AI Experts Caught Using Hallucinated Citations at NeurIPS Conference
A recent scan reveals that top AI researchers at the prestigious NeurIPS conference accidentally include fake, AI-generated citations in their papers. While the statistical impact remains low, the irony highlights the hidden flaws of relying on language models for academic writing.
AI detection startup GPTZero scans all 4,841 accepted papers from the prestigious NeurIPS conference and discovers 100 completely fabricated citations spread across 51 publications. Although this number represents a statistically insignificant fraction of the tens of thousands of references in the overall dataset, the presence of AI-generated "slop" at such a high-level machine learning event sparks a conversation about academic rigor.
The inclusion of fake citations poses a distinct problem for the research community because citations act as a professional currency that measures a scholar's influence. NeurIPS relies on a rigorous peer-review process where multiple reviewers explicitly check for hallucinations, but the sheer volume of submissions easily overwhelms these human gatekeepers. An inaccurate reference does not necessarily invalidate the core scientific findings of a paper, but it subtly degrades the reliability of the academic record.
This situation creates a highly ironic scenario where the world's foremost AI experts fail to fact-check the automated outputs of the very technology they build. Researchers easily know which actual papers influence their work, making it puzzling that they do not catch these LLM errors before submission. Ultimately, this revelation serves as a stark warning about the subtle ways unverified AI assistance sneaks into professional workflows and strains established verification pipelines.