions, since
patients believe that AI providers are unable to take into account each patient’s unique situation. Social psychologists should seek solutions
to this problem based on the findings of this study.
Luzsa, Robert, and Susanne Mayr. 2021. False consensus in the echo chamber: Exposure to favorably biased social media news
feeds leads to increased perception of public support for own opinions. Cyberpsychology: Journal of Psychosocial Research on
Cyberspace 15: Article 3.
In this paper the authors conducted two experiments to confirm that online echo chambers can influence the false consensus effect.
Although the result is not surprising, the magnitude is alarming. The sample size is relatively small (n1 = 331, n2 = 207), and thus social
psychologists can consider replicating this type of research.
Ognibene, Dimitri, Rodrigo Wilkens, Davide Taibi, et al. 2022. Challenging social media threats using collective well-being-aware
recommendation algorithms and an educational virtual companion. Frontiers in Artificial Intelligence 5:654930.
In contrast to popular belief that social media users are passively manipulated, these authors asserted that the impact of social media is
resulted from the complex interplay between the users’ interactions and the intelligent components of the platform.
Pariser, Eli. 2012. The filter bubble: What the Internet is hiding from you. London: Penguin Books.
Pariser documented how tech giants, including Google, Facebook, Apple, and Microsoft, fed selective information to users in order to
maximize their profit. Although this book was published before the advance of AI, the concept of “filter bubble” led to subsequent debates
and research on echo chambers, resulting from AI-enabled recommendation systems.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 18/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
Sindermann, Cornelia, Jon Elhai, Morten Moshagen, and Christian Montag. 2020. Age, gender, personality, ideological attitudes
and individual differences in a person’s news spectrum: How many and who might be prone to filter bubbles and echo chambers
online? Heliyon 6:e03214.
Based on a sample of 1,681 participants, this study showed how filter bubbles and echo chambers affect users psychologically. Despite the
fact that the authors used traditional statistics instead of machine learning methods in their analysis, clear patterns were still revealed by
their data visualization.
Zhang, Zaixuan, Zhansheng Chen, and Liying Xu. 2022. Artificial intelligence and moral dilemmas: Perception of ethical decision-
making in AI. Journal of Experimental Social Psychology 10: Article 104327.
This empirical study found that people perceive AI to make more utilitarian choices than humans do. This finding has many practical
implications. For example, in the past the trolley problem was just a thought experiment, but the invention of AI-enabled self-driving vehicles
has turned it into a real problem.
Industrial-Organizational Psychology
User interface or user experience is one of the major domains of industrial-organizational (I/O) psychology. With the advance of AI robots,
this research tradition has extended to research on human-AI interaction. Specifically, I/O psychologists study challenges and opportunities
of human-AI collaboration, including issues related to trust, communication, and decision-making. For example, Nam and Lyons 2020
argues that trust is a crucial factor in the success of human-robot interaction, and thus designing for trust is crucial for developing effective
and socially acceptable AI-enabled robots. Takayama 2012 reports on extensive research on the psychology of human-computer interaction
(HCI) and human-robot interaction (HRI). Takayama advocates a human-centered approach to design HCI and HRI systems, such as
focusing on the needs, preferences, and behaviors of users in order to create systems that are intuitive, engaging, and effective. She also
emphasized the importance of considering the ethical and social implications of new technologies so that they can promote human well-
being and autonomy. Bias in AI is another hot topic sought after by I/O psychologists. Today, AI-powered predictive models are prevalent,
but sometimes their transparency and accuracy are questionable. For example, recommendation systems are often used for high-stakes
decision-making, such as résumé screening, hiring decisions, college admission, parole review, and many others, but in many situations the
outcomes are incorrect and unfair. Bias due to sampling errors is a good example. Earlier facial recognition systems were not capable of
analyzing darker faces because the databases store more white people’s facial features than the information of other ethnic groups. By the
same token, in the first version of the AI art generation tool DALLE, the system returned images of all white males when the user requested
pictures of a lawyer. These cases are documented in Buolamwini and Gebru 2018; Gonzalez, et al. 2019; Kantayya 2020; O’Neil 2016; and
Marcinkowski, et al. 2020. In response, psychologists have investigated how these biases happen and how they can be mitigated or
eliminated. For example, Landers and Behrend 2023, by two I/O psychologists, argues that AI-empowered predictive models for high-stake
decision support must go through rigorous and thorough audits. The auditing process needs to involve multiple auditors, including internal
auditors, external auditors from consulting firms, and independent auditors from a regulatory authority. In addition, the criteria should
encompass various aspects of the AI system, including data quality, model design, model development, model validation, generalizability,
cultural context, and many others.
