Combining Psychology With Artificial Intelligence: What Could Possibly Go Wrong?

Abstract

The current AI hype cycle combined with psychology’s various crises make for a perfect storm. Psychology, on the one hand, has a history of weak theoretical foundations, a neglect for computational and formal skills, and a hyperempiricist privileging of experimental tasks and testing for effects. Artificial intelligence, on the other hand, has a history of conflating artifacts for theories of cognition, or even minds themselves, and its engineering offspring likes to move fast and break things. Many of our contemporaries now want to combine the worst of these two worlds. What could possibly go wrong? Quite a lot. Does this mean that psychology and artificial intelligence can best part ways? Not at all. There are very fruitful ways in which the two disciplines can interact and theoretically contribute to cognitive science, for instance, by studying the scope and limits of computational models of human cognition. But to reap the fruits, one needs to understand how to steer clear of potential traps.

Psychology has been living through various crises that have left the field grappling with its scientific status. Crises can lead to positive change, for instance, when they stimulate a reflective reimagining of theoretical foundations and epistemological practices. However, crises can also leave a field vulnerable to false prophets that promise illusory quick fixes. Right when psychology is vulnerable, society is going through an AI hype cycle. This AI hype cycle’s impact seems even worse than the ones that came before. With its devastating ecological harm and the exploitation of hidden labor, hyped AI serves to amplify discrimination and other social, economic, and environmental injustices. Psychological scientists predominantly prefer to ignore such real-world harms and instead ask: “How can AI benefit us?” ¹ This is understandable. ² After all, hyped AI promises to deliver candidate theories and statistical inferences through automated processes, also known as “machine learning.” This is music to the ears of psychology, which wants good theories and good statistical practices but whose scientific practitioners predominantly lack advanced theoretical, computational, and statistical skills. ³ Hyped AI even promises that it can replace human participants with “artificial minds” that are amenable to standard psychological experimental methods. This creates new opportunities for psychology to just keep doing what it has been doing—such as focusing on statistical inference and effects hunting (for a glossary of these and other terms used in this article, see Appendix Table A1; Cummins, 2000; van Rooij & Baggio, 2021)—but now in experiments with “artificial” participants in addition to real participants. Thus, no deep reimagining of our discipline’s foundations and epistemology may appear to be needed. However, these appearances and promises are all false. In this article we explain how and why. We end with guidance on how a fruitful interface between psychology and artificial intelligence is possible that does not fall in these traps and that can benefit cognitive science, the interdisciplinary study of cognition. This alternative approach disavows attempts to automate science—using machine learning or otherwise—and instead centers scientists’ own cognition.

Traps to Avoid

AI systems are not minds

If one reads the news and advertisements from the technology industry (sometimes disguised as scientific articles), one could be led to believe that we are on the verge of creating genuine artificial minds. For decades the domain generality of cognition was recognized as making cognition hard—and perhaps impossible—to explain, model, or replicate computationally (Fodor, 2000; Pylyshyn, 1987; Ryle & Tanney, 2009; van Rooij et al., 2019). But these days many people have come to believe that by training on massive amounts of human data it is possible to create AI systems that can think and act in a domain-general way, just like humans. The intuition seems to be that as long as one has enough human data to train one’s AI systems, those systems will asymptote to human-level behavior. Such envisioned human-level AI is also known as artificial general intelligence (AGI).

Van Rooij, Guest, et al. (2024) recently showed that training AI systems to scale up to human-level cognition is intractable (see Fig. 1 and Box 1). This has two implications. First, creating AGI through machine learning inherently consumes astronomical amounts of resources—sooner will the sun die out than that we will create AGI, and in the meantime we will just be polluting our planet and exploiting un(der)paid labor. Second, any AI systems that can be created in the short term are but decoys—these systems can trick us into thinking they are like human minds, but they are anything but (Guest & Martin, 2023). Even though these decoys can appear impressive and trick us, they are not hard to unmask through careful tests (e.g., Dentella et al., 2024) or even commonsense probes.

Fig. 1.

Illustration of why AI systems cannot realistically scale to human cognition within the foreseeable future. (b) Human cognitive capacities (such as reasoning, communication, problem solving, learning, concept formation, planning) can handle unbounded situations across many domains, ranging from simple to complex. (a) Engineers create AI systems using machine learning from human data. (d) In an attempt to approximate human cognition, AI systems consume a lot of data. (c) Making AI systems that approximate human cognition is intractable (van Rooij, Guest, et al., 2024), that is, the required resources (e.g., time, data) grow prohibitively fast as input domains become more complex, leading to diminishing returns. (a) Any existing AI system is created in limited time (hours, months, or years—not millennia or eons). Therefore, existing AI systems cannot realistically have the domain-general cognitive capacities that humans have. (Figure made with elements from freepik.com.)

