AI peer reviewers are fine with AI-fabricated papers

Peer review systems that use artificial intelligence recommend the acceptance of unsound papers created entirely by AI more than 80% of the time, a new study has found. The authors of the analysis, posted Oct. 20 on arXiv without peer review, created 600 fabricated manuscripts using only GPT-5 and had three other large language model (LLM) tools — o3, o4-mini, and GPT 4.1—review them. All four generative LLMs were built by OpenAI. Despite discovering integrity problems with some papers, the AI reviewers recommended accepting studies up to 82% of the time, the study found. “AI can be misused to attack this vulnerable system,” says study author Fengqing Jiang, a computer scientist at the University of Washington who hasn’t yet released the code underlying the project, BadScientist, fearing manipulation. While GPT-5 produced computer science papers for the study, it’s not difficult to tweak the model to produce chemistry papers, Jiang says. Jiang submitted the study last month to AI Agents for Science, the first conference where all the submitted studies were written and reviewed exclusively by AI. The event explored whether AI can independently create novel scientific hypotheses, methodologies, and results while preserving quality. It sought to spur conversations around what role AI should have in academic research and identify areas in which it can have the strongest impact. But Jiang and colleagues worry that science may get caught up in AI-only loops—AI-generated research reviewed solely by LLMs—that will result in questions about the soundness of the research. Safeguards against AI-only loops could include mandating human oversight, scoring papers