On the misuse of LLMs as models of mind: A case study of Centaur

In the field of NeuroAI, many strong claims are made about the similarity between artificial neural networks (ANNs) and the human mind. A recent study by (Binz et al., 2025) is a case in point. The authors transcribed a large database of psychological experiments and fine-tuned a large language model to predict human responses on a trial-by-trial basis. This model, Centaur, outperformed domain-specific state-of-the-art cognitive models. We highlight several major problems with Centaur. Through simulation, we demonstrate that the model is not subject to basic cognitive constraints and that its behaviour is best characterised as a form of role-playing. More generally, we discuss the inferential fallacy of mistaking predictive power for explanatory power and the importance of subjecting models to experimental manipulations to test specific hypotheses. These critical points are often overlooked in NeuroAI, but need to be addressed if the field is to progress.