New ‘Mind-Reading’ AI Predicts What Humans Will Do Next, And It’s Shockingly Accurate
(Image of Metamorworks on Shutterstock)
In a word
- Scientists have created an AI called Centaur who can predict human behavior through any psychological experience with unprecedented precision
- AI has outperformed specialized models of several decades and successfully predicted behavior in completely new scenarios that he had never seen before
- The internal functioning of Centaur has become more aligned with the activity of the human brain simply by learning to predict our choices, potentially revolutionizing our understanding of cognition
Munich – An artificial intelligence system can now predict your next movement before doing so. We are not only talking about knowing if you click on “Buy now” on this Amazon cart, but rather how to sail in complex decisions, acquire new skills or explore an unexplored territory.
Researchers have developed an AI called Centaur who predicts human behavior precisely in practically all psychological experiences. It even surpasses the specialized IT models that scientists have been using for decades. Trained on data of more than 60,000 people making more than 10 million decisions, Centaur captures the underlying models of the way we think, learn and make choices.
“The human mind is remarkably general,” write researchers in their article, published in Nature. “Not only do we regularly make trivial decisions, such as the choice of a breakfast cereal or the selection of an outfit, but we also take on complex challenges, such as finding how cancer or exploring space.”
An AI that really understands human cognition could revolutionize marketing, education, treatment of mental health and product design. But that also raises uncomfortable questions about confidentiality and manipulation when our digital fingerprints reveal more about us than ever.
How scientists have built a digital reader of a AI
The research team started with an ambitious objective: to create a single AI model that could predict human behavior in any psychological experience. Their approach was surprisingly simple but required a massive scale.
Scientists have assembled a set of data called Psych-101 containing 160 experiences covering memory tests, learning games, risk-taking scenarios and moral dilemmas. Each experience has been converted into simple English descriptions that AI could understand.
Rather than building from scratch, the researchers took the Meta Llama 3.1 language model (the same type supplying the Chatppt) and gave it specialized training on human behavior. They used a technique that allows them to modify a small fraction of the AI programming while keeping most of it unchanged. The entire training process only took five days on a high -end computer processor.
Centaur dominates traditional cognitive models
When tested, Centaur completely crushed the competition. In head-to-head comparisons with specialized cognitive models that scientists have spent decades to perfect, Centaur has won in almost all experiences.
The real breakthrough occurred when the researchers tested a centaur on completely new scenarios. The AI has successfully predicted human behavior even when the history of experience has changed (transforming a space treasure hunt into a magic carpet adventure), when the structure has been modified (adding a third option to a two -choice task), and when entirely new domains have been introduced (logical reasoning tests that were not in its training data).
Centaur could also generate realistic human behavior when performing simulations. In a test involving exploration strategies, AI has achieved performance comparable to real human participants and has shown the same type of decision -making guided by the uncertainty that characterizes how people behave.
Neuronal alignment: Centaur imitates the activity of the human brain
In a surprising discovery, the internal functioning of Centaur had become more aligned with the activity of the human brain, even if it has never been explicitly formed to match neural data. When the researchers compared the internal states of AI to the brain analyzes of people performing the same tasks, they found stronger correlations than with the original and not formed model.
Learning to predict human behavior has apparently forced AI to develop internal representations that reflect how our brain really processes information. AI is essentially inverted aspects of human cognition simply by studying our choices.
The team also demonstrated how Centaur could accelerate scientific discovery. They used AI to analyze human behavior models, leading to the discovery of a new decision -making strategy that has surpassed existing psychological theories.
“We have created a tool that allows us to predict human behavior in any situation described in natural language – as a virtual laboratory,” said the main author Marcel Binz in a press release.
What is the next step for human behavior AI?
Although impressive, this research represents only the beginning. The current version is mainly focused on learning and decision -making, with limited coverage of areas such as social psychology or intercultural differences. The set of data also biaise towards Western and educated populations, a common limitation of psychological research.
The team plans to expand its data set to include more diverse domains and populations, imagining a complete model that could serve as a unified theory of human cognition. They made their data set and their AI model publicly accessible to the public so that other researchers can rely.
“We combine research on AI with psychological theory – and with a clear ethical commitment,” adds Binz. “In a public research environment, we have the freedom to pursue fundamental cognitive issues that are often not the emphasis in industry.”
For the first time, we have an artificial system that can predict human behavior through the complete spectrum of psychological research with unprecedented precision. Whether this development excites or concerns, you can depend on how we can make sure that these tools are used in responsibility.
Paper summary
Methodology
The researchers created Centaur by adjusting the Llama 3.1 70B language model of Meta on a set of data called PSYC-101, which contains behavioral test data by trial of 160 psychological experiences involving more than 60,000 participants making more than 10 million choices. They converted all the experiments into natural language format and used an economical training technique into parameters called QLORA which only modified 0.15% of the model parameters. The training focused specifically on the prediction of human responses while masking other parts of the experimental instructions.
Results
Centaur has surpassed the cognitive models specific to the existing field in almost all experiences when forecasting the behavior of the participants held. The AI has also managed to generalize with modified coverage stories, changes in structural tasks and entirely new areas such as logical reasoning. In open loop simulations, Centaur generated realistic human behavior models and has performed comparable to real humans in exploration tasks. In addition, the internal representations of the model have become more aligned with human neural activity compared to the basic model.
Boundaries
The current data set mainly focuses on the areas of learning and decision -making, with limited coverage of social psychology, intercultural studies and individual differences. The biaise participants basin towards Western and educated populations typical of psychological research. The format of natural language also introduces the selection bias in relation to the experiments which cannot be easily expressed in the text, and the researchers note the need for possible expansion in multimodal data formats.
Financing and disclosure
The research was supported by the Max Planck Society, the Humboldt Foundation, the Volkswagen Foundation and the Nomis Foundation. An author has consultation relationships and property interests in several biotechnology companies. The researchers made their data set and their model accessible to the public 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 Man Centered on Man, Helmholtz Center Munich, with collaborators of institutions such as Princeton University, University of Tübingen, Max Planck Institute for Biological Cybernetics and others.
