Study finds a surge of fake citations after AI writing tools spread
Reviewed by Helen Jones
A 24 September 2026 paper reports an audit of 111 million references across 2.5 million papers in arXiv, bioRxiv, SSRN, and PubMed Central. After large language models became common writing tools, non-existent bibliographic entries rose sharply. The authors give a conservative estimate of 146,932 hallucinated citations in 2025 alone. The typical pattern was a few bad references per paper, concentrated in fields with fast AI uptake, and more common among smaller or early-career teams. Those papers also showed language that looks like LLM writing. Authors linked to hallucinated citations had far fewer prior papers, about 62 percent fewer on arXiv and bioRxiv and up to about 73 percent fewer on SSRN, then posted much more by 2025, with relative output rising about 1.3 to 3.1 times depending on the corpus. Credit also tilted toward already prominent, male scholars. The study infers this from timing and writing patterns, not from tool logs, and it covers scholarly citations, not workplace files.
Before this audit, fake citations were treated as an anecdote from a sloppy draft. A manager could still believe that a polished report with named sources was basically trustworthy, and that checking every reference was extra work rather than part of the job. What changed is the scale of a checkable error. The model is not only inventing facts. It is inventing the paper trail that makes those facts look reviewed. That is the same failure as a briefing that cites a policy, a vendor, or a past memo that does not exist. Speed without a source check is not a workflow. It is a slot machine with footnotes.
Analysis
Treat this as a change to act on, not a trend to watch from the sidelines. Before the next research, competitor, or meeting-prep workflow goes to the team, add one mandatory review step: every named source must be opened and confirmed, or it is stripped from the draft.
Source note
Pulse published by Collab365 Spaces, reviewed by Helen Jones on . Cite as "Study finds a surge of fake citations after AI writing tools spread", Collab365 Spaces. 1 source referenced.