Buolamwini, Joy, and Timnit Gebru. 2018. Gender shades: Intersectional accuracy disparities in commercial gender classification.
Proceedings of Machine Learning Research 81:1–15.
This study indicated that darker-skinned females are more likely to be misclassified by facial recognition algorithms, and the error rates are
up to 34.7 percent. It is noteworthy that the second author, Gebru, was a Google researcher. She claimed that once Google forced her to
withdraw a paper on bias of language models.
Gonzalez, Manuel, John Capman, Frederick Oswald, Evan Theys, and David Tomczak. 2019. Where’s the I-O? Artificial intelligence
and machine learning in talent management systems. Personnel Assessment and Decisions 5:3.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 19/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
The authors present empirical evidence that AI bias can negatively impact job applicants. To rectify the situation, the authors advocate
collaboration among I/O psychologists, computer scientists, legal scholars, and other professionals in developing, implementing, and
evaluating AI/ML applications. This paper will be of interest to readers interested in multidisciplinary research.
Kantayya, Shalini, dir. 2020. Coded bias. Brooklyn, NY: 7th Empire Media.
This documentary exposed the hidden biases embedded in AI algorithms that perpetuate discrimination and inequality. One infamous case
is that facial recognition technology was less accurate in identifying darker-skinned individuals and women. There are many thought-
provoking case studies in this film that will be of interest to social and I/O psychologists.
Landers, Richard, and Tara Behrend. 2023. Auditing the AI auditors: A framework for evaluating fairness and bias in high stakes AI
predictive models. American Psychologist 78:36–49.
These authors argue that it is difficult to counteract AI bias because different fields define these concepts differently. For example, traditional
statistics equates bias with sampling errors while ML focuses on the trade-off between bias and variance. They suggest AI systems must be
thoroughly inspected in various aspects through multiple perspectives.
Marcinkowski, Frank, Kimon Kieslich, Christopher Starke, and Marco Lünich. 2020. Implications of AI (Un-)fairness in higher
education admissions: The effects of perceived AI (Un-)fairness on exit, voice and organizational reputation. In Proceedings of
the Conference on Fairness, Accountability, and Transparency. Edited by Mireille Hildebrandt and Carlos Castillo, 122–130. New
York: Association for Computing Machinery.
Using survey data (n = 304) obtained from a German university, these authors address fairness concerns in academic settings. The authors
suggest that it is essential to consider possible fairness issues and their further implications when universities want to use AI systems for
selection and decision support.
Nam, Chang, and Joseph B. Lyons, eds. 2020. Trust in human-robot interaction. Cambridge, MA: Academic Press.
The comprehensive book includes contributions from various researchers in the field of I/O psychology and engineering to examine how
trust influences robot design and development. It was found that humans trust robots based on their appearance, behavior, and
performance, among other factors.
O’Neil, Cathy. 2016. Weapons of math destruction: How big data increases inequality and threatens democracy. New York: Crown.
The author studied how algorithms contribute to social inequalities and unfairness. She claimed that some algorithms in educational,
criminal justice, and job recruitment sectors have led to the creation of “weapons of math destruction.” It is noteworthy that not all examples
cited in the book are related to AI/ML.
Takayama, Lelia. 2012. Perspectives on agency: Interacting with and through personal robots. In Human-computer interaction:
The agency perspective. Edited by Marielba Zacarias and José Valente de Oliveira, 195–214. New York: Springer.
This book chapter examines the concept of agency in the context of human-robot interaction. Agency is defined as the ability of a robot to
act autonomously and make decisions based on its own internal processes and external environment. The information is beneficial to both
I/O psychologists, social psychologists, and industrial engineers.
Quantitative Psychology
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 20/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
Traditional statistical methods for quantitative psychology have been criticized for several shortcomings, as documented in Yu 2022.
Machine learning (ML) methods can benefit quantitative psychology by either overcoming those limitations or supplementing traditional
statistics for triangulation. For example, the conclusion based on hypothesis testing is inevitably dichotomous: either to reject or not to reject
a null hypothesis according to the p value. When alpha =.05 is adopted, not only is the final decision dichotomous, but also the evidence is
binary. In contrast, ML seeks to identify the patterns and trends in data, instead of following a definite cutoff. In traditional statistics, usually
the researcher starts with a preformulated hypothesis or a strong model, and then collects data to confirm the model. This model-based
approach is prone to confirmation bias and also requires parametric assumptions derived from mathematical models. In contrast, most ML
methods are data-driven and assumption-free, which is in alignment with the methodological traditions of Breiman 2001, Chambers 1993,
Cleveland 2001, and Tukey 1977. The starting point in traditional statistics is usually a single model, and then the model is improved by
remediate assumption violations, if there are any. On the other hand, data scientists usually use a variety of ML modeling techniques, such
as neural networks, support vector machines, boosting, and bagging. In the end, the final model is obtained by model averaging or model
selection. In traditional statistics, usually the entire sample is included in the modeling process. As a result, this approach is prone to
overfitting, thus resulting in the replication crisis in psychology, as documented in Open Science Collaboration 2015. To compensate for this
shortcoming, most ML methods employ cross-validation, bootstrapping, or both for repeated analyses, thus enhancing generalizability.
