AI Agents Master Research Engineering, Fail at Open-Ended Science: Princeton

A Princeton-led study found that state-of-the-art AI agents can perform research engineering tasks but fail to produce credible scientific research that domain experts accept, despite extensive compute resources and time. The study introduced 'shadow evaluations' to test AI agents on genuinely unpublished, open-ended research questions, resulting in paper rejections by expert reviewers.

This Wire brief sits within Fusion42's coverage of AI & ML, and 1 source has reported it.