AI agents given six days, $3,000 and GPU credits to do real research got rejected by the papers' own authors.

A Princeton-led study with Stanford, Berkeley, Johns Hopkins, Toronto, Georgetown and the UK AI Security Institute handed Claude Opus 4.8 and GPT-5.6 Sol Ultra the central research question from unpublished NeurIPS 2026 submissions. The original authors graded the output as reviewers and rejected both, citing weak experimental design, unsupported conclusions and impenetrable prose. Neither agent spent its full budget. (MIT Technology Review) Kirgis and Kapoor argue the failure was judgment, not engineering: the agents could run experiments and write LaTeX, they just couldn't decide which hypotheses deserved compute.