When Pearson, Fisher, and other statistical theorists developed their methodologies during the late nineteenth and early twentieth
centuries, these methods were suited to small-sample studies. Today, psychology is increasingly using big data, but as demonstrated by
Fan, et al. 2014, certain classical statistical procedures are no longer applicable to large samples. Nevertheless, modern ML algorithms,
such as neural networks, bagging, and boosting, are capable of big data analytics. Further, while traditional statistics focuses on structured
data, natural language processing, a subdomain of AI and computational linguistics, is extensively used in text mining for analyzing
unstructured data. For more information about text mining, consult Ignatow and Mihalcea 2017.
Breiman, Leo. 2001. Statistical modeling: The two cultures. Statistical Science 16:199–231.
In this seminal article, Breiman states that there are two cultures in statistical modeling. The traditional approach looks for inferring from
sample statistics to the population parameters based on the assumption that the data are stochastically generated. Conversely, the
prediction-centric approach uses algorithmic models without assuming the data mechanism.
Chambers, John. 1993. Greater or lesser statistics: A choice for future research. Statistics and Computing 3:182–184.
In this article, Chambers argues that lesser statistics is confined to academia; its focus is mathematical techniques, and collaboration with
other disciplines is rare. In contrast, greater statistics is inclusive, closely related to other disciplines, and practiced by analysts outside of
academia. This notion paves the way to modern DSML.
Cleveland, William. 2001. Data science: An action plan for expanding the technical areas of the field of statistics. International
Statistical Review 69:21–26.
In this seminal article, Cleveland proposed a new discipline that synthesizes modern computing methods and statistics. He suggested that
statisticians should incorporate advanced computing into data analysis and learn how to engage with subject matter experts. This article is
one of the precursors of data science and machine learning.
Fan, Jianqing, Fang Han, and Han Liu. 2014. Challenges of big data analysis. National Science Review 1:293–314.
These authors demonstrated that many assumptions in most statistical methods applied to big data cannot be validated. As a result, they
can lead to incorrect statistical inferences and consequently wrong scientific conclusions. Their arguments are understandable to those who
are trained in traditional statistics.
Ignatow, Gabe, and Rada Mihalcea. 2017. Text mining: A guidebook for the social sciences. Thousand Oaks, CA: SAGE.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 21/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
This book covers a variety of topics related to text mining, including Web scraping, sentiment analysis, and natural language processing. As
this book is intended for social scientists, psychologists should find it useful.
Open Science Collaboration. 2015. Estimating the reproducibility of psychological science. Science 349.6251.
This project indicated that in many replicated studies the results were not as strong as the original reports in terms of significance (p-values)
and magnitude (effect sizes). Specifically, 97 percent of the original studies reported significant results (p <.05), but only 36 percent of the
replicated studies yielded significant findings.
Yu, Chong Ho. 2022. Data mining and exploration: From traditional statistics to modern data science. Boca Raton, FL: CRC Press.
This book discusses the limitations of traditional statistics and how machine learning can remediate those shortcomings. The book
introduces various ML techniques, including neural networks, support vector machines, boosting, bagging, and text mining. The book is
conceptual and procedural, rather than mathematical.
