AI That Thinks Like Us? New Human Cognitive Model Sparks Hope and Debate in Psychology

Can AI Think Like Humans? Insights from Recent Nature Research

Artificial intelligence (AI) that mimics human thinking could offer fresh insights into how people make decisions, especially under different psychological conditions like depression or anxiety. A recent paper published in the journal Nature explores this idea, suggesting AI might bring a new perspective to human health research.

The research, led by the Helmholtz Munich team, presents a cognitive foundation model named Centaur. This AI model predicts human decisions across various tasks—gambling, memory, problem-solving—with accuracy surpassing traditional psychological theories. The team believes Centaur opens avenues for deeper exploration of human cognition and improvements in psychological models, potentially aiding clinical studies involving mental health.

Simulated Experiments and Clinical Applications

Centaur enables researchers to run experiments in a virtual environment rather than relying solely on human participants. This is particularly helpful when recruiting specific groups like children or patients with mental health issues is challenging. As Marcel Binz, the study’s lead author, notes, this approach accelerates research progress where traditional methods may lag.

Experts like Giosuè Baggio from the Norwegian University of Science and Technology highlight the excitement around machine-assisted cognitive science. The data-driven discovery of general cognitive models, as demonstrated by Centaur, shows promise for evolving unified theories of human cognition.

Training AI on Human Decisions: The Making of Centaur

Psychology has long struggled to develop models that both explain thought processes and reliably predict behavior. Centaur bridges this gap by combining interpretability with strong predictive power. It identifies common decision strategies and adapts flexibly to new situations, even predicting reaction times with notable accuracy.

The model was trained on a dataset called Psych-101, which includes over 10 million decisions from 60,000+ participants across 160 psychological experiments. These experiments cover a broad range of behaviors—from risk-taking and reward learning to moral dilemmas.

Centaur was fine-tuned using the LoRA method on Llama, a large language model. After training on 90% of the data, it was tested on the remaining 10%, performing better than task-specific cognitive models. For instance, in a two-armed bandit decision task, Centaur’s outputs aligned more closely with human choices than specialized models.

Interestingly, Centaur generalized to new tasks outside its training data, such as introducing a third option in the bandit task. This suggests it could serve as a virtual lab for developing and testing psychological theories before applying them to human participants.

Blueprint for Future Cognitive Science

The researchers demonstrated how Centaur and Psych-101 could guide the creation of interpretable, predictable cognitive models. This method can serve as a general template for scientific discovery across various experimental settings. Centaur also shows potential in automated cognitive science applications, like computer simulation prototyping.

Eric Schulz, director of the Human-Centered AI Institute at Helmholtz Munich, emphasized the model's potential to reveal computational patterns linked to decision-making processes. Future work will focus on differentiating cognitive strategies between healthy individuals and those with mental health challenges, aiming to responsibly deepen human cognition research.

Criticism and Skepticism from the Scientific Community

Despite promising results, Centaur has faced criticism. Blake Richards, a computational neuroscientist at McGill University and Mila, expressed skepticism, noting the model doesn’t truly simulate human cognition and its outputs may not reliably match real human behavior.

Jeffrey Bowers from the University of Bristol described some of Centaur’s behaviors as "absurd." In particular, the model’s short-term memory far exceeds human capacity, recalling 256 digits compared to the human average of 7. It also demonstrated "superhuman" reaction speeds in tests, which challenges its ability to generalize meaningfully beyond training data.

Bowers and others argue that while Centaur produces human-like outputs, its internal mechanisms differ fundamentally from human thought. Federico Adolfi of the Max Planck Institute noted the model’s potential fragility and warned that 160 experiments, although large, represent only a tiny fraction of the complexity of human cognition.

Recognizing the Value Amid Limitations

Some researchers acknowledge the value in the Psych-101 dataset itself. Rachel Heaton from the University of Illinois points out that it provides a valuable benchmark for testing other cognitive models. Richards also sees worth in investigating Centaur’s inner workings further.

Katherine Storrs of the University of Auckland observes that despite some overreaching conclusions, the significant effort behind the dataset and model could yield meaningful scientific contributions over time.

Conclusion

Centaur represents a novel attempt to model human decision-making with AI trained on extensive psychological data. While it shows potential to support research and clinical applications, its limitations and differences from actual human cognition must be carefully considered. The ongoing debate underscores the need for responsible use and further validation of AI models in cognitive science.

For researchers interested in AI applications in cognitive science and psychology, exploring advanced training and model evaluation methods can be valuable. Resources such as Complete AI Training’s latest AI courses offer practical pathways to deepen expertise in this area.