NeurIPS desk-rejected 178 papers as AI-generated — then the detector flagged the chairs' own papers too
CSProfKGD · x · 2026-09-19
The NeurIPS Position Paper Track screened all 969 submissions through AI-detection tool Pangram in June 2026, desk-rejecting 178 papers (18.4%) with no appeal and demanding version histories from 123 more.
Key controversies:
- Desk-rejected author Sergey Berezin ran the track chairs' own recent papers through the same detector: scores of 24%, 36%, 45% and 69%
- The tool used was Pangram 3.3.2; the supposedly more accurate Pangram 4 shipped a month after decisions were made
- Earlier reporting: 42.7% of submissions initially scored 90-100% AI-generated and 273 hit 100%; narrowing the text window alone dropped the flag rate by 70%
With ICLR reportedly seeing 50k submissions, AI-era review scale and detector reliability are becoming industry-wide problems.
More from AGI Musings
- "Have you tried turning suffering off?" X user mocks EA's endless genre of naive objections — dioscuri · 2026-09-19
- Every AI agent needs a human sponsor liable for its actions, argues tech commentator — binarybits · 2026-09-19
- Sovereign AIs may be impossible to regulate, argue researchers — leaving 'nuisance' path open — binarybits · 2026-09-19
- EA community slammed for dropping life-saving interventions over optics concerns — EigenGender · 2026-09-19
- Social engineering, not hacking, is ASI's likeliest escape route from secure labs — ronbodkin · 2026-09-19
- Gary Marcus: AI bubble is worse than tulips, which didn't need circular financing — miniapeur · 2026-09-19