Cognitive Psychology
AI and cognitive psychology have a symbiotic relationship, with each field enriching the other. By providing insight into how the brain
processes information, cognitive psychology can help AI algorithms mimic intelligence. Conversely, cognitive psychologists can use AI to
model and simulate brain functions, allowing them to investigate complex cognitive processing. Adaptive Control of Thought-Rational (ACT-
R), developed by John Anderson (see Anderson 1995; Anderson 2012; and Anderson, et al. 2004), is a well-known cognitive theory that
describes how humans perceive, learn, and reason based on the interaction between declarative and procedural knowledge. In cognitive
psychology, ACT-R has been widely used to develop computational models of human cognition for testing and refining theories about how
humans process information, as well as predicting and testing human performance on a variety of cognitive tasks. Additionally, by modeling
the cognitive processes of learning and problem-solving, ACT-R has been used for developing intelligent tutoring systems that can provide
customized feedback and guidance to different learners. Moreover, by anticipating how users will interact with complex systems, ACT-R has
been employed for the design of human-machine interfaces. Parallel Distributed Processing (PDP), a connectionist model developed by
Rumelhart and McClelland also made a significant impact on cognitive psychology, especially in the areas of perception, memory, and
language (see Rumelhart and McClelland 1986a and Rumelhart and McClelland 1986b). According to PDP, representations of information
are distributed throughout numerous nodes of the entire cognitive network. When ambiguous pieces of a puzzle are being put together in a
parallel fashion instead of serial processing, the whole picture will be clear from the induction. The idea that problem-solving as filling in the
gaps of partial information can be traced to German psychologist Otto Selz. By modeling cognitive processes as a network of nodes, PDP
can explain mental processes that were challenging to previous theories, such as how the brain processes ambiguous or incomplete
information. Furthermore, by modeling memory as a network of interconnected nodes, PDP reveals how memories can be stored and
retrieved based on patterns of activation within the network. In addition, PDP shows how language processing can be influenced by both
top-down processes and bottom-up processes. State, Operator, And Result (SOAR), introduced in Newell 1990 and further developed in
Laird 2019, is another AI-related cognitive architecture that led to many practical applications in cognitive psychology. SOAR inspired
cognitive psychologists to treat problem-solving as a process of searching through the possible solution space. It was also utilized in
building intelligent training systems for air-combat simulations.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 22/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
This book provides a comprehensive overview of the ACT-R cognitive architecture, including its theoretical foundations, key components,
and applications. The cognitive model is called “adaptive” because Anderson believed that the complexity of intelligence is developed in
response to the complexity of the environment. First published 1983.
Anderson, John. 2012. The atomic components of thought. London: Psychology Press.
This sequel provides an in-depth exploration of the building blocks of the ACT-R architecture, including the different types of memory, the
processing mechanisms, and the learning algorithms. The atoms of thought include processes like comparing pieces of information. These
simple processes can be combined to enable complex thought. First published 1998.
Anderson, John, Daniel Bothell, Michael Byrne, Scott Douglas, Christian Lebier, and Yulin Qin. 2004. An integrated theory of the
mind. Psychological Review 111:1036–1060.
This article presents a detailed description of the ACT-R architecture and its components, including the declarative and procedural memory
systems and the problem-solving module. The authors argued that most human learning involves tuning sub-symbolic processes. These
mental submodules can function alone or in concert with others.
Laird, John. 2019. The SOAR cognitive architecture. Cambridge, MA: MIT Press.
This book provides an overview of the SOAR architecture, including knowledge-intensive reasoning, reactive execution, hierarchical
reasoning, planning, and learning from experience. SOAR represents knowledge as production rules that operate on a working memory to
transform one state to another. The goal is to create a general computational system that leads to AGI. First published 2012.
Newell, Allen. 1990. Unified theories of cognition. Cambridge, MA: Harvard Univ. Press.
In this classic book, Newell argued that cognitive processes are a collection of information-processing operations that can be simplified by
working memory, production rules, and chunking. Newell’s unified theory consists of a long-term memory for knowledge storage, a working
memory for situational data, and a decision cycle for rule selection.
Rumelhart, David, and James McClelland. 1986a. Parallel distributed processing. Vol. 1, Explorations in the microstructure of
cognition: Foundations. Cambridge, MA: MIT Press.
This book shed light on AI research by proposing the massively parallel architecture of the human mind. It laid the foundation for
connectionism, which challenged the symbolic approach to AI and cognitive psychology.
Rumelhart, David, and James McClelland. 1986b. Parallel distributed processing. Vol. 2, Psychological and biological models.
Cambridge, MA: MIT Press.
While Volume 1 outlined the theory of connectionism, Volume 2 applied it to specific topics in cognitive psychology and neuroscience, such
as perception, memory, language, and thought.
Clinical Psychology
Through a systematic review, Shatte, et al. 2019 identifies four domains of mental health that have been benefited by AI: detection and
diagnosis; prognosis, treatment and support; public health applications; and research and clinical administration. First, traditional diagnostic
methods highly depend on expert opinions, and as a result, its inter-rater reliability is often challenged by skeptics. To rectify the situation, AI
has been utilized by mental health professionals to diagnose patients based on a variety of data, such as clinical records, behavioral data,
and many other nontraditional data sources. For example, Bzdok and Meyer-Lindenberg 2018 found that analyzing complex patterns in
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 23/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
brain, behavior, and genes by machine learning methods, such as support vector machines, neural networks, and cross-validation, can
better predict treatment outcomes than both the Diagnostic and Statistical Manual of Mental Disorders (DSM) and the International
Classification of Diseases (ICD). In a similar vein, Perna, et al. 2018 argues that AI can be used to facilitate personalized psychiatry.
