Fact-Check Corrects NeurIPS 2026 Rejection Criteria and Pangram False Positive Rates

A fact-check of the NeurIPS 2026 Position Track rejections reveals that the automated screening tool Pangram v3.3.2 produced false positives on human-written content polished by AI. The original narrative that desk-rejects were based on a simple "0.8 score + solo author" rule was incorrect. The actual criteria for the 79-paper tier required an AI detection score of 0.8 plus either multiple solo-authored submissions or an author with another desk reject. The 22-paper tier included authors who left the AI declaration blank, not only those who denied AI use.

Pangram Performance and False Positives

Pangram’s own internal report benchmarks v3.3.2, the specific version used by NeurIPS, showing it incorrectly classified 14.9% of human-written, AI-polished peer reviews as fully "AI" on an easy subset, and 4.5% on a harder subset. This performance was worse than both v3.0 and v4. The track's policy explicitly allowed AI polishing. An ICML 2026 paper notes that per-window false-positive rates should not be extrapolated to whole papers in either direction.

Clarifications on Rejection Criteria and Bias Claims

The fact-check corrects several misconceptions:

• Rejection Tiers: The 79-paper tier required an AI detection score of 0.8 plus multiple solo-authored submissions or another desk reject. The 22-paper tier included authors who left the AI declaration blank.
• ESL Bias Claims: The widely cited 61% false positive rate for non-native English speakers (ESL) comes from a 2023 study testing seven other detectors. Pangram’s v3.3.2 report claims 0 false positives out of 89 on the same TOEFL essays, though this is vendor data and not independently replicated.
• Independent Researchers: The reference to "independent researchers" referred to one person, Sergey Berezin, who explicitly makes no claim about how the chairs' papers were written.

Current Status

The total number of conditional papers cleared after the initial desk rejects remains undisclosed. The full analysis is available on the author’s blog.