Portuguese 
English 
Spanish 
4 min of reading 
0 comments 
European Researchers Train a Language Model to Simulate Human Behavior in Cognitive Tests, Revealing Surprising Similarities to Real Mind and Reigniting the Debate on the Limits of Artificial Intelligence in Understanding Cognition.
European researchers have managed to train an artificial intelligence capable of simulating human behavior in psychological experiments.
Named “Centaur,” the system is a customized version of a large language model similar to ChatGPT, capable of responding like humans in cognitive tests.
The study was published in March 2025 in the prestigious journal Nature, raising debates among experts about the limits of artificial cognition.
This new AI represents more than a technological advance.
-
An 18-year-old transforms a bedroom-built app into a millionaire machine generating US$1.4 million per month with artificial intelligence and a visionary strategy.
-
5-year-old boy creates complete game with artificial intelligence and reveals how children are mastering technology even before learning traditional programming
-
Powering ChatGPT and the world’s data centers in 2026 will require more than 1,000 TWh of electricity — equivalent to Japan’s entire consumption.
-
Google bought 115 MW and six banks injected US$ 421 million: why AI data centers are choosing geothermal energy instead of solar and wind
It proposes a tool to better understand how the human mind works, while also expanding the boundaries of what machines can simulate.
The project is led by Marcel Binz, a researcher at Helmholtz Zentrum München in Germany, and includes a multidisciplinary team of cognitive scientists and AI engineers.
Race for General Artificial Intelligence
Technology giants like OpenAI and Meta are in an ongoing competition to develop the so-called artificial general intelligence (AGI) — an AI that can act like a human in any context.
Despite the enthusiasm, there is no consensus definition of what AGI would be.
In simplified terms, the goal is to create a system that is indistinguishable from human cognition.
Today, AI models are already performing impressive tasks.
They play chess matches masterfully, solve complex chemistry problems, and even write poems.
However, these abilities are almost always restricted to specific contexts.
A system that plays chess perfectly, for example, does not know how to drive a car.
Meanwhile, a human participating in a chess tournament can easily drive to the competition venue.
In other words, machines lack generalization.
The Birth of Centaur
Intrigued by the potential of large language models (LLMs) like ChatGPT, Marcel Binz began investigating how far these tools could simulate human cognition.
His curiosity found a viable path in 2023 when Meta released LLaMA, an open-source model that researchers could freely customize.
That’s when the Centaur project was born, named after the mythological creature that combines a horse’s body with a human torso — a clear analogy to the fusion of computation and cognition.
The objective of Centaur was clear: to act like a human in psychological experiments.
To achieve this, the team compiled over 160 scientific studies, drawing data from more than 60,000 volunteers, totaling over 10 million responses.
This data included tasks such as memorizing word lists, piloting ships in search of fictional treasures, playing slot machines, and making risk-based decisions.
The model was trained to imitate the decisions made by humans in these tests.
AI Reproduces Human Patterns Accurately
One of the first proofs that Centaur had absorbed aspects of the human mind was its ability to generalize cognitive strategies, even in modified environments.
When a spaceship game was transformed into a flying carpet game, the AI employed the same strategies used by humans.
This indicated an adaptation capacity — an essential characteristic of cognition.
Furthermore, the Centaur also demonstrated logical reasoning similar to humans, accurately answering questions that volunteers typically answered correctly and getting the same questions considered difficult wrong.
This pattern shows that the AI not only copies answers, but appears to partially understand the challenges presented, according to researchers.
In more specific tests, such as “rock-paper-scissors” style games, Centaur was able to predict human players’ behavior but struggled to understand moves generated by statistical algorithms.
This detail is crucial, as it suggests that the model absorbed real nuances of human thought and not just numerical patterns.
Not All Scientists Are Convinced
Despite the enthusiasm around Centaur, not all of the scientific community agrees with the interpretations of the results.
Researcher Olivia Guest from Radboud University in the Netherlands criticized the absence of a formal theory of cognition behind the model’s training.
For her, predicting answers does not mean understanding how the mind works.
Another critic, scientist Gary Lupyan from Indiana University, reinforced this point by stating that cognitive science seeks theories that explain, not just predict behaviors.
Marcel Binz acknowledges these limitations.
He admits that Centaur does not yet represent a complete theory of the human mind.
However, he believes that the model can serve as a valuable reference for comparing different hypotheses about cognition and stimulate the development of new theories.
The Future of Centaur
The team behind Centaur is already planning the next steps.
They are expanding the dataset used for training and want to increase the number of experiments included in the system by five times.
The expectation is that, with more data, the model will be able to imitate even more aspects of the human mind, offering unprecedented tools for cognitive science.
“We hope that, with this dataset, Centaur will be able to do even more things,” said Binz.
The idea is not just to improve the AI, but also to deepen the understanding of the human mind itself, something that remains one of science’s greatest mysteries to this day.
Impressive Advance, but with Reservations
Centaur represents an unprecedented advance at the intersection of artificial intelligence and psychology.
It still does not understand the human mind, but it can imitate it with impressive accuracy.
Even though there is still skepticism in the scientific community, the model establishes itself as a new tool for cognitive exploration.
And who knows?
Perhaps it is the first concrete step toward the long-dreamed general artificial intelligence — or, at the very least, a powerful lens for studying ourselves.
Be the first to react!