AnalysisAI ModelsSeptember 13, 2026

Study: AI agents can't yet do open-ended AI research

Read original source →technologyreview.com

A Princeton-led study by Peter Kirgis and Sayash Kapoor used "shadow evaluation" — asking agents to answer research questions from unpublished NeurIPS papers — and found agents solved engineering tasks but lacked the judgment to produce conference-caliber original research. Anthropic's Claude Opus 4.8 was among the models tested.

People · Peter Kirgis, Sayash Kapoor

2 sources

More stories today

Open the live feed