Second, ML-based analytics can help clinicians assess the effectiveness of treatment outcomes. For example, although cognitive-
behavioral therapy (CBT) alone or in combination with pharmacotherapy is the first choice for treating patients suffering from panic disorder
with agoraphobia, Sundermann, et al. 2017 found that the use of CBT did not reliably predict anxiety reduction based on support vector
machine analysis. Third, AI can be well-applied to demographic-oriented public health analysis. For example, utilizing Web scraping, K-
means clustering, and text mining of Twitter data, Chary, et al. 2017 found that mentions of misuse of prescription opioids on Twitter by state
is strongly correlated with the geographical patterns revealed by National Surveys on Drug Usage and Health. Last, AI enhances mental
health research and administration by improving research methodologies and resource allocation. In the past, research conducted by
mental health researchers was limited by small samples, narrow sources, and structured data. These barriers had been broken by the
availability of big data and the advance of ML methods that can accommodate various data types. According to Graham, et al. 2019, the
most common data types used in AI studies of mental health are electronic health records, mood rating scales, brain imaging data, novel
monitoring systems (e.g., smartphone, video), and social media platforms (e.g., Twitter). Further, for cost management, Roysden and Wright
2015 used random forests to develop a decision support system that can detect patients who may benefit from referral.
Bzdok, Danilo, and Andreas Meyer-Lindenberg. 2018. Machine learning for precision psychiatry: Opportunities and challenges.
Biological Psychiatry: Cognitive Neuroscience and Neuroimaging 3:223–230.
According to the authors, there is growing suspicion that traditional disease categories may misrepresent the causes of mental disturbance.
The paper provides clinicians and researchers with an overview of the opportunities and challenges of improving diagnosis using AI.
Chary, Michael, Nicholas Genes, Christophe Giraud-Carrier, Carl Hanson, Lewis Nelson, and Alex Manini. 2017. Epidemiology
from tweets: Estimating misuse of prescription opioids in the USA from social media. Journal of Medical Toxicology 13:278–286.
In this study the authors addressed a well-studied public health issue: abuse of opioids. What set this paper apart from others is that by
employing AI/ML methods to analyze social media data, they were able to canvas a larger segment of the population and yield timely
insights for syndromic monitoring.
Graham, Sarah, Colin Depp, Ellen Lee, Camille Nebeker, Xin Tu, Ho-Cheol Kim, and Dilip Jeste. 2019. Artificial intelligence for
mental health and mental illnesses: An overview. Current Psychiatry Reports 21: Article 116.
The panel reviewed twenty-eight mental health studies using AI/ML. It was concluded that while some studies revealed high accuracies and
might be promising, most should be treated as early proof-of-concept works.
Perna, Giampaolo, Massimiliano Grassi, Daniela Caldirola, and Charles Nemeroff. 2018. The revolution of personalized psychiatry:
Will technology make it happen sooner? Psychological Medicine 48:705–713.
The authors advocate using high technologies to support clinical decision-making based on individual profiles. While AI/ML is taken into
consideration, they also covered other data-related technologies, such as big data and real-time data streaming. Readers who are
interested in a new and promising clinical paradigm are encouraged to read the paper.
Roysden, Nathaniel, and Adam Wright. 2015. Predicting health care utilization after behavioral health referral using natural
language processing and machine learning. AMIA Annual Symposium Proceedings 2015:2063–2072.
These authors used random forest classifiers for predicting whether a patient needs a referral, based on structured administrative data, lab
results, and free text notes as input. While this approach is beneficial to cost management, using AI to make high-stake decisions is
contentious.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 24/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
Shatte, Adrian, Delyse Hutchinson, and Samantha Teague. 2019. Machine learning in mental health: A scoping review of methods
and applications. Psychological Medicine 49:1–23.
This is a systematic and comprehensive review of how AI/ML has been applied to mental health. Many of the references are ML-related, but
some are actually based on traditional statistics, such as regression analysis.
Sundermann, Benedikt, Jens Bode, Ulrike Lueken, et al. 2017. Support vector machine analysis of functional magnetic resonance
imaging of interoception does not reliably predict individual outcomes of cognitive behavioral therapy in panic disorder with
agoraphobia. Frontiers in Psychiatry 8.
Unlike traditional hypothesis testing that yields a dichotomous answer (reject or not to reject the null hypothesis), this study used ML
methods to indicate predictive accuracy. Additionally, the results were verified with cross-validation and permutation testing.
Educational Psychology
Artificial intelligence has been applied to educational psychology in various ways to improve the effectiveness of teaching and learning.
