Human-like AI might unlock the secrets of how we think

Researchers at Helmholtz Munich have created an AI model capable of accurately mimicking human behavior. Named Centaur, this language model was trained on more than ten million decisions drawn from psychological studies, enabling it to make choices in ways that closely mirror those of real people. This advancement paves the way for deeper insights into human cognition and the refinement of psychological theories.

For years, psychology has aimed to capture the complexity of human thinking. However, existing models typically excelled at either explaining mental processes or predicting behavior—rarely both. The team, led by Dr. Marcel Binz and Dr. Eric Schulz at the Institute for Human-Centered AI at Helmholtz Munich, has now developed a model that achieves both goals. Centaur was trained on Psych-101, a carefully assembled dataset containing over ten million decisions from 160 behavioral experiments.

Centaur stands out for its ability to predict human behavior not only in familiar scenarios but also in completely new ones it hasn't seen before. It recognizes common patterns in decision-making, adjusts to different contexts with ease, and can even estimate reaction times with impressive accuracy. “We’ve developed a tool that can forecast human behavior in any situation described using natural language – essentially functioning as a virtual lab,” explains Marcel Binz, the study’s lead author.

READ ALSO: Summit supercomputer draws molecular blueprint for repairing damaged DNA

READ ALSO: Lockheed Martin Delivers S-70i Black Hawk Helicopters to the Philippines

Its potential uses range from reanalyzing well-known psychological studies to modeling individual decision-making in clinical settings, such as for patients with depression or anxiety. Centaur offers promising opportunities in health research by shedding light on how people with various mental health conditions make choices. The dataset behind the model is also being expanded to incorporate demographic and psychological traits.

Centaur: Bridging Theory and Prediction

Centaur unites two areas that were traditionally separate: clear, interpretable theories and strong predictive accuracy. It highlights the limitations of classical models and offers ways to enhance them. This creates exciting opportunities for both scientific research and practical applications in fields like healthcare, environmental studies, and social sciences. “We’re only at the beginning, but the potential is already remarkable,” says Eric Schulz, the institute’s director. Marcel Binz emphasizes the importance of keeping such systems transparent and manageable—such as by using open-source, locally hosted models that ensure complete control over data.

READ ALSO: From Microbes to Complex Cells: Tracing the Origins of Life

READ ALSO: Microsoft unveils Majorana 1, the world’s first quantum processor powered by topological qubits

The next step for the researchers is to explore Centaur’s inner workings more closely: Which computational patterns align with specific decision-making behaviors? Can these patterns help us understand how people process information—or how decision-making differs between mentally healthy individuals and those with psychological disorders? The team believes that, if used responsibly, such models could significantly advance our understanding of human cognition.

It’s no accident that this research is being conducted at Helmholtz Munich instead of within major tech firms. “We integrate AI development with psychological theory and a strong ethical framework,” says Binz. “In a public research setting, we have the freedom to investigate core questions about cognition that are often overlooked in the private sector.”

What is Psych-101?

Psych-101 is a specially curated dataset created by Marcel Binz and his team to train the Centaur AI model. It includes over ten million individual decisions from more than 60,000 participants, collected across 160 psychological studies. These studies span a broad spectrum of human behaviors, such as risk-taking, reward learning, and moral decision-making. To make the data usable by a language model, the researchers carefully processed and standardized it. As a result, Psych-101 stands out as a valuable resource for systematically modeling human behavior using natural language inputs.