GPT plus confiant sur les tâches difficiles où ils se trompe le plus, selon un preprint USC/Berkeley

In three linesGPT-4o, ChatGPT, and GPT-o3 display confidence exceeding their actual accuracy, with the gap widening on difficult tasks where they make the most mistakes. A USC/Berkeley preprint reveals growing divergence between stated confidence and real performance.

Summary generated by Claude — human-verified