NeurIPS Audit Finds 100 Hallucinated Citations, Exposes AI Risks in Academic Publishing

GPTZero audited all 4,841 papers accepted to NeurIPS (Dec 2025) and confirmed 100 hallucinated (fabricated) citations across 51 papers. While the number is small relative to the tens of thousands of references submitted, the finding highlights systemic strain on peer review caused by a surge in submissions and growing use of large language models (LLMs) for drafting and formatting. Fabricated references—plausible-looking but nonexistent—are difficult to spot without manual verification or automated cross-checking. NeurIPS said core research is not automatically invalidated, but fabricated citations erode scholarly rigor and trust. Experts frame the issue as a crisis of scale, not intentional fraud, and expect responses including stricter submission rules, mandatory disclosure of AI assistance, automated pre-screening tools, and new verification services. The case underscores broader risks of unverified LLM outputs in high-stakes contexts and will likely influence conference and journal policies through 2026.