When researchers unveiled an AI system called Centaur earlier this year, the pitch was bold: a single model that could predict how humans would behave in virtually any psychological experiment, just from a plain-language description of the scenario. Built on Meta’s Llama 3.1 70B large language model and fine-tuned on data from 101 classic psychology studies, Centaur was framed not as a narrow prediction tool but as a “domain-general computational model of human cognition.” A press release distributed through EurekAlert went further, describing it as able to “predict human behavior in any situation described in natural language.”
Now, a pointed critique published in May 2026 argues that Centaur’s impressive benchmark scores mask a fundamental problem: the system mimics the surface of human behavior through statistical pattern-matching without capturing the cognitive processes underneath. The challenge, titled “Not Yet AlphaFold for the Mind,” tested Centaur not just on whether it could predict average outcomes but on whether its generated responses matched the full texture of real human data. The authors found systematic gaps.
What Centaur was designed to do
The original Centaur work, published in Nature, describes a system that takes Meta’s 70-billion-parameter Llama 3.1 architecture and fine-tunes it on a purpose-built collection called Psych-101. That dataset draws behavioral data from 101 experiments spanning decision-making, memory, perception, and moral judgment. A companion arXiv preprint provided additional technical detail and reported that Centaur matched or exceeded human-level prediction accuracy on many held-out tasks that were not part of its training data.
The ambition went beyond forecasting. The Centaur team positioned the model as a “participant simulator” that could eventually stand in for real human subjects in future research, potentially reducing the cost and time of running large-scale psychology experiments. That framing set a high bar and invited scrutiny from researchers who study cognition for a living.
Where the critique says it breaks down
The “Not Yet AlphaFold for the Mind” paper, posted as a preprint and not yet peer-reviewed, tested Centaur’s outputs against the full distribution of human responses rather than just summary statistics. The title itself is a deliberate provocation: AlphaFold, DeepMind’s protein-folding model, succeeded because it predicted molecular structures with atomic-level precision, earning recognition as a genuine scientific breakthrough. The critique’s authors argue that Centaur has not cleared a comparable bar for cognition.
Their findings point to specific failure modes. In some experiments, Centaur produced overly consistent answers where real human participants showed wide variability. In others, it failed to reproduce minority response patterns, the kinds of outlier behaviors that are often theoretically important in psychology. A model that nails the group average but flattens the range of individual responses may look accurate on a leaderboard while missing what makes human cognition interesting and unpredictable.
The critique’s central argument is blunt: high predictive accuracy on held-out test data does not guarantee that a model produces genuinely humanlike behavior when it generates new responses from scratch. A system can learn regularities in the mapping from stimuli to average outcomes without capturing the underlying cognitive mechanisms or the ways individuals depart from the mean. From this perspective, Centaur functions more like a sophisticated curve-fitting tool than a window into human thought.
Commentary from the University of Bristol’s psychology department reinforced that distinction. Researchers there noted that Centaur can produce outputs resembling human responses, but its internal processes may differ fundamentally from the cognitive mechanisms people actually use. A credible cognitive model, they argued, must not only match observable behavior but also align with established theories about representation, memory, and inference.
Open questions neither side has settled
As of June 2026, the Centaur team has not issued a formal response to the critique’s specific findings about generative divergence. The original authors’ claims rest on evaluation metrics reported in the Nature paper and the broader preprint, but neither document addresses the particular failure modes the “Not Yet AlphaFold” paper identifies. Without a direct rebuttal, it remains unclear whether the Centaur team views these divergences as minor edge cases, artifacts of how the critique was conducted, or evidence of a deeper limitation.
The Psych-101 dataset itself raises questions that independent researchers have not yet fully explored. If the 101 experiments over-represent certain paradigms, such as the forced-choice tasks common in psychology labs, the model could learn shortcuts that perform well on similar benchmarks but fail in open-ended or real-world settings. This possibility is consistent with the pattern-matching interpretation, though it has not been tested in a systematic, published study.
Individual differences present another challenge. Many psychological effects depend on factors like age, culture, or prior knowledge. The available descriptions of Centaur’s training do not indicate that it was explicitly conditioned on such variables, raising the possibility that it mostly learns a kind of “average participant” profile. The critique’s authors suggest this may explain why the model sometimes matches mean outcomes while missing important subgroups.
There is also a disciplinary fault line running through the debate. The Centaur preprint lists authors from computationally oriented departments; the Bristol commentary comes from experimental psychologists with different methodological commitments. Model builders tend to prioritize predictive performance and scalability. Experimentalists focus on explanatory adequacy and alignment with existing theory. Whether these camps can converge on shared evaluation standards is a question no current publication resolves.
Finally, it is not clear how far the critique generalizes beyond Centaur. The “Not Yet AlphaFold” authors imply that any large language model trained primarily on outcome data, without constraints from cognitive theory, may fall into similar pattern-matching traps. But whether alternative architectures or training approaches could close the gap between predictive accuracy and genuinely humanlike generative behavior remains untested.
What this means for researchers and readers
The strongest evidence on both sides comes from the research papers themselves, not from press releases or university news items. The Nature paper provides the primary technical account of Centaur’s architecture, training, and benchmark performance. The arXiv critique offers the most detailed counterargument, with specific experimental results showing where Centaur’s generated behavior departs from human data. Readers evaluating these competing claims should weight those two documents most heavily, paying close attention to how each defines success and failure.
The practical stakes extend well beyond academic turf. If Centaur or similar models gain adoption as synthetic research participants, replacing or supplementing real human subjects in psychology experiments, the difference between pattern-matching and genuine cognitive modeling becomes a question of scientific validity. A model that predicts average responses accurately but generates unrealistic individual-level behavior could lead researchers to false conclusions about how interventions, products, or policies affect real people.
For teams considering whether to use Centaur or related tools in their own work, the immediate takeaway is caution. Predictive accuracy on standard benchmarks is necessary but not sufficient. Any researcher planning to use an AI model as a synthetic participant should test it against the full distribution of human responses in their specific experimental context, not just against summary statistics. The divergences reported in the critique suggest that such checks may reveal important mismatches even when headline accuracy numbers look impressive.
A notable advance, not yet a revolution
Over time, the Centaur debate may help sharpen what cognitive scientists expect from AI-based models of the mind. One possible outcome is a shift toward hybrid evaluation regimes that combine aggregate prediction metrics with detailed tests of generative behavior, individual differences, and theoretical alignment. Another is a more modest framing of what current large language models can offer: powerful tools for exploring hypotheses and simulating certain behavioral patterns, but not yet full explanations of how humans think. The safest reading of the evidence, as it stands, is that Centaur marks a real step forward in behavioral prediction while falling short of the transformative breakthrough that early coverage implied.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
