21% of peer reviews at ICLR 2026 were fully AI-generated and nobody caught it during the review cycle

At the International Conference on Learning Representations (ICLR) 2026, an analysis by Pangram Labs of all 75,800 peer reviews found that 15,899 reviews (21%) were fully AI-generated. More than half of all peer reviews showed signs of AI involvement. The discovery only happened because dozens of academics noticed red flags -- hallucinated citations, unusually verbose feedback, non-standard requests -- and raised alarms on social media. Carnegie Mellon professor Graham Neubig offered a bounty for analysis, and Pangram's team parsed every review within 12 hours. The conference's own systems caught none of it. The consequence is not just embarrassment for the AI research community. Each AI-generated review represents a decision point: a paper accepted or rejected based on feedback that no human actually read or evaluated. Researchers who spent months on a paper had their careers shaped by a language model's output. Junior researchers denied acceptance may lose conference presentation slots that are critical for job market visibility, tenure cases, and grant competitiveness. Senior researchers whose mediocre work was rubber-stamped by an AI reviewer gain unearned credibility. The peer review system's entire legitimacy rests on the assumption that expert humans are evaluating work -- when that assumption fails at 21%, the signal-to-noise ratio of the entire venue collapses. The structural cause is a volume crisis. ICLR submissions nearly tripled in three years: 7,304 papers in 2024, 11,672 in 2025, and 19,814 in 2026. Each reviewer was assigned multiple papers on tight deadlines with no compensation. AI-generated reviews are a rational individual response to an irrational collective workload. Conferences have no reliable detection mechanism, no penalties for AI-assisted reviews, and no way to reduce submission volume without gatekeeping that contradicts open science principles. The reviewer pool cannot scale with submissions because qualified reviewers are a finite resource, but submission volume is effectively unbounded.