Fake Citations Climb in Scientific Papers
A Columbia audit finds fabricated references rising sharply, with the team suspecting AI tools are involved, though that link is unconfirmed.
Updated on Nov. 2, 2026 in Environmental
Live Poll
Do you trust scientific research to be accurate as AI tools spread?
A Columbia University team found 1,135 likely fabricated references across 619 papers in physics, computer science, materials science, engineering, and chemistry, after checking 21.6 million references from 500,000 studies indexed between January 2023 and March 2026. The findings, published in the Proceedings of the National Academy of Sciences, are research-stage analysis of published literature rather than a new product or policy.
Why it matters
The fabricated-reference rate rose from 6.3 per 10,000 studies in 2023 to 23 per 10,000 in the first quarter of 2026, a trend the researchers link to growing use of AI models, though they have not confirmed the cause. The pattern signals a growing verification problem for scientific publishing as machine-written text becomes more common.
The audit covered 21.6 million references drawn from 500,000 studies indexed between January 2023 and March 31, 2026. Using title matching across seven sources with a language model filtering formatting false positives, the team flagged 1,135 references as fabricated, with most affected papers containing only one or two.
The players
Columbia University
A research university whose team, led by Topaz, conducted the reference audit across physics, computer science, materials science, engineering, and chemistry literature.
Cornell University
A research university whose researcher Yin, uninvolved in the study, commented on the findings.
PubMed Central
The National Library of Medicine's free full-text archive of biomedical and life sciences literature, which a separate preprint from the team examined for AI-involvement evidence.
The details
Researchers extracted every reference from each paper and compared the cited title against the title at the included DOI link, a unique identifier for each publication. A large language model, a type of AI system trained on text, was used to remove false positives caused by formatting differences. Titles that could not be matched in any of seven sources, including scholarly databases and search engines, were classified as fabricated.
Timeline
2023: fabricated-reference rate in biomedical papers was four per 10,000.
January 2023 to March 31, 2026: window of studies sampled from OpenAlex.
May 2026: Topaz team published its biomedical fabricated-reference findings.
End of 2025: biomedical fabricated-reference rate reached 51.3 per 10,000.
The Tech Race
The findings place the verification problem alongside the separate preprint from the same team, which found many AI models still fabricate references. Together they show that detection of machine-generated citations is now a research front in its own right, with the Columbia audit serving as a measured baseline for future tracking.
Journal editors, reviewers, and readers who rely on reference lists to check claims face a rising share of citations that do not resolve to real publications. Publishers and indexing services may need to add verification steps, though no specific tools or timelines have been announced.
The takeaway
The fabricated-reference rate is the trajectory to watch, and the researchers' planned work on references lacking DOIs will show whether the hidden share is larger than the current count. Readers can track whether follow-up audits in the Proceedings of the National Academy of Sciences confirm the rising trend beyond the first quarter of 2026.
Further reading
For broader context on how research integrity is tracked across the field, see Environmental coverage.
Source note: This article includes information reported by The Scientist.
Live Poll
Do you trust scientific research to be accurate as AI tools spread?
