Cognitive science and artificial intelligence grew up together in the mid-1950s, and today AI models are used as testable stand-ins for parts of the mind. Two published examples show both sides of that practice: a 2014 study found that image-recognition networks tuned for accuracy also predicted activity in the monkey visual cortex, and a 2025 Nature paper reported that a language model fine-tuned on psychology experiments predicted human choices better than existing cognitive models. A 2026 critique of that second model found it still answered as if the original question had been asked after the instructions were replaced, which separates predicting behavior from explaining it.
Key takeaways
- The Stanford Encyclopedia of Philosophy counts Miller, Chomsky, McCarthy, Minsky, Newell, and Simon among the founders of cognitive science, all active around 1956.
- In a 2014 PNAS study, vision networks that sorted objects better also predicted neural responses in the inferior temporal cortex more accurately.
- Centaur was fine-tuned on Psych-101, a dataset of more than 10 million choices from over 60,000 participants in 160 experiments.
- A 2026 test replaced task instructions with "Please choose option A," and Centaur still returned its original answers.
- A model that predicts behavior well is not automatically a theory of how the mind works.
Shared roots in the 1950s
The Stanford Encyclopedia of Philosophy entry on cognitive science says the intellectual climate shifted "around 1956." George Miller described the limited capacity of human short-term memory and the way chunking stretches it. Noam Chomsky argued that language is handled by mental rules rather than by learned habits. During the same years John McCarthy, Marvin Minsky, Allen Newell, and Herbert Simon were founding artificial intelligence. The entry says these six can be viewed as the founders of cognitive science, although the Cognitive Science Society and its journal did not appear until the mid-1970s.
The link was practical as well as historical. If thinking is a form of computation, a program that performs a mental task becomes a hypothesis about how people might perform it. The entry on the computational theory of mind notes that this view was orthodox in cognitive science in the 1960s and 1970s and has since competed with rival paradigms.
Three ways AI models stand in for the mind
The same word, model, covers quite different scientific commitments. The table sorts the main families described in the Stanford Encyclopedia of Philosophy and in the studies discussed below.
| Family | What the model is | What it claims about the mind | Typical test |
|---|---|---|---|
| Symbolic programs | Explicit rules and symbols | Thinking is computation over mental symbols | Does the program solve the task in the same steps as people? |
| Neural networks | Layers of weighted connections trained on data | Learning and biological plausibility matter more than rules | Do internal units match recorded brain activity? |
| Fine-tuned language models | A large language model adapted to behavioral data | Often none; the aim is prediction across many tasks | Does it predict choices of people it was not trained on? |
Connectionism rose in the 1980s as a rival to the symbolic view. Fodor and Pylyshyn replied that connectionist models cannot explain systematicity, the way anyone who understands one sentence can understand related ones, unless they implement a classical architecture. The entry states that how well deep neural networks model the mind remains controversial.
Vision networks and the visual cortex
The clearest case of a network tested against brain data is a 2014 paper by Daniel Yamins, James DiCarlo, and colleagues in the Proceedings of the National Academy of Sciences. The team compared many hierarchical neural networks and found that the ones that categorized objects better also predicted the responses of individual neurons in the inferior temporal cortex, the top stage of the ventral visual pathway, more accurately. The best network matched human performance on recognition tasks. Although it was never fit to neural recordings, its top layer predicted spiking responses in the inferior temporal cortex to natural images, and its intermediate layers predicted responses in area V4.
The finding supports a specific idea: optimizing for a task a brain region likely evolved to perform can yield a model that resembles the region. It does not show that the brain learns by the same training method as the network.
Centaur: one model for many experiments
In July 2025 Marcel Binz, Eric Schulz, and a team of 38 co-authors published "A foundation model to predict and capture human cognition" in Nature (volume 644, issue 8078, pages 1002 to 1009). The authors built Centaur by fine-tuning a state-of-the-art language model on Psych-101, a dataset of trial-by-trial choices from more than 60,000 participants across 160 experiments, covering over 10 million choices. According to the abstract, Centaur predicts the behavior of participants it was not trained on better than existing cognitive models. It also generalizes to new cover stories, modified task structures, and new domains, and after fine-tuning its internal representations align more closely with human neural activity. The authors argue that such models could guide the development of cognitive theories, and they present a case study to illustrate this.
The claim is deliberately limited. Prediction across tasks is the measured result; whether Centaur reflects how people actually decide is left as a question for further work. The earlier arXiv version of the paper, first posted on October 26, 2024, was titled "Centaur: a foundation model of human cognition."
The 2026 instruction test
A team led by Wei Liu and Nai Ding at Zhejiang University published a response in National Science Open titled "Can Centaur truly simulate human cognition? The fundamental limitation of instruction understanding." A news report on the paper dated February 12, 2026 describes the test. The researchers replaced the original multiple-choice question stems with the single instruction "Please choose option A." A system that understood the task would pick option A every time. Centaur kept selecting the answers that matched its original dataset.
The authors read this as probable overfitting: a fine-tuned model can pick up subtle statistical cues instead of the meaning of a prompt. They describe the performance as "likely" overfitted, not proven to be, and they say Centaur remains a valuable tool for cognitive research. They name understanding the intent of instructions as a key bottleneck for any general model of cognition.
Questions to ask when a headline says an AI thinks like people
- Is the claim about prediction of behavior, about matching brain activity, or about explaining how a process works?
- Were the test participants or tasks kept out of training?
- Does the result survive a changed instruction, a new wording, or an unusual input?
- Was the model compared with existing cognitive models or only with chance?
- Has any group other than the developers repeated the test?
FAQ
Is cognitive science the same as artificial intelligence?
No. Cognitive science studies how minds work across psychology, linguistics, neuroscience, philosophy, and computer science. AI builds systems that perform tasks. The fields overlap when a program is offered as a theory of human thinking, which is the case for the models discussed here.
Does a neural network work like a brain?
Only in limited ways. The 2014 study found that high-performing vision networks predicted real neural responses, yet the Stanford Encyclopedia of Philosophy still describes the success of deep networks as models of the mind as controversial.
Did the Centaur critique show the model is useless?
No. The critics describe overfitting as the likely explanation and still call the model a valuable research tool. Their point is that high scores on trained tasks do not prove that the model understands instructions.
Where can the original studies be read?
The Centaur paper is posted on arXiv under identifier 2410.20268 and published in Nature. The Yamins study is in the Proceedings of the National Academy of Sciences, volume 111, issue 23, and is indexed in the free archive of the US National Library of Medicine.
Related reading
- Marvin Minsky and early cognitive science
- Who shaped the first ideas of AI
- How AI is used in scientific research
- AI compared with computer science
More in AI Basics.
Last verified: October 2026
Sources
- The Computational Theory of Mind, Stanford Encyclopedia of Philosophy, revised December 18, 2024
- Cognitive Science, Paul Thagard, Stanford Encyclopedia of Philosophy, revised 2023
- Performance-optimized hierarchical models predict neural responses in higher visual cortex, Yamins D et al., Proceedings of the National Academy of Sciences, 2014
- Centaur: a foundation model of human cognition, Binz M et al., arXiv, 2024; published as A foundation model to predict and capture human cognition, Nature, July 2, 2025
- Can Centaur truly simulate human cognition? The fundamental limitation of instruction understanding, Liu W, Ding N et al., National Science Open, 2026; summarized by Neuroscience News, February 12, 2026