The assumption science runs on

Peer review does not re-run every experiment it approves. It runs on a narrower, quieter assumption: that a citation points to something real.

A reader does not usually check whether reference #47 exists. Reviewers do not usually verify all 60 entries in a reference list. Editors do not usually re-derive a paper's evidence chain from scratch. The system works because everyone assumes someone upstream already confirmed the sources are real.

A study published in The Lancet in May 2026 shows that assumption is now measurably failing, at a rate that has grown sharply since 2023 ^[1].

The audit: what was actually measured

A team led by Maxim Topaz (Columbia University School of Nursing, Data Science Institute), with co-authors Nir Roguin, Pallavi Gupta, Zhihong Zhang and Laura-Maria Peltonen, built an AI-assisted verification pipeline and ran it across the PubMed Central Open Access subset — not the whole of PubMed, a distinction the study is explicit about ^[1,5].

The verification method itself matters: candidate references were checked against PubMed, Crossref, OpenAlex and Google Scholar; a reference was only labelled "fabricated" if it survived filtering for ordinary bibliographic noise (typos, formatting variants, incomplete metadata) and still could not be located in any of these databases ^[1]. The authors report precision at ~91%, but validated precision rather than recall — meaning the true fabrication count could be higher than what the audit caught, not lower ^[8].

The trend is the finding, not the raw count

4,046 fabricated references out of 97.1 million verified is, on its face, a small fraction. The reason this made The Lancet rather than a specialist newsletter is the trajectory ^[1,4,6,8,9]:

Two further breakdowns, independently corroborated across sources ^[8,9]:

• 91% of affected papers contained only 1–2 fabricated references; 246 papers had 3 or more.
• Review articles showed a fabrication rate 57% higher than other publication types (16.7 per 10,000 vs. 10.6 per 10,000) — notable because review articles are precisely what other researchers, and eventually clinical guidelines, build on.
• Per STAT News's reporting on the same letter, over one-third of the fabricated citations traced back to just two large open-access publishers — a concentration finding not broken down by publisher name in public reporting ^[7].

What a "fabricated reference" actually looks like

This is not typo-level sloppiness. The study's own characterization, echoed across independent coverage, is that fabricated references were "not obviously wrong" — topically specific, correctly formatted, and attributed to real researchers who never wrote the cited work ^[8].

"Your doctor could be making decisions around treatment based on studies that never existed." — Columbia University release ^[6]

Columbia's own release cites one example: a paper found to contain 18 fabricated references out of 30 total, with some already propagating into systematic reviews that inform clinical practice ^[6]. Topaz himself, describing his own experience checking a citation, said:

"I was deeply embarrassed: I checked for that, and it still almost happened to me." — Maxim Topaz ^[7]

The preprint pathway: a second, related finding

A follow-up paper by the same core team, published in the Journal of Internal Medicine (DOI: 10.1111/joim.70139), examined fabricated references specifically in medRxiv preprints and reports these appear roughly twice as often as in the peer-reviewed PMC corpus — stated directly in the paper's own title ^[3].

Correlation, not proof of cause

The authors are explicit that their data shows when the fabrication rate rose, not definitively why ^[1,8]. Three candidate mechanisms are named in the paper itself: paper-mill activity, intentional misconduct, and uncritical use of generative AI writing tools — the authors state they "could not determine cause" between these ^[8].

Independent commentators add nuance rather than certainty: Misha Teplitskiy (University of Michigan) frames the finding as early evidence of "AI slop" entering the literature; Mohammad Hosseini (Northwestern University) frames it as evidence that citation practices are becoming more superficial ^[7].

Fabricated references increased sharply during the same period generative AI tools became widely available. AI hallucination is a plausible contributor. It has not been shown to be the sole, or even the dominant, cause for every fabricated reference in this dataset.

Why this matters more than a citation-formatting problem

A citation is not decoration. It is a claim that evidence exists. Biomedical evidence typically flows through a chain:

A fabricated reference introduced early in that chain, and not caught, does not stay contained. A review author cites it in good faith. A systematic reviewer may spend real time trying to retrieve a study that was never written. Downstream authors copy the citation without re-checking the original. By the time it reaches a clinical guideline, it looks exactly as credible as every real reference around it.

What this study does not show — limitations, stated plainly

• Covers the PubMed Central Open Access subset only — not subscription-access biomedical journals, and not other fields ^[1,5,8].
• Measured precision, not recall — the true fabrication rate could be higher than reported ^[8].
• Cannot attribute any single fabricated reference to a specific cause without further investigation ^[1,8].
• The granular medRxiv follow-up statistics rest on secondary characterization, not independent confirmation ^[3].

What researchers, reviewers and publishers can actually do

Authors: treat any AI-suggested citation as a search lead, not a verified source. Open the original publication. Resolve the DOI and confirm it points to the exact title and authors listed — not merely to a paper.

Reviewers: manually verifying 40–100 references per manuscript does not scale. Automated pre-screening — does this DOI resolve, does it match this exact title and author list, does the source exist in Crossref/OpenAlex/PubMed at all — can and should happen before a manuscript reaches a human reviewer's desk.

Publishers: reference-existence checking belongs in the same automated pre-submission pipeline that already screens for plagiarism and image manipulation ^[1,7].

None of this requires abandoning AI tools in research. It requires treating every AI-suggested citation as unverified until checked against a primary source — the same discipline already expected for a data point or a statistical claim.

The ORBIXER angle: reference verification is a distinct problem from journal verification

The layered verification chain this data points toward:

The first two links are where most existing tools currently concentrate. The last three are comparatively underbuilt — and this audit is direct evidence of why they matter.

The research gap

The Lancet audit answers a question about the PMC Open Access corpus, in biomedical literature, over one specific 38-month window. It leaves open:

• Whether comparable fabrication rates exist in subscription-access biomedical journals, or in non-biomedical fields.
• What proportion of fabricated references trace to AI hallucination specifically versus paper-mill production or deