According to the latest report by THE DECODER, AI detection company GPTZero found that among 4,841 papers accepted at the 2025 NeurIPS (Conference on Neural Information Processing Systems), 51 papers contained at least 100 confirmed fake citations. These papers, despite undergoing rigorous peer review, included fabricated literature sources generated by AI.

QQ20260123-100129.png

The investigation revealed a phenomenon called "fake original citation" (Vibe Citing): the cited content is highly deceptive in format and style, including fictional authors, URLs, or invalid DOIs. For example, some papers directly listed "John Doe" as an author, or provided incorrectly formatted preprint IDs (such as arXiv:2305.XXXX). These errors were mainly concentrated in research from top universities like New York University and large companies such as Google.

QQ20260123-100140.png

GPTZero pointed out that the root cause of this problem lies in the "tsunami-like" increase in paper submissions in recent years — the number of submissions to NeurIPS rose from 9,467 in 2020 to 21,575 in 2025, an increase of over 220%. The heavy system load forced organizers to recruit a large number of inexperienced reviewers, and some reviewers were even accused of violating rules by using AI tools instead of reading papers manually, leading to serious damage to academic integrity. Currently, NeurIPS has clearly stated that fake citations will be a reason for rejection or withdrawal of papers.