Researchers Report Peer Reviews Largely AI-Generated as Papers Rejected with Low Scores

On November 6, researcher nsaphra revealed that a paper she considers possibly her most practically impactful was rejected with low review scores of 2/2/3/2. Running the four reviews through the AI detection tool Pangram, three were flagged as 100% AI-generated; she later added that even the shortest review was 43% AI-generated. The same day, author TuhinChakr reported a similar experience: three of his ICLR reviews tested at 91%, 93%, and 91% AI-generated respectively, with one being almost a direct copy-paste from Claude, offering only the most superficial nitpicks. The incidents sparked widespread concern about the quality of academic peer review.

Confirmed

• nsaphra's paper was rejected with review scores of 2/2/3/2; three of four reviews were judged 100% AI-generated by Pangram, and the shortest one 43% AI-generated (m2, m4, m5)
• TuhinChakr's three ICLR reviews showed AI-generation ratios of 91%, 93%, and 91%, one being nearly a direct Claude copy (m3)
• nsaphra's view: a 43% AI share would have been acceptable if at least one review had been human-written; she lamented that academia likely can't afford to ban such reviewers from continuing to review (m4, m5)

Why it matters

• nsaphra pointed out that poor review quality is nothing new—"lottery-style" reviews caused by reviewers not understanding a paper or disliking it have always existed; what's new is reviewers facing mountains of AI slop submissions and simply giving up without even reading the manuscript, which she called a new low for peer review (m6)
• The two authors' independent tests corroborate each other, showing that AI-generated reviews are genuinely appearing in top-conference reviewing, and authors doing serious research may be rejected by perfunctory or even machine-generated feedback (m3, m4)

2026-11-06 ~ 2026-11-06 · 6 related posts

• [source] Researcher's paper rejected 2/2/3/2 — Pangram flags 3 of 4 reviews as 100% AI-generated — nsaphra · 2026-11-06
• Peer review hit a new low: reviewers now skip reading papers amid a mountain of slop — nsaphra · 2026-11-06
• [source] Author finds 3 ICLR reviews were 91-93% AI-generated per Pangram detector — TuhinChakr · 2026-11-06
• Reviewer scandal: Pangram flags 3 of 4 peer reviews as 100% AI-generated on a rejected paper — nsaphra · 2026-11-06
• [source] Follow-up: the shortest peer review in the AI-reviewer scandal was 43% AI-generated — nsaphra · 2026-11-06

1 near-duplicate retellings: nsaphra