First, educators can use AI to develop personalized learning based on individual student data, including their learning styles and
competence. This technology can also provide real-time feedback to students and teachers, allowing for ongoing formative evaluation and
immediate adjustments to instruction. For example, utilizing student’s ability and attention scores, Wei, et al. 2021 developed a personalized
online learning recommender to improve students’ learning outcomes. The theoretical framework of such a system is Zone of Proximal
Development, which aims to provide students with learning materials that are neither too easy nor too difficult. Student engagement, which
refers to the level of interest, curiosity, enthusiasm, and passion during the learning activities, is an important concept of educational
psychology. Traditionally, engagement monitoring is done by manually analyzing videotaped lectures or self-report data provided by
students. In contrast, AI has the capability of automating this tedious process, as well as producing more accurate results. For example, the
machine vision–based approach developed by Goldberg, et al. 2019, which is based on gaze, head pose, and facial expressions, provides
good estimations of the manual ratings of student engagement. In addition, educational psychologists can utilize ML methods to analyze
large data sets to identify patterns and trends of student behavior and performance. This information can be used to develop new
interventions to improve learning outcomes. For example, Yu, et al. 2019 analyzes the Programme for International Assessment of Adult
Competencies data set (n = 126,324) using ML procedures, such as clustering, bagging, and boosting. It was found that learning outcomes
are strongly predicted by cultural engagement, readiness to learn, and social trust. While educators are aware of the preceding merits of AI,
they are also concerned with whether AI technologies are used ethically and responsibly. ChatGPT, a large language model released in
2022, is capable of writing articles, generating codes, and solving complex math problems. While some educators have reported the
advantages of ChatGPT and even provided guidelines on how to use it in the classroom, some schools have banned its use because many
students used ChatGPT to cheat on assignments and exams. As quick answers are easily accessible through ChatGPT, some educational
psychologists fear that it could hinder students’ cognitive development. For a discussion of pros and cons of using ChatGPT in educational
settings, see Tlili, et al. 2023.
Goldberg, Patricia, Ömer Sümer, Kathleen Stürmer, et al. 2019. Attentive or not? Toward a machine learning approach to
assessing students’ visible engagement in classroom instruction. Educational Psychology Review 33:27–49.
These researchers developed a machine vision–based approach to evaluate students’ attentiveness and engagement. It is noteworthy that
the study was conducted in three lessons, and thus it provides only a proof of concept.
Tlili, Ahmed, Boulus Shehata, Michael Agyemang Adarkwah, et al. 2023. What if the devil is my guardian angel: ChatGPT as a case
study of using chatbots in education. Smart Learning Environments 10:1–24.
This study, which includes different perspectives from different parties, is a fair and balanced assessment of the pros and cons of ChatGPT
in educational settings. It presents several research directions that should be taken into consideration to ensure a responsible and safe
adoption of AI tools in education.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 25/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
Wei, Xin, Shiyun Sun, Dan Wu, and Liang Zhou. 2021. Personalized online learning resource recommendation based on artificial
intelligence and educational psychology. Frontiers in Psychology 12: Article 767837.
In this study both AI technology and educational psychology theory are applied for developing a personalized online learning
recommendation system to enhance academic performance. While recommendation systems are widely used in e-commerce and social
media, this application is relatively new in education.
Yu, Chong Ho, Hyun Seo Lee, Emily Lara, and Siyan Gan. 2019. Adult learning across USA, Canada, and New Zealand: A cross-
cultural study of PIAAC. International Education Studies 12:1–16.
The sample size of this study is large (n = 127,757) and thus it is challenging to traditional statistics. Using DSML methods that can avoid
overfitting, the authors identified cultural and social variables contributing to test performance of literacy, numeracy, and technology-based
problem-solving in PIAAC.