Box 1.

Intractability and exponential growth.

Input-output mappings can differ in the amount of resources (e.g., time, space) needed to compute them. For instance, finding a number (output) in a list of length $n$ (input) takes on the order of $n$ steps because one needs only to inspect each position in the list at most once to find the number one is looking for. Sorting a list of numbers instead takes no more than $n\times n=n^2$ steps because one needs to go through the list at most $n$ times to find the next smallest number. Mappings like these that are computable in $n$ , $n^2$ , etc., steps are said to be tractable. In contrast, some input-output mappings are intractable. For instance, when for each small increase in $n$ the number of steps doubles, then the growth is exponential (i.e., $2\times 2\times \dots \times 2=2^n$ ). Such exponential growth may appear deceivingly slow at first but grows astronomically large quickly. People tend to be poor in intuiting intractable resource demands, leading them to underestimate resource demands needed for scaling up to larger $n$ . The water-lilies problem is a simple example: If a water lily doubles every 24 hours, after 60 days the lilies will already cover more than the surface of the earth (van Rooij, 2018). For implications of intractability, see the main text and Figure 1.

It is thus worrisome that some researchers seem to believe that AI systems can replace human participants in experimental research (e.g., Dillion et al., 2023; cf. Crockett & Messeri, 2026). Confusing AI systems for human minds is not only a category error (Ryle & Tanney, 2009) and dehumanizing (Bender, 2024; Erscoi et al., 2023) but also a recipe for shoddy science (Guest & Martin, 2023). After all, AI systems are decoys that cannot realistically approximate human cognition and behavior in any reliable domain-general way (van Rooij, Guest, et al., 2024). The methodological crisis in psychology has been bad enough (Flis, 2019). There is no benefit in making it worse by replacing the people whom we wish to study with decoys.

AI systems are not theories

So AI systems cannot be minds. But can these engineered systems be theories of how cognition works? It seems prima facie that training neural networks—or other cognitively inspired ⁴ computational architectures—on cognitive tasks and/or human data produces viable computational theories of how cognition works. After all, if such an AI system can mimic human behavior and predict ⁵ human performance on cognitive tasks, then human cognition must work more or less analogously to how the AI system works, right? Many people seem to believe that this implication holds. However, no matter how intuitively appealing, the belief is fundamentally mistaken for several reasons.

Prediction is not explanation

Being able to predict human behavior does not imply being able to explain the why or how of that behavior. That prediction and explanation are dissociable is easily illustrated by considering the tides (Cummins, 2000; see also Blokpoel & van Rooij, 2021–2025, Chapter 2): We could predict the tides long before we could explain them (i.e., in terms of the gravitational pull of the moon). Even today, we use tide tables, that is, large lookup tables that map dates and times of day to positive or negative levels of the tide. Using such tide tables for different locations on earth, one can quite precisely predict the tides at any time of the day. Yet no one would claim that tide tables explain the tides. Similarly, a huge lookup table that more or less predicts what behaviors people will display in different situations and conditions does not provide an explanation for why the behavior is as it is or of how cognition works (van Rooij, 2022).

Correlation does not imply cognition

Some may object that, although the prediction of outward behavior is insufficient for a model to be explanatory, surely when internal parameters of the model correlate with brain data that shows the model matches the mechanistic workings of brains/minds (Guest & Martin, 2023). Unfortunately, this inference is mistaken, too. As shown by Guest and Martin (2023), it is invalid to infer from correlations between a model and brain data that the model works like the brain. This follows from multiple realizability ⁶ : Just like both a digital clock and an analog clock can tell the time, and one will be able to correlate parts from one with the other, they operate in fundamentally different ways (see Guest & Martin, 2023, Fig. 2).

