Introduction

  • TL;DR: GPTZero reported 100 confirmed fake references across 51 accepted NeurIPS 2025 papers, and the incident spotlights how AI-generated “reference slop” can slip through elite peer review.
  • In the first place, hallucinated citations are not “minor typos”—they break the verifiability chain that science relies on.

Why it matters: Citations are the audit trail. If the trail is fabricated, readers can’t reproduce or validate claims.

What was found

GPTZero said it scanned the full set of accepted NeurIPS papers and confirmed 100 hallucinated citations across 51 papers. Some reports mention a slightly different paper count (e.g., “at least 53”), which typically reflects differences in counting criteria or update timing.

Why it matters: Even if the percentage is small, the failure mode is systemic—and it happened at the top of AI research.

NeurIPS policy: tools allowed, responsibility stays with authors

NeurIPS explicitly allows LLM usage, but states authors are responsible for the entire paper content, including references. The NeurIPS Board later emphasized (in a statement to The Register) that incorrect references do not automatically invalidate a paper’s content—while still acknowledging the issue needs deeper assessment.

Why it matters: “The model did it” is not a defense. Your workflow must include reference verification as a mandatory gate.

Why peer review misses it

Reviewers prioritize novelty, experiments, and proofs; reference lists are easy to de-prioritize under time pressure. The broader publishing ecosystem is already strained by increased manuscript volume and AI-assisted writing.

Why it matters: You can’t rely on reviewers to do full bibliographic audits. Verification must be automated before submission.

A practical verification pipeline (Crossref + Semantic Scholar + OpenAlex)

Use DOI resolution + metadata matching as the default. Fall back to title-based matching across at least two databases.

  • Crossref REST API for DOI/metadata checks
  • Semantic Scholar API for graph-based metadata checks
  • OpenAlex works endpoint for references/citation graph
  • DOI resolution via doi.org

Why it matters: Hallucinated references are “plausible strings.” Databases turn them into pass/fail signals.

Conference policy spectrum (real examples)

  • CVPR 2026: prohibits LLM/chatbot use in any part of reviewing (confidentiality + integrity reasons).
  • ICLR 2026: allows LLM assistance, but requires disclosure when usage is significant (or risk desk rejection).

Why it matters: Policies vary, but every venue converges on the same constraint: you must protect confidentiality and correctness.

Conclusion

  • Build a reference-lint gate (CI) before submission.
  • Treat any LLM-generated BibTeX as untrusted until verified against canonical databases.
  • Use at least two sources (Crossref + Semantic Scholar/OpenAlex) to avoid coverage false positives.

Summary

  • Hallucinated citations reached accepted NeurIPS papers (100+ fake references across 51 papers).
  • NeurIPS allows LLM tools but keeps authors accountable for references.
  • Automated verification using Crossref/Semantic Scholar/OpenAlex is the most reliable fix.

#hallucinatedcitations #neurips #peerreview #researchintegrity #crossref #semanticscholar #openalex #llmgovernance #academicwriting

References

  • (Irony alert: Hallucinated citations found in papers from NeurIPS, 2026-01-21)[https://techcrunch.com/2026/01/21/irony-alert-hallucinated-citations-found-in-papers-from-neurips-the-prestigious-ai-conference/]
  • (GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers, 2026-01)[https://gptzero.me/news/neurips/]
  • (LLM Policy - NeurIPS, 2025)[https://neurips.cc/Conferences/2025/LLM]
  • (AI conference’s papers contaminated by AI hallucinations, 2026-01-22)[https://www.theregister.com/2026/01/22/neurips_papers_contaiminated_ai_hallucinations/]
  • (NeurIPS research papers contained 100+ AI-hallucinated citations, 2026-01-21)[https://fortune.com/2026/01/21/neurips-ai-conferences-research-papers-hallucinations/]
  • (Science Is Drowning in AI Slop, 2026-01)[https://www.theatlantic.com/science/2026/01/ai-slop-science-publishing/685704/]
  • (Crossref REST API docs, 2020-04-08)[https://www.crossref.org/documentation/retrieve-metadata/rest-api/]
  • (Semantic Scholar Academic Graph API, 2026)[https://www.semanticscholar.org/product/api]
  • (Works - OpenAlex docs, 2025-12-13)[https://docs.openalex.org/api-entities/works]
  • (DOI Resolution Documentation, 2026)[https://www.doi.org/the-identifier/resources/factsheets/doi-resolution-documentation]
  • (ICLR 2026 Author Guide, 2026)[https://iclr.cc/Conferences/2026/AuthorGuide]
  • (CVPR 2026 Reviewer Guidelines, 2026)[https://cvpr.thecvf.com/Conferences/2026/ReviewerGuidelines]