In a major leap for artificial intelligence and decision making, researchers have developed a revolutionary AI system called Centaur AI, capable of accurately predicting human behavior across a wide…
In a major leap for artificial intelligence and decision making, researchers have developed a revolutionary AI system called Centaur AI, capable of accurately predicting human behavior across a wide range of scenarios. This cutting-edge decision-making AI outperforms many traditional psychological theories and models, offering powerful new insights at the intersection of AI and cognitive psychology.
Centaur AI Trained on Massive Psychology Dataset
To build this advanced decision AI, the team trained Centaur on data from 160 psychology experiments, covering over 10 million decisions made by 60,000 people in tasks such as gambling, memory tests, and problem-solving challenges. Unlike typical computational models — like DeepMind’s AlphaGo, which can only play Go — Centaur AI predicts human choices across multiple diverse contexts and even in completely new tasks it has never encountered before.
Unmatched Performance in Predicting Human Decisions
During testing, Centaur was evaluated on participants who were not part of the training data. In 31 out of 32 tasks, it outperformed Llama and 14 other cognitive or statistical models in predicting people’s choices, demonstrating the impressive potential of artificial intelligence cognitive psychology. The only exception was a task involving grammatical judgments.
Centaur also showed remarkable generalization skills, accurately predicting behavior in modified and entirely new tasks, including logical reasoning exercises. These results highlight the powerful role of artificial intelligence in decision making, showing that AI can uncover the underlying structure of human behavior.
Transforming Cognitive Research with Decision AI
Marcel Binz, a cognitive scientist at the Helmholtz Institute for Human-Centered AI in Munich, Germany, explained:
“Instead of recruiting human participants for every study, we can conduct experiments virtually with Centaur, which could greatly accelerate research, especially in studies involving children or psychiatric patients”.
Giosuè Baggio, a psycholinguist at the Norwegian University of Science and Technology, added:
“It’s exciting to see what new insights might emerge with help from this ai and decision making technology”.
Current Limitations and Future Development
Despite its impressive abilities, Centaur AI still has important limitations. For example, it focuses only on predicting what choice a person will make, but cannot estimate how long it will take them to decide — a key aspect of many studies in ai and cognitive psychology.
Another challenge involves the diversity of the training data: most behavioral data came from Western, educated, industrialized populations, which could affect Centaur’s accuracy across more diverse groups. To overcome this, the team is working to expand the dataset to four times its current size, aiming for better generalization.
The Future of AI in Decision Making and Beyond
With Centaur AI now freely available to researchers worldwide, its developers hope it will be tested and refined across different contexts, including applications like AI for investment decisions, where decision-making AI could help analyze and predict investor behavior with unprecedented precision.
“What we have now is just the starting point” says Binz “Centaur will only become more accurate and capable as we continue to develop it” pointing to a bright future for artificial intelligence and decision making.
This AI System Mimics Human Decision-Making After Training on 160 Psychology Experiments
A cutting-edge artificial intelligence (AI) system is now capable of forecasting how people will act across a wide range of scenarios — often surpassing traditional psychological theories of decision-making.
Named Centaur, the AI was refined using a massive dataset drawn from 160 separate psychology experiments, which recorded over 10 million decisions made by 60,000 individuals across tasks including gambling, memory challenges, and problem-solving.
Whereas most computational models and cognitive theories are confined to single tasks — such as Google DeepMind’s AlphaGo, which can only play Go, or prospect theory, which only predicts choices involving gains and losses — Centaur stands out by accurately simulating human behavior in multiple, diverse contexts. Remarkably, during testing, it could even anticipate decisions in tasks it had not seen before. The details of Centaur’s creation appear in a paper published today in Nature.
Researchers believe Centaur could become a powerful asset for cognitive science. “Instead of recruiting human participants for every study, we can run experiments virtually,” says co-author Marcel Binz, a cognitive scientist at the Helmholtz Institute for Human-Centered AI in Munich, Germany. He notes this approach could speed up research, especially for studies involving children or people with psychiatric conditions, where recruitment is challenging.
“It’s exciting to see what new insights might emerge with help from AI,” adds Giosuè Baggio, a psycholinguist at the Norwegian University of Science and Technology.
Overcoming Traditional Barriers
Historically, scientists have found it difficult to generalize human behavior across different tasks because most computational models were narrowly focused. Binz and his team aimed to break this limitation by fine-tuning Meta’s Llama large language model (LLM) using a vast behavioral dataset dubbed Psych 101. Over five days, they trained the model to predict not just average behaviors but also the typical range of decisions seen across populations.
To evaluate Centaur’s performance, they tested it on participants outside the training data. In 31 out of 32 tasks, Centaur surpassed both Llama and 14 other cognitive or statistical models at predicting participants’ choices. The only exception was a task involving grammatical judgments.
Moreover, Centaur demonstrated strong generalization skills by accurately predicting behavior in modified tasks and entirely new tasks it had never encountered, including logical reasoning exercises.
“This work highlights the underlying structure in human behavior,” says Russell Poldrack, a cognitive neuroscientist at Stanford University. “It raises the standards for what models in psychology should aim to achieve.”
Current Limitations and Next Steps
Despite its broad abilities, Centaur still has notable constraints. For instance, it focuses solely on language-based predictions: it can suggest what choice someone will make, but not how long they might take to decide, notes Poldrack.
Another limitation involves the dataset’s diversity — most of the behavioral data used for training comes from Western, educated, industrialized populations, which could affect the AI’s predictive accuracy across more varied demographics.
To improve Centaur, Binz and his team are expanding the dataset to quadruple its size, aiming for broader coverage and better generalization. As the model is freely accessible, they also invite researchers worldwide to validate its performance in different contexts. “What we have now is just the starting point,” says Binz. “Centaur will only become more accurate and capable as we continue to develop it.”
العودة إلى الأخبار
