New 'Mind-Reading' AI Predicts What Humans Will Do Next

Via StudyFinds,

MUNICH — An artificial intelligence system can now predict your next move before you make it. We’re not just talking about whether you’ll click “buy now” on that Amazon cart, but rather how you’ll navigate complex decisions, learn new skills, or explore uncharted territory.

(Image by metamorworks via Shutterstock)

Researchers have developed an AI called Centaur that accurately predicts human behavior across virtually any psychological experiment. It even outperforms the specialized computer models scientists have been using for decades. Trained on data from more than 60,000 people making over 10 million decisions, Centaur captures the underlying patterns of how we think, learn, and make choices.

“The human mind is remarkably general,” the researchers write in their paper, published in Nature. “Not only do we routinely make mundane decisions, such as choosing a breakfast cereal or selecting an outfit, but we also tackle complex challenges, such as figuring out how to cure cancer or explore outer space.”

An AI that truly understands human cognition could revolutionize marketing, education, mental health treatment, and product design. But it also raises uncomfortable questions about privacy and manipulation when our digital footprints reveal more about us than ever before.

How Scientists Built a Digital Mind Reader AI

The research team started with an ambitious goal: create a single AI model that could predict human behavior in any psychological experiment. Their approach was surprisingly straightforward but required massive scale.

Scientists assembled a dataset called Psych-101 containing 160 experiments covering memory tests, learning games, risk-taking scenarios, and moral dilemmas. Each experiment was converted into plain English descriptions that an AI could understand.

Rather than building from scratch, researchers took Meta’s Llama 3.1 language model (the same type powering ChatGPT) and gave it specialized training on human behavior. They used a technique that allows them to modify only a tiny fraction of the AI’s programming while keeping most of it unchanged. The entire training process took only five days on a high-end computer processor.

Centaur could mark a new turning point in AI in its unprecedented ability to understand the human mind. (Image by Shutterstock AI Generator)

Centaur Dominates Traditional Cognitive Models

When tested, Centaur completely crushed the competition. In head-to-head comparisons with specialized cognitive models that scientists spent decades perfecting, Centaur won in almost every single experiment.

The real breakthrough came when researchers tested Centaur on completely new scenarios. The AI successfully predicted human behavior even when the experiment’s story changed (turning a space treasure hunt into a magic carpet adventure), when the structure was modified (adding a third option to a two-choice task), and when entirely new domains were introduced (logical reasoning tests that weren’t in its training data).

Centaur could also generate realistic human-like behavior when running simulations. In one test involving exploration strategies, the AI achieved performance comparable to actual human participants and showed the same type of uncertainty-guided decision-making that characterizes how people behave.

Neural Alignment: Centaur Mimics Human Brain Activity

In a surprising discovery, Centaur’s internal workings had become more aligned with human brain activity, even though it was never explicitly trained to match neural data. When researchers compared the AI’s internal states to brain scans of people performing the same tasks, they found stronger correlations than with the original, untrained model.

Learning to predict human behavior apparently forced the AI to develop internal representations that mirror how our brains actually process information. The AI essentially reverse-engineered aspects of human cognition just by studying our choices.

The team also demonstrated how Centaur could accelerate scientific discovery. They used the AI to analyze human behavior patterns, leading to the discovery of a new decision-making strategy that outperformed existing psychological theories.

“We’ve created a tool that allows us to predict human behavior in any situation described in natural language – like a virtual laboratory,” says lead author Marcel Binz in a statement.

What’s Next for Human Behavior AI?

While impressive, this research represents just the beginning. The current version focuses primarily on learning and decision-making, with limited coverage of areas like social psychology or cross-cultural differences. The dataset also skews toward Western, educated populations, a common limitation in psychological research.

The team plans to expand their dataset to include more diverse domains and populations, envisioning a comprehensive model that could serve as a unified theory of human cognition. They’ve made both their dataset and AI model publicly available for other researchers to build upon.

“We combine AI research with psychological theory – and with a clear ethical commitment,” adds Binz. “In a public research environment, we have the freedom to pursue fundamental cognitive questions that are often not the focus in industry.”

For the first time, we have an artificial system that can predict human behavior across the full spectrum of psychological research with unprecedented accuracy. Whether that development excites or concerns you may depend on how confidently we can ensure such tools are used responsibly.

Paper Summary

Methodology

The researchers created Centaur by fine-tuning Meta’s Llama 3.1 70B language model on a dataset called Psych-101, which contains trial-by-trial behavioral data from 160 psychological experiments involving over 60,000 participants making more than 10 million choices. They converted all experiments into natural language format and used a parameter-efficient training technique called QLoRA that modified only 0.15% of the model’s parameters. The training focused specifically on predicting human responses while masking out other parts of the experimental instructions.

Results

Centaur outperformed existing domain-specific cognitive models in almost every experiment when predicting behavior of held-out participants. The AI also successfully generalized to modified cover stories, structural task changes, and entirely new domains like logical reasoning. In open-loop simulations, Centaur generated realistic human-like behavior patterns and achieved comparable performance to actual humans in exploration tasks. Additionally, the model’s internal representations became more aligned with human neural activity compared to the base model.

Limitations

The current dataset focuses primarily on learning and decision-making domains, with limited coverage of social psychology, cross-cultural studies, and individual differences. The participant pool skews toward Western, educated populations typical of psychological research. The natural language format also introduces selection bias against experiments that cannot be easily expressed in text, and the researchers note the need for eventual expansion to multimodal data formats.

Funding and Disclosures

Research was supported by the Max Planck Society, the Humboldt Foundation, the Volkswagen Foundation, and the NOMIS Foundation. One author has consulting relationships and ownership interests in several biotech companies. The researchers have made their dataset and model publicly available for scientific use.

Publication Information

“A foundation model to predict and capture human cognition” was published in Nature on July 2, 2025. The study was led by Marcel Binz at the Institute for Human-Centered AI, Helmholtz Center Munich, with collaborators from institutions including Princeton University, University of Tübingen, Max Planck Institute for Biological Cybernetics, and others.