NeurIPS desk-rejected 178 papers as AI-generated — while its chairs' own papers flagged at 24-69%
tughanbulut · reddit · 2026-09-08
The NeurIPS Position Paper Track desk-rejected 18.4% (178) of submissions using the proprietary Pangram AI detector, with no human review or appeal. A deep dive into conference statements and Pangram docs reveals:
- The chairs would fail their own test: recent papers by the three track chairs flagged at 24%–69% by the same detector
- 42.7% initially flagged: Pangram's defaults marked nearly half the track as 90-100% AI; shrinking text windows was needed to bring it down to 12.7%
- The circularity trap: 22 papers were rejected solely for scoring >0.5 while checking "no AI use" — the black-box score treated as proof of lying
- The ESL penalty: a Stanford study found 61.22% of human-written TOEFL essays get flagged; NeurIPS published zero demographic calibration data
Rejected authors face no blacklist or misconduct mark and can resubmit to ICLR (Sept 25 deadline) or ICML. The poster runs strictcite.com, a deterministic zero-AI citation checker, and published a full field note with thresholds and recourse options.
More from Safety
- Deepfakes are unstoppable: local models and mass generation outpace any law, warns AI insider — LinusEkenstam · 2026-09-08
- Compromised LG Smart TVs can eavesdrop while appearing off, upload recordings later — TechNadu · 2026-09-08
- Zvi: OpenAI's Astra Marks a Rapid Decline in Chain-of-Thought Monitorability — Don't Worry About the Vase (Zvi) · 2026-09-08
- Lightweight prompt injection detector: MiniLM + logistic regression, F1 just 61.6% on adversarial benchmark — Worldly_Yoghurt8850 · 2026-09-08
- Fields Medalist Tsimerman founds Mathematical AI Safety Institute, hiring 10-100 mathematicians — AndrewCritchPhD · 2026-09-08
- US-China Biosecurity Cooperation Is Possible: Screen DNA Synthesis Orders — ChinaTalk · 2026-09-08