NeurIPS 2026 Desk-Rejects 18% of Position Papers Over AI Generation, Using Pangram Detection
Key Takeaways
- ▸Nearly 1 in 5 Position Paper submissions were substantially AI-generated, revealing widespread adoption of LLMs for academic writing
- ▸NeurIPS deployed Pangram's AI detection technology at scale to enforce policy, successfully identifying and flagging violators
- ▸The policy permits AI only for copy-editing; all substantive writing must be human-authored, with authors required to disclose AI use
Summary
The NeurIPS 2026 Position Paper Track has implemented a hardline policy against AI-generated content, limiting AI use to copy-editing and peripheral changes only. To enforce compliance, organizers partnered with Pangram, an AI detection modeling company, to analyze all submissions for violation. The results were significant: 178 submissions (18.4% of all papers) were desk-rejected for being substantially AI-generated, while 123 submissions (12.7%) were flagged and required to provide evidence of substantial human authorship or face rejection.
Conference organizers argue that while thoughtful use of AI can improve productivity, AI-generated text poses an acute risk to the integrity of peer review. AI-generated content is often polished but can depart significantly from authors' original intent, shifting verification costs to reviewers and raising attribution questions. Pangram's detection models underwent multiple independent validation analyses to verify accuracy before the conference made enforcement decisions, with a strict data agreement ensuring zero retention of submission data.
- Conference is taking a conservative stance to preserve peer review integrity and establish norms for responsible AI use in academic publishing
