Does NeurIPS 2025's LLM disclosure policy match what readers actually need?

NeurIPS 2025 sorts LLM use by where it lands in the paper. Use that shapes the method has to be described: the policy asks authors to document their methodology and to describe LLM use "in the experimental setup section (or equivalent) if it is an important, original, or non-standard component of the approach." Use only for writing, editing or formatting, which "does not impact the core methodology, scientific rigorousness, or originality," needs no declaration; authors are asked instead to report it in a "statistical analysis survey." Responsibility stays with people whatever the tool. "Only humans are eligible to be authors," and each author is "fully responsible for all the content in your paper, including text, figures, and methodology, regardless of what tools (e.g., LLMs) you have used."

The reasons given are about risk rather than style. Some tools "may retain input data for further model training purposes," which the policy treats as a privacy consideration, and high-level instructions "could potentially result in hallucinations when generating plots, risking scientific integrity." Enforcement is retrospective: NeurIPS "reserves the right to revoke the paper's publication status" for violations, and its example is "using references generated by an LLM without conducting the due diligence to verify correctness, existence and appropriateness." Reviewers get a different and stricter rule. They may not share information about a submission "with anyone or any LLMs," but they may consult LLMs about concepts and phrasing if they take care not to leak submission content.

The ICML 2026 experiment in Does banning LLM use in peer review change review outcomes? tests the reviewer side of this question. Banning LLM use and allowing limited use both barely moved scores, decisions or confidence, and a substantial share of reviewers broke the rules either way. That is evidence on whether stated reviewer rules hold, which the NeurIPS page asserts but does not measure. On the author side, the method-versus-prose line meets the reader evidence in Do readers and writers differ on AI disclosure necessity?. There, readers judged disclosure more necessary than writers did, the judgment rose when AI text was directly adopted, and effort had no significant effect. That points the other way from a writing-only exemption, which covers direct adoption of AI-written text so long as the method is untouched. The policy's documentation demand also lines up with Does iterative prompt engineering undermine scientific validity?, which argues that prompt iteration itself breaks scientific method. NeurIPS asks only that LLM use be described when it is a non-standard part of the approach, a narrower claim.

The excerpt does not establish what happens in practice. It is a policy page. It reports no count of declared or undeclared LLM use, no compliance rate, and no test of whether authors apply the "important, original, or non-standard" test consistently; that judgment is left to the authors. The risks it names (retained inputs, hallucinated plots, unverified references) are stated as concerns, not measured. What follows is narrower than the policy itself: NeurIPS has drawn a clear division of responsibility and a rule for where disclosure belongs, but whether that rule changes what readers and reviewers see is open. A reader who cannot reliably tell LLM-written prose from human prose, as in Can readers tell LLM abstracts from human ones?, cannot see the writing-only use the exemption leaves undeclared. So the policy's line rests on the separate usage survey and on each author's own judgment.