Major AI conference flooded with peer reviews written fully by AI (www.nature.com)

Concerns that large language models had been drafting peer reviews at the International Conference on Learning Representations (ICLR) prompted a mass audit: Pangram Labs scanned 19,490 submissions and 75,800 peer reviews for ICLR 2026 after researchers reported hallucinated citations, long vague feedback, and incorrect numerical claims. Using its own LLM-detection tool and scripts to extract text, Pangram flagged roughly 21% of reviews as fully AI-generated and found that over half showed signs of AI use. On the manuscript side, 199 submissions (~1%) appeared fully AI-generated, 61% were mostly human-written, and 9% had more than 50% AI-generated text. The team posted results and a preprint to document their methodology and examples. The episode is significant because it’s the first large-scale, quantified instance of LLM-written peer review targeting a flagship ML conference, exposing a trust gap in a process that influences acceptance decisions and reputations. Technically, the findings show current LLMs can produce plausible but error-prone critiques (hallucinated citations, incorrect metrics, odd requests), which can sway reviewers or program chairs. ICLR organizers plan automated screening for policy breaches; the community now faces urgent questions about detection reliability, reviewer guidelines, auditability of reviews, and how to adapt review workflows and incentive structures to prevent AI-enabled gaming or degradation of review quality.