A new critique of Centaur, the language model presented in 2025 as a system able to predict and simulate human behavior, cuts into one of AI's favorite borrowed costumes. The original Nature study reported that Centaur could anticipate decisions across psychological experiments with striking accuracy after training on over 10 million human choices from 160 studies and 60,000 people. A later paper in National Science Open argues that the performance may have come from overfitting. The model learned statistical shortcuts inside the training material instead of acquiring a working grasp of human decision-making.

The test was disarmingly plain. Researchers Nai Ding and Wei Liu modified multiple-choice prompts with an instruction asking Centaur to choose option A. A system following the task should obey that instruction across the altered questions. Centaur kept selecting the previously correct answers, as if the pattern in the exam outweighed the new command. That result does not settle the whole debate over machine reasoning, but it changes the burden of proof. Accuracy can no longer stand alone as evidence of thought. A model may reproduce the contour of human choice while missing the operation that gives a choice its meaning. The machine passes through the answer sheet and leaves the question almost untouched.