4,841 papers, checked against the citations they claim
GPTZero ran its Hallucination Check tool against 4,841 of the roughly 5,290 papers NeurIPS 2025 accepted, the largest venue it had audited this way. The tool flags a citation it cannot verify against academic databases and the open web, checking the claimed authors, title, venue and link against what actually exists, and every flagged citation is then reviewed by a human before it counts as confirmed. [1]
The scan turned up 100 confirmed hallucinations spread across 53 papers, or about 1.1 percent of the papers checked. Every one of those papers had already cleared NeurIPS's peer review and appeared in the final conference proceedings. [1][2]
Atlas interpretation: One percent sounds small until it is read against NeurIPS's acceptance rate. The conference took 24.52 percent of submissions in its main track, so each of these 53 papers had already beaten out roughly 15,000 others before a single reviewer would have had reason to check whether its citations were real. [1][2]
Not typos. Citations to work that was never written
GPTZero's report distinguishes what it calls a vibe citation, a reference generated by a language model that looks correctly formatted but does not correspond to a real publication, from an ordinary citation mistake such as a wrong page number or a mistyped year. Cases the audit confirmed included wholly invented papers with fabricated authors and titles, real papers credited to coauthors who never wrote them, fake arXiv identifiers and DOIs, and citations built by mixing details from two unrelated real papers into one that does not exist. [1][2]
Atlas interpretation: The reason these get through peer review is the same reason they are easy for a tool to catch mechanically: a reviewer reads a paper for its argument and skims the bibliography for plausibility, while a fabricated citation is built to look plausible on that kind of read. Verifying that a specific title, author list and venue actually exist together is a lookup, not a judgment call, which is exactly the kind of check peer review was never set up to do at scale and a citation-matching tool can run on thousands of papers in one pass. [1]
NeurIPS says it needs more time; GPTZero says the moment already arrived
GPTZero CEO Edward Tian described the finding as a milestone for the field: papers with fabricated citations "survived peer review, and were published in the final conference proceeding. So it's definitely a big moment." NeurIPS's board acknowledged the report and said significantly more effort is required to determine the implications, while noting that a citation error does not by itself invalidate a paper's underlying research. [2]
Atlas interpretation: Those two positions are not actually in conflict. GPTZero is making a claim about detection, that fabricated citations are now findable at conference scale. NeurIPS's board is making a claim about consequence, that finding one does not settle what should happen to the paper it sits in. The report only establishes the first of those; adjudicating retractions, corrections or nothing at all is a separate process NeurIPS had not yet worked through at the time of this report. [2]