Forensic Psychology
AI can be applied to forensic psychology in various ways to improve the accuracy and effectiveness of forensic evaluations and
investigations, including predictive analytics, risk assessment, textual analytics, and data mining. In predictive analytics for policing, AI is
used to analyze geographical patterns and temporal trends of criminal behavior so that police officers can be deployed to hot spots for
crime prevention. With regard to using AI for risk assessment, individuals with a history of criminal behavior can be examined for the
likelihood of violent behavior and reoffending. Risk assessment provides decision support to a variety of tasks, including sentencing after
conviction, assignment to prison security levels, and parole board decisions. In recent years, risk assessment has shifted from traditional
regression analysis to machine learning, thereby improving forecasting accuracy. According to Berk 2021, random forests and gradient
boosting are two of the most effective ML methods for risk assessment. Further, Tortora, et al. 2020 reports that the combination of AI and
neuroimaging has led to AI neuroprediction, a technique that uses structural and functional brain parameters together with machine learning
methods for enhancing risk assessment. AI can also be used to analyze written texts and speeches to identify deception, sentiment, and
psychological state. The findings can be used to assess the credibility of witnesses, analyze confessions, and identify potential risk. For
instance, using text mining based on natural language processing, Ashraf, et al. 2022 develops a social media forensic model for identifying
people with abusive, aggressive, frustrated, and offensive behaviors. Cybercrime and cyber-security agencies can use this model to
shortlist profiles containing certain behaviors to prevent future crimes. Additionally, AI can be used to analyze criminal behavior and other
relevant information to identify complex patterns and relationships that may not be discovered by traditional statistics. For example,
Hofmann, et al. 2022 argues that traditional hypothesis testing does not accommodate psychiatric syndromes with their complex and often
highly interdependent multifactorial relationships. Using support vector machines, they found that factors related to psychopathology and
antisocial behavior are the most important predictors of aggressive behaviors among offender patients with schizophrenia spectrum
disorders. Although AI has the potential to revolutionize the field of forensic psychology, there have been criticisms of accuracy, fairness,
and potential biases in certain algorithms, such as Correctional Offender Management Profiling for Alternative Sanctions (COMPAS).
Consult Angwin, et al. 2016 for more information about COMPAS.
Angwin, Julia, Jeff Larson, Surya Mattu, and Lauren Kirchner. 2016. Machine bias. ProPublica, 23 May 2016.
COMPAS is an ML-based program to predict the likelihood of a criminal offender committing a future crime. The program has been used by
many US courts to inform decisions about pretrial detention, sentencing, and parole. This report used ample evidence to show that the
algorithm is biased against black offenders.
Ashraf, Noorulain, Danish Mahmood, Muath Obaidat, Ghufran Ahmed, and Adnan Akhunzada. 2022. Criminal behavior
identification using social media forensics. Electronics 11: Article 3162.
Based on the assumption that aggressive behavior leads to criminal activities, the researchers utilized AI to analyze negative comments,
abusive language, hate speech, and online bullying to detect potential criminal behavior. This is a proof-of-concept study.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 26/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
Berk, Richard. 2021. Artificial intelligence, Predictive policing, and risk assessment for law enforcement. Annual Review of
Criminology 4:209–237.
While exploring how AI can be applied to improve predictive policing and risk assessment, the author also discusses potential pitfalls of AI
for law enforcement, such as inaccuracy, bias, and lack of operational transparency.
Hofmann, Lena, Steffen Lau, and Johannes Kirchebner. 2022. Advantages of machine learning in forensic psychiatric research—
Uncovering the complexities of aggressive behavior in schizophrenia. Applied Science 12: Article 819.
The authors realized that linear regression methods might not be suitable to analyze complex psychiatric phenomena, and thus they turned
to ML methods. Multiple ML methods, including support vector machines, classification trees, random forests, KNN, Naive Bayes, and
gradient boosting, were employed, and the best model emerged from model comparison.
Tortora, Leda, Gerben Meynen, Johannes Bijlsma, Enrico Tronci, and Stefano Ferracuti. 2020. Neuroprediction and A.I. in forensic
psychiatry and criminal justice: A neurolaw perspective. Frontiers in Psychology 11.
These authors suggest combining AI and neo-imaging to improve the accuracy of risk assessment. While neuroprediction seems promising,
they maintain that additional research is needed for validating this approach. The paper is conceptual and accessible.
