How many accepted conference papers contain hallucinated citations?

GPTZero reports that it scanned 4841 of the 5290 papers accepted to NeurIPS 2025 and found "noticeable traces of AI authorship and hundreds of vibe citations." The company, which sells an AI-text detector and a citation checker called Hallucination Check, frames this as a flaw in the peer review pipeline, not a charge against named organizers or reviewers. Its account of the pressure behind it: NeurIPS submissions rose "from 9,467 to 21,575" between 2020 and 2025, and at a 24.52% main-track acceptance rate each accepted paper "beat out 15,000 other papers despite containing one or more hallucinations." The title's headline count is 100 new hallucinations. The body never gives that number; it says only "hundreds."

The mechanism is a citation check. Hallucination Check uses an in-house agent to flag any citation it cannot find online. GPTZero is explicit that a flag is not a verdict: many archival or unpublished works "can't be matched to an online source," so flags "indicate which sources require further human scrutiny." It defines a vibe citation as one that "likely resulted from the use of generative AI," with errors such as fabricated authors, titles, DOIs or containers, and first names extrapolated from initials. The company claims the tool catches "99 out of 100 flawed citations," with a higher false positive rate because any citation it cannot verify gets flagged. The examples were checked by people: "each of the hallucinations presented here has been verified by a human expert."

Set against the nearest notes, this excerpt is about what the process let through, not how reviewers behaved. Does banning LLM use in peer review change review outcomes? measured reviewer conduct under policy; GPTZero checks the output and its citations, not the reasoning. Can inference scaling help reviewers catch errors humans miss? spends heavy compute per manuscript on proofs and experiments. Citation matching is shallower, but it is mechanical and runs against public records, which is why one pass can cover 4841 papers. The excerpt still routes its flags to people, and Can people reliably spot content made by AI? suggests human judgment is a weak check for AI content on its own, though these flags are checked against online records, not by eye. The vendor-claim problem also recurs: How often do legal AI tools actually hallucinate citations? tested "hallucination-free" marketing independently, while the 99-in-100 figure here is GPTZero's own and untested in the excerpt.

The excerpt does not establish how many flagged citations are confirmed hallucinations. The title's 100 is never reconciled with the body's "hundreds," and the human verification it describes covers "the hallucinations presented here," not the full scan. Its growth figure also needs care: 21,575 over 9,467 is a ratio of about 2.28, or growth of about 128 percent, so "more than 220%" reads as a ratio mislabeled as growth. There is no false positive rate, no per-paper breakdown, no comparison with other venues, and no NeurIPS response. The defensible implication is narrow: unverifiable citations appear often enough in accepted ML papers to justify venue-level checks, and those checks are cheap to run. The scan does not measure how many accepted papers contain fabricated sources. Its figures are the vendor's own until an independent audit reproduces them.