Princeton: AI Agents Fail At Real Science — Both NeurIPS Submissions Rejected

Peter Kirgis and Sayash Kapoor from Princeton University provided Claude Opus 4.8 with problems taken from unreleased NeurIPS 2026 papers. They had access to a $3,000 API allowance and six days to complete the tasks. While the AI did a good job solving engineering problems, it often ended up getting stuck on incorrect hypotheses and couldn’t move away from a dead end. It submitted two papers that were subsequently rejected because they were poorly written and lacked novelty.

We’ve all been waiting for the moment when AI starts self-improving recursively, i.e., without human intervention. But this experiment demonstrated how important the difference between an AI’s level of accuracy and its ability to make discoveries still is. As Anthropic co-founder Jack Clark notes, this is a “bearish signal,” so don’t hold your breath for rapid progress.