Artificial Psychology
In comparison with the well-established subfields of psychology already discussed, artificial psychology is relatively new. Artificial
psychology seeks to understand and model both human and machine mental processes in order to make AI systems behave more like
humans. To accomplish this goal, it is imperative for artificial psychology to be interdisciplinary, incorporating neuroscience, psychology,
physiology, philosophy, human engineering, linguistics, information science, computer science, and artificial intelligence. According to
Wang, et al. 2016, almost all psychological activities are influenced by emotion. Therefore, artificial psychology is concerned primarily with
the emotional stimulation of mental activities and the computerized implementation of such activities. Kansei engineering, which started in
1970 at Hiroshima University and flourished during the 1990s in Japan, is credited with giving rise to artificial psychology. The term “Kansei”
refers to responsiveness to feelings, emotions, and aesthetic experiences. The purpose of Kansei engineering is to bridge the gap between
human emotions and product design by including the sentimental element of user experience in product development. For more information
about Kansei engineering, consult Nagamachi 2016. Besides Kansei engineering, there are a few other historical roots of artificial
psychology. For example, Picard 1997 introduced the concept of affective computing, which focuses on developing technological systems
that can recognize, interpret, and respond to human emotions. It involves developing technologies to detect emotional cues from facial
expressions, vocal intonations, physiological signals, and other nonverbal behaviors (e.g., heart rate, skin conductance, etc.). Similarly,
Wang and Xie 1999 proposed a theory of artificial psychology that uses both AI and information science research methods to better
understand the human mind. Moreover, Minsky 2006 argued that emotions arise from the interactions of simpler, interconnected
subprocesses. Thus, understanding these components could lead to the development of machines that are capable of experiencing and
expressing emotions. More recent developments in artificial psychology are contributed by Crowder and Friess 2013; Crowder, et al. 2015;
and Crowder, et al. 2020. According to Crowder and colleagues, different reasoning methods such as inductive, deductive, abductive,
analogical, and comparative reasoning are needed for AI systems, whereas psychological impacts on both humans and the AI system need
to be considered for developing AI. To achieve this objective, they propose a modular artificial cognitive architecture with different types of
neural components for perception, attention, behavior, and expectation. This artificial cognitive architecture aims to mimic emotion’s
pervasive role in human cognition and create more adaptive AI systems. Artificial emotions are most useful in software agents where
sophisticated learning is a top priority.
Crowder, James, John Carbone, and Shelli Friess. 2020. Artificial psychology: Psychological modeling and testing of AI systems.
New York: Springer.
Crowder, Carbone, and Friess suggested that in order to make an AI system possess an “artificial consciousness,” it must model human
cognition, intuition, and other capabilities. Psychologists should also study how an AI system is received and perceived by humans, as well
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 27/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
as how the system will react to humans.
Crowder, James, and Shelli Friess. 2013. Artificial psychology: The psychology of AI. Journal on Systemics, Cybernetics and
Informatics 11.8: 64–68.
Based on the work of Dan Curtis, the authors argue that for an AI system to be considered intelligent, it must meet three conditions: (1) it
must make decisions independently using new, abstract, incomplete information; (2) it can reprogram itself based on new information; and
(3) it can meet the preceding conditions in novel situations.
Crowder, James, Shelli Friess, and John Carbone. July 2015. Artificial psychology revisited: Constructs for modeling artificial
emotions. In Proceedings of the 17th International Conference on Artificial Intelligence. Edited by Hamid R. Arabnia, 421–426. Las
Vegas, NV: CSREA Press.
The article presented modular neural architectures, weighted memories, and genetic learning as methods for modeling emotions in AI. It
aims to replicate the benefits of emotions in learning, motivation, and adaptation. The information presented is highly technical.
Minsky, Marvin. 2006. The emotion machine: Commonsense thinking, artificial intelligence, and the future of the human mind.
Cambridge, MA: MIT Press.
In this accessible book, Minsky argues that emotions arise from the interaction of simpler mental functions, or “agents.” These agents
interact and combine in many ways, resulting in the rich tapestry of human emotions. Emotion can be modeled as a computational process
by breaking it down into its fundamental elements.
Nagamachi, Mitsuo. 2016. Kansei/affective engineering. Boca Raton, FL: CRC Press.
This book explained how the set theory, AI-enabled expert systems, neural networks, statistical analysis, and psychological methods can be
applied to facilitate Kansei and affective engineering to support designers’ decisions and customers’ choices. Examples cited in the books
include automotive, home electrics, appliances, cosmetics, packaging, and e-commerce business.
Wang, Zhiliang, and Lun Xie. July 1999. Artificial psychology: An attainable scientific research on the human brain. In
Proceedings of the Second International Conference on Intelligent Processing and Manufacturing of Materials. Vol. 2. Edited by
John A. Meech, 1067–1072. New York: Institute of Electrical and Electronics Engineers.
Wang and Xie propose a new theory of AI-based artificial psychology, which uses information science methods to study human
psychological activities comprehensively. They expand the range of psychological characteristics to include low-level psychological activities
and high-level psychological processes. By reflecting objective reality, artificial psychology gained new meaning and broader content.
Wang, Zhiliang, Lun Xie, and Ting Lu. 2016. Research progress of artificial psychology and artificial emotion in China. CAAI
Transactions on Intelligence Technology 1:355e365.
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 28/29
11/2/23, 11:18 AM Artificial Intelligence, Machine Learning, and Psychology - Psychology - Oxford Bibliographies
This review covers the concepts, history, current state, and future directions of research on artificial psychology. Despite progress, there are
still many challenges left to overcome to develop more advanced emotion models and improve the interaction between humans and
machines.
back to top
https://www.oxfordbibliographies.com/display/document/obo-9780199828340/obo-9780199828340-0323.xml?rskey=OFJxbI&result=24 29/29