Helmholtz Munich's Centaur AI accurately predicts human decisions

For decades, cognitive scientists have hit the same frustrating wall: models that explain *why* humans think a certain way can rarely predict what they will actually do, while predictive machine learning models act as black boxes with zero psychological grounding.

Now, researchers at Helmholtz Munich may have bridged that divide.

Led by Marcel Binz and Eric Schulz, the team unveiled Centaur, an AI foundation model capable of predicting human decision-making with startling accuracy. The study was published in Nature on July 2.

A virtual lab powered by 10M decisions

Centaur gets its human-like intuition from a massive, newly created dataset called Psych-101.

The researchers compiled data from 160 classic and modern behavioral experiments, spanning over 60,000 human participants and more than 10 million individual choices. These trials cover everything from risk-taking and reward learning to complex moral dilemmas. Binz’s team manually processed and standardized the data so it could be fed into a natural language model.

The payoff is remarkable generalization. Centaur doesn’t just replicate scenarios it has seen before; it accurately forecasts human choices in completely novel environments described in plain English. Strikingly, it can even predict human reaction times with fine-grained detail.

Binz calls it a "virtual laboratory"—a system that can simulate human experiments on demand without needing to recruit fresh human cohorts for every hypothesis.

Peeking into the clinical mind

The team is now looking inside Centaur to map its internal computational patterns to human cognitive processes.

One of the most promising applications is in mental health research. By expanding Psych-101 to include detailed demographic and clinical data, the researchers hope to simulate and understand how decision strategies diverge in individuals suffering from depression or anxiety disorders.

Binz noted that carrying out this work at a public research institution rather than a Big Tech lab was intentional. It allowed them to pursue foundational cognitive questions with clear ethical boundaries, advocating for open-source, locally hosted architectures that protect data sovereignty.

Why it matters

Centaur effectively brings cognitive psychology into the foundation model era.

For researchers and behavioral scientists, this is a massive leap forward. Instead of spending months recruiting human participants to test early-stage hypotheses, they now have a digital sandbox capable of simulating real-world reactions, biases, and even hesitation times.

However, a model that accurately predicts human irrationality is a double-edged sword. While Helmholtz Munich is focused on clinical insights and fundamental science, the commercial and political incentives behind this capability are enormous. If an AI can reliably forecast how a human will react under moral stress or financial risk, using it to optimize product addiction, financial predation, or micro-targeted influence is dangerously straightforward. We think the real test will be whether open academic frameworks can set the safety baseline before commercial actors weaponize human predictability.