A study conducted by researchers from Princeton, the UK AI Security Institute, Stanford, and the University of Toronto evaluated whether frontier AI agents could independently conduct original AI research. The agents received the research questions from two unpublished NeurIPS 2026 papers, with six days, thousands of dollars in API credits, GPU resources, and internet access to produce a conference-quality paper. Both AI-generated papers were rejected by the original authors, who concluded that the systems failed to produce original scientific contributions worthy of publication at a top machine learning conference. The study identified five recurring failure modes and suggests that current AI agents can automate many engineering tasks but continue to struggle to generate original scientific work.
Source: Read the original article
