AI That Simulates Human Behavior Could Enhance Understanding of Cognition

Researchers from Helmholtz Munich have introduced Centaur, a new artificial intelligence (AI) model capable of simulating human decision-making with remarkable precision. Trained on over ten million decisions drawn from psychological experiments, Centaur mirrors human behavior across a wide range of scenarios. This breakthrough could reshape psychological theories and offer new insights into human cognition.

Centaur: A step forward in understanding decision-making

Centaur represents a major advancement over traditional psychological models, which often face a trade-off between offering clear explanations of human thought processes or reliably predicting behavior. In contrast, the Centaur model developed by Dr. Marcel Binz and Dr. Eric Schulz at the Institute for Human-Centered AI combines both elements. Trained on the Psych-101 dataset—a collection of over ten million decisions from 160 behavioral experiments—Centaur has the capacity to predict decisions in familiar tasks and novel situations alike.

One of the key features of Centaur is its ability to adapt to new contexts and make accurate predictions in unseen scenarios. The model identifies common decision-making patterns and can even predict the reaction time of individuals with notable precision. The researchers emphasize that Centaur functions as a virtual laboratory capable of simulating human decision-making processes in a variety of contexts. Potential applications extend to analyzing classic psychological experiments, as well as investigating individual decision-making processes in clinical settings, such as depression or anxiety.

Applications and potential impact

The researchers envision Centaur playing a significant role in fields like health research, particularly in understanding decision-making among people with different psychological conditions. The model’s creators are also working to enhance it by adding demographic and psychological data to the dataset, which could lead to more personalized predictions and a deeper understanding of human behavior.

Centaur’s ability to bridge two important aspects of psychological research—interpretability and predictive power—sets it apart from previous models. It provides insights into the limitations of classical models and points toward improvements. This opens up new avenues for research, with potential applications in diverse areas ranging from clinical psychology to environmental science.

Ethical considerations and next steps

While the promise of Centaur is significant, the researchers stress the importance of transparency and control in using AI systems like this. For instance, Centaur’s developers are ensuring that the model is open-source and can be hosted locally, preserving data sovereignty and ethical standards. This approach is integral to Helmholtz Munich’s commitment to advancing AI research in conjunction with psychological theory while remaining mindful of the ethical implications.

Looking forward, the team plans to examine the computational patterns within Centaur’s decision-making process. The goal is to better understand how individuals process information and how decision-making strategies differ between healthy individuals and those with mental health conditions. By doing so, the team hopes to offer valuable insights into human cognition and the factors influencing decision-making.

What is Psych-101?

Psych-101 is the dataset used to train Centaur, containing more than ten million individual decisions made by over 60,000 participants in 160 psychological experiments. These experiments cover a wide variety of human behaviors, including risk-taking, reward learning, and moral decision-making. The data was meticulously processed and standardized, ensuring it could be interpreted by a language model. Psych-101 is a unique resource for systematically modeling human behavior and offers unprecedented insights into how humans make decisions.

Reference: Binz M, Akata E, Bethge M, et al. A foundation model to predict and capture human cognition. Nat. 2025. doi:10.1038/s41586-025-09215-4

This content includes text that has been generated with the assistance of AI. Technology Networks' AI policy can be found here.