Capacities are not tasks

Even if multiple realizability cannot be ruled out, it may seem that AI systems that can (be made to) perform like humans on cognitive tasks provide at least possible theories of how cognition could work. However, this inference is not licensed either. Computational models of tasks, and task performance, are not yet theories of cognition (Guest & Martin, 2021; Morrison & Morgan, 1999). Theories of cognition should at a minimum provide possible explanations of one or more substantive human cognitive capacities, such as vision, decision-making, reasoning, concept formation, learning, or communication (Cummins, 2000; Egan, 2018; van Rooij & Baggio, 2021). “A capacity is a more or less reliable ability (or disposition or tendency) to transform some initial state (or ‘input’) into a resulting state (‘output’)” (van Rooij & Baggio, 2021, p. 684). A cognitive capacity is thus the ability to transform cognitive states (e.g., percepts, preferences, beliefs) into other cognitive states (e.g., beliefs, plans, decisions, actions). Although it is true that in (hyperempiricist) psychology cognitive capacities are typically studied by having people perform various tasks, computationally (or, more often, statistically) modeling task performance does not yield explanatory theories of the full-blown capacities. This is not only because tasks do not map one-to-one, or in any other straightforward way, to substantive cognitive capacities, but also because even if they could, the models would not be able to scale up to situations of real-world complexity.

AI systems cannot scale

For computational models of cognition to be able to scale up from toy scenarios (such as studied in the psychological labs or such as form the bases of training AI systems through machine learning), these models should minimally be computationally tractable (van Rooij, 2008). That is, the input-output mappings hypothesized by the computational models should in principle be computable using a realistic amount of resources (Box 1). However, when existing computational models are mathematically analyzed for their resource demands it turns out that such models are either tractable but limited in their input domains, or they are domain-general but then the models are intractable (for a textbook introduction to this methodology, see van Rooij et al., 2019). Consequently, no computationally tractable account exists to date for substantive and domain-general cognitive capacities, such as reasoning, communication, decision-making, planning, analogizing, categorization, and concept formation (van Rooij et al., 2019). Moreover, there is no good reason to believe that such tractable accounts will be forthcoming via machine learning (van Rooij, Guest, et al., 2024) or otherwise (Rich et al., 2021). Hence, if someone claims their concrete AI system is a “theory” of human cognition, they are more likely than not overstating the scope and capacities of the system (van Rooij et al., 2019) and obfuscating the system’s limits, including the human in the loop needed to make such systems “work” (Guest & Martin, 2025).

Cognitive science cannot be automated

Psychology has undergone some important cultural changes because of the so-called replication crisis. ⁷ Instead of improving psychology’s theoretical foundations and epistemological practices by adopting conceptual, computational, and formal tools from computational cognitive science (Guest, 2024; Guest & Martin, 2021; Navarro, 2019, 2021; van Rooij & Baggio, 2020, 2021; van Rooij, Devezer, et al., 2024), the overwhelming response in the mainstream has been to push for statistical methodological reform that centers a rigid proceduralization of empirical research (for critiques, see also Devezer et al., 2021; Szollosi et al., 2020). This move has further cemented hyperempiricism into psychology and has left the field vulnerable to a view that science can be proceduralized and maybe, to a large extent, even automated.

Right at the time that psychology is vulnerable, hyped AI enters the scene and makes false and misleading promises that both theory generation and scientific inference can be automated using machine learning. For psychology such promises are very attractive, especially because its practitioners often lack theory-building and advanced statistics skills. However, a common expression applies here: “If something seems too good to be true, then it probably is.” Cognitive science cannot be automated because theory generation is provably intractable (Rich et al., 2021; van Rooij, Guest, et al., 2024), and scientific inference cannot be reduced to statistical inference (Guest & Martin, 2023; Navarro, 2019).

Hyped-AI promises are not harmless (Guest, 2024). Although automation may give the false impression of rigor and efficiency, it leads to conceptual and scientific deskilling, deteriorates theorizing, and can make us blind to important scientific paths we would need to go down (Rich et al., 2021). In addition, because building theories of substantive (human-level) cognitive capacities is computationally intractable, any efficient proceduralized way of generating theories can produce only decoys, leading to the other traps (van Rooij, Guest, et al., 2024). Last but not least, automated scientific inferences can cause deep scientific inconsistencies and theoretical confusion (Guest & Martin, 2023; Guest et al., 2025) and can give false credibility to harmful pseudoscientific ideas and practices (Birhane & Guest, 2021; Spanton & Guest, 2022).

A Possible Path Forward

In this article we have focused on what all can go wrong when combining psychology with AI in thoughtless ways. We realize the reader may appreciate guidance for traveling the winding, branching, and open-ended road that is cognitive science without falling into said traps. Thus, we provide a summary of the nature of each of the traps in Table 1, as well as what can be done to avoid them and the problems that arise when one does not. Step 1 in avoiding the traps is to be aware of them and to be able to recognize them “in the wild” (e.g., in the literature or in scientific practice). Step 2 is to develop a research approach that removes the root causes of the traps and prevents them from arising in the first place. We (and others) have proposed that such a cognitive science is possible even within a computationalist framework if we reconceptualize artificial intelligence (or computer science more broadly) as “a provider of computational tools (frameworks, concepts, formalisms, models, proofs, simulations, etc.) that support theory building in cognitive science” (van Rooij, Guest, et al., 2024, p. 616), but without confusing the theoretical possibility of explaining human cognition computationally with the practical feasibility of (re)making human cognition in factual computational systems (i.e., makeism; see van Rooij, Guest, et al., 2024, Box 2). We coined this alternative computationalist approach “nonmakeism.”

