Abstract
In the preface of Artificial Psychology: Learning from the Unexpected Capabilities of Large Language Models, Clayton Lewis poses a provocative question to both cognitive science and artificial intelligence: what if the stunning capabilities of Large Language Models (LLMs) force us to radically reconsider our own minds? The book argues that the core mechanism of statistical prediction in these systems is not merely a technical achievement but also a potential clue to human cognition. Drawing on his experience as a cognitive scientist, Lewis situates this claim against the backdrop of longstanding debates about representation, structure, and learning in psychology. He contends that the success of LLMs challenges foundational psychological models that rely on rigid, pre-structured knowledge such as rules, scripts, and symbolic logic. Instead, he proposes that human thought itself may be understood as a powerful, multi-modal predictive process, a notion he develops through the imaginative construct of the “Prediction Room” (PR). This framing positions the book within broader theoretical discussions about the nature of mind, while also raising questions about how far analogies between artificial and human systems can be sustained.
The book starts with a comparison between human and LLM trains of thought in problem solving and reasoning. In the face of Einstellung, analogical reasoning, inference, and prediction problems, the human brain and predictive models show different patterns. Referring to scientists and philosophers like Brandom, Lewis points out that LLMs tend to follow linear logic, while human “thinking” can revise both premises and predicted results simultaneously. However, as LLMs rely on a much broader field of knowledge, and monotonic logic can enact some patterns of human thought, the models can seem more intelligent in specific contexts. Lewis’ explanations of these mechanisms, as well as the problems evoked by them, offer a clear and accessible entry point into ongoing theoretical debates about rationality and flexibility in cognition. At the same time, the discussion occasionally remains descriptive, and further elaboration of the underlying theoretical commitments would strengthen the argument, particularly for readers less familiar with the philosophical background.
The second part focuses on “memory”; for predictive models like ChatGPT, that means storage and retrieval work. Different from what some of us understand of human memory—featuring a library of knowledge, while recalling is like taking a book from its shelf—this metaphor arguably resembles how LLMs’ memory systems operate. Mentioning psychological theories of proposition and association, the author suggests that they may be describing behavioral regularities rather than underlying mechanisms, while the prediction model is presented as closer to the structure of cognition. A novel point of view is thus put forward: knowledge may not be stored as discrete symbols but may emerge from distributed patterns of regularity within predictive systems. For human learning and memory, this chapter refers to research such as John Anderson’s (2005) ACT-R project and Cohen and Squire’s (1981) mirror-reading experiment for support. While these connections are suggestive, the extent to which predictive models can fully account for the richness of human memory remains open to debate, particularly given ongoing discussions about embodiment, affect, and contextual variability in psychological theory.
In the following sections, Lewis extends his framework to language and action. Instead of relying on the notion of innate Universal Grammar, he argues that language is better understood through general cognitive capacities such as intention reading, predictive modeling, and constraint-based decision making. These capacities are presented as shaping all human communication. The discussion also draws on Optimality Theory, which is introduced as a powerful, though imperfect, framework for capturing cross-linguistic regularities through ranked constraints, while acknowledging that its flexibility raises questions about what is genuinely universal (Smolensky et al., 2022). By comparing human cognition with emerging mechanisms in AI, the author shows how predictive systems—biological or artificial—combine evolved biases with learned experience. This integrative perspective is one of the book’s strengths, as it connects psychology, linguistics, and AI in a coherent framework. At the same time, some readers may find that the treatment of competing theories is relatively brief, and a more sustained engagement with alternative accounts would enhance the balance of the discussion.
By developing these comparisons, Lewis argues that both human cognition and predictive models rely on learning and prediction, and that LLMs can offer insight into underlying mental processes. This claim provides the basis for further exploration of behavior, including emotion and belief-desire psychology. According to Lewis, these domains can be understood as forms of contextual prediction and can be modeled within the PR framework. The book thus presents a unified account in which emotions, intuition, and belief-desire reasoning are interpreted as aspects of predictive systems. This synthesis is conceptually ambitious and contributes to ongoing theoretical discussions about unification in psychology. However, the proposal also raises important questions: whether predictive modeling alone can capture the normative, social, and phenomenological dimensions of these processes, and whether the PR construct risks oversimplifying complex psychological phenomena.
In the final section, Lewis underlines both the potential and the limits of predictive models. As cognitive systems, they can solve problems, recall information, use concepts, master language, and form intuitions, suggesting that prediction may serve as a unifying lens on mental life. At the same time, the book argues that prediction can function without internal symbolic structures, a position that resonates with Harold Garfinkel’s (1969) view. Characterized by simplicity and flexibility, predictive frameworks are used to explain phenomena such as the formation of extreme beliefs and the development of epistemic practices. Nevertheless, challenges remain, particularly regarding analogical reasoning and flexible updating. Lewis acknowledges that predictive modeling, like earlier theoretical approaches, may illuminate certain aspects of cognition while leaving others insufficiently explained. This reflexive stance adds nuance to the book’s overall argument, though some of these limitations could be explored in greater depth.
In conclusion, this book offers an innovative and thought-provoking reimagining of human cognition through the lens of predictive modeling. Its central message—that many aspects of mental life, from language and memory to intuition, emotion, and belief, can be illuminated by examining predictive processes—is presented with clarity and conceptual ambition. At the same time, the book’s arguments occasionally remain programmatic, and its critical engagement with alternative theoretical perspectives could be further developed. By situating LLMs within broader debates in theoretical psychology, Lewis opens new avenues for inquiry while also inviting careful evaluation of the assumptions underlying predictive approaches. The book is likely to be of interest to readers in psychology, linguistics, and AI, particularly those concerned with the conceptual foundations of cognition, and it contributes to ongoing discussions about how artificial systems can inform our understanding of the human mind.
