NeurIPS accepted research papers with 100 AI-hallucinated citations (fortune.com)

At the recent NeurIPS 2025 conference, a significant number of research papers—53 in total—were found to contain AI-hallucinated citations, as revealed by the Canadian startup GPTZero. They analyzed over 4,000 accepted papers and identified that citations were either fabricated or subtly altered versions of real references. This revelation is crucial for the AI/ML community as NeurIPS is a leading venue for cutting-edge research, where authors' credibility hinges on peer-reviewed citations. Such hallucinations raise serious concerns about the integrity of academic publishing, especially since even a single fabricated reference can undermine a paper's validity. The emergence of AI-generated citations that passed through multiple peer reviewers highlights the growing challenges in maintaining rigorous standards in academic research, especially with the increasing volume of submissions—21,575 for this year's conference alone. GPTZero's analysis, which included both automated checks and human verification of flagged citations, underscores the risks involved in AI-assisted research writing. These findings not only jeopardize academic reputations but also complicate reproducibility in a field where reliable citations are essential for verifying results. As AI continues to influence research practices, the need for robust detection methods and ongoing scrutiny of AI contribution in academic writing is more pressing than ever.