Can AI Discover Psychology Without Human Subjects?

Centaur, AUTOCOG, and the Possibility of In Silico Cognitive Science

Psychology faces a recurring problem when AI models are used in place of human participants. Even if an AI produces human-like answers, a theory derived from those responses may describe the simulator rather than human cognition. In “Sparks of In Silico Cognitive Science: Theories from Simulated Data Can Generalize to Humans,” Akshay K. Jagadish and colleagues test this problem directly. They let Centaur, a foundation model trained to predict human behavior, act as the experimental population. New theories were discovered using only Centaur’s responses and were then evaluated on actual human data. The result is striking: theories discovered from simulated participants predicted human decisions better than established benchmark theories.

An AI Designs Experiments, Judges Theories, and Rewrites Them

At the center of the study is AUTOCOG, the Automated Cognitive Scientist. It begins with two competing theories and runs a four-stage cycle. An LLM first designs experiments intended to distinguish one theory from another. Centaur then completes the trials as a substitute for human participants. The competing theories generate predictions, an LLM arbiter determines which theory better matches the simulated responses, and the weaker theory is revised or replaced through program synthesis.

The researchers ran this cycle 25 times. What matters here is that AUTOCOG is not merely selecting from a fixed menu of theories. It can alter the computational mechanism itself. The search therefore operates over possible explanations of how decisions are made, rather than merely tuning parameters inside an existing theory.

The experimental task involves choosing between two products described by up to five attributes with different validities. The initial theories were two familiar decision heuristics: Take-The-Best (TTB) and Tallying. TTB inspects cues in order of validity and stops once it encounters a cue that distinguishes the options. Tallying simply counts the number of attributes on which each option wins. One relies on a decisive cue; the other aggregates wins across cues.

The Discovered Theories Change Their Decision Strategy with Context

After 25 cycles, two theories survived: Relative Contextual Salience Lexicographic theory (RCSL) and Contextual Relative Advantage Normalization (CRAN). Both recombine mechanisms familiar from decision science, but neither was present among the initial theories.

RCSL produced the strongest generalization to human data. Like TTB, it examines more reliable cues first. But encountering a difference does not automatically terminate the search. Instead, the probability of stopping depends on how strongly that cue distinguishes the two options. A large difference can dominate the decision, while a small difference encourages the process to continue to other cues. If no cue becomes decisive, remaining evidence is integrated through a weighted compensatory calculation.

This produces an interesting picture of human decision-making. Rather than assuming that a person consistently uses either a one-cue heuristic or a weighted integration strategy, RCSL allows the decision process to move between these modes depending on the structure of the current choice.

CRAN introduces context in another way. It evaluates each attribute difference relative to the largest difference present in the current stimulus. When one cue contains a dominant advantage, smaller differences become less influential. When no extreme difference exists, several smaller advantages can accumulate. The scale against which evidence is judged therefore changes with the options currently on the table.

The Crucial Test: Theories Found in AI Data Survived Human Data

The central result appears in Figure 1D on page 2. The researchers tested the theories discovered through Centaur against ten held-out experiments with actual human participants.

RCSL achieved a mean squared error of 0.021, and CRAN 0.034. The benchmark Take-The-Best model scored 0.110, while Tallying scored 0.161. By comparison, theories discovered by running AUTOCOG directly with human participants reached about 0.018. Thus, RCSL, discovered without using human participants in the discovery loop, approached the performance of theories generated from human experimental data itself.

A Simulator May Not Need to Reproduce Humans Perfectly

The paper’s most consequential idea concerns what scientific simulations actually need to get right.

A simulator used for precise behavioral estimation would need to reproduce human responses closely. Theory discovery imposes a different requirement. AUTOCOG repeatedly asks which of two competing theories better explains the observed behavior. For this purpose, the simulator may only need to preserve the behavioral regularities that distinguish those theories.

Centaur therefore does not need to reproduce the entire human behavioral distribution with equal accuracy. It needs sufficient fidelity along the dimensions that determine which candidate theory survives the competition.

This distinction changes the possible role of synthetic participants in science.

The authors propose a hybrid workflow. Early stages of theory search could be conducted with simulated participants while candidate theories remain far apart and experiments stay near the simulator’s training distribution. Human participants could enter later, once competing theories begin producing similar predictions or experiments move beyond regions where the simulator can be trusted.

The resulting model of scientific work resembles a funnel. Simulation explores a wide landscape of possible mechanisms at low cost. Human experiments then test the smaller set of theories that survive. In this version of in silico cognitive science, AI does not remove humans from psychology. It changes where human evidence becomes most valuable.