Table 1.

Typology of Traps, What Goes Wrong if They Are Not Avoided, and How They Can Be Avoided

Trap	Problems if not avoided	How to avoid
Believing AI systems are minds	Dehumanization, shoddy science	Realize AI systems are decoys
Believing AI systems are theories	Theoretical deterioration, pseudoexplanations, fallacious metatheoretical calculi, overclaiming, obfuscation of humans in the loop	Realize AI systems are decoys, prediction is not explanation, correlation does not imply cognition, tasks are not capacities, and AI systems cannot scale
Believing cognitive science can be automated	Theoretical cul-de-sacs, pseudoscience, deskilling, proliferation of decoys, creation of the other traps	Think deeply, slow science, take computationalism seriously, and be “nonmakeist”

Note: These traps in a sense constitute category errors (Ryle & Tanney, 2009), and the success-to-truth inference (Guest & Martin, 2023) is an important driver in most, if not all, of them.

Nonmakeist AI takes computationalism more seriously than makeist AI (Guest et al., 2025) because it bites all the bullets implied by computationalist axioms, such as multiple realizability of computation and fundamental limits imposed by computational intractability. “Cognitive science is itself a cognitive activity” (Rich et al., 2021, p. 3034). It thus follows from computationalism that cognitive scientists’ explanations, inferences, and theory building are all limited by computational constraints as well. Lacking any efficient reliable procedure for generating explanatory theories, all we can do is postulate (often blatantly wrong) theories and rigorously analyze their explanatory scope and limits. In other words, nonmakeist AI distinguishes itself from the currently hyped AI by disavowing the automation of science and instead centering the role of scientists’ own cognition in the thoughtful analyses of computational models of cognition. By using whatever insights we may draw from such analyses we advance our scientific understanding (or our lack of understanding) one small step at a time. Good science is slow (Stengers, 2018), and if cognitive science wants to take AI as theoretical psychology back on board, then it needs to take computationalism seriously.

Conclusion

Psychology and artificial intelligence (or, more broadly, computer science) are two of the six traditional disciplines constituting cognitive science (the other four being philosophy, linguistics, neuroscience, and anthropology; see Fig. 2). Over the last 3 decades psychology came to dominate cognitive science with its hyperempiricist tendencies (Gentner, 2010, 2019), whereas artificial intelligence retracted from the field (Forbus, 2010), taking most computational theory-building tools with it. We are currently witnessing a rapprochement of the two disciplines. Although theoretical strengthening of cognitive science is welcomed, great caution is needed to prevent a new status quo that is worse than the old one. Computational concepts remain valuable for carefully crafting theories in cognitive science (Guest & Martin, 2021; van Rooij & Baggio, 2021), but they can only flourish if we (a) do not confuse AI systems for minds or theories, (b) do not confuse machine learning for the scientific method, and (c) understand that our computational models can track only the scope and limits of our understanding.

Fig. 2.

Visual depiction of the connections between the cognitive sciences. Solid lines denote stronger interdisciplinary ties, and dashed lines denote weaker ones. This figure is derived from the original put forth by the Sloan Foundation in 1978 and reproduced from Figure 4 in Pléh and Gurova (2013). Different versions of it over time have used “artificial intelligence” instead of “computer science,” and vice versa (cf. Miller, 2003).

Recommended Reading

Guest, O. (2024). (See References). Proposes criteria for evaluating theories, including the criterion of minimizing harm.

Guest, O., & Martin, A. E. (2021). (See References). Argues for the benefits of formal modeling for psychology.

Guest, O., & Martin, A. E. (2023). (See References). Explains the logic of why correlation is not cognition.

van Rooij, I., & Baggio, G. (2021). (See References). Counters hyperempiricism and proposes the theory cycle as a complement to the empirical cycle.

van Rooij, I., Guest, O., Adolfi, F., de Haan, R., Kolokolova, A., & Rich, P. (2024). (See References). Argues that machine learning cannot realistically recreate human-level cognition in AI systems and proposes nonmakeism as an alternative approach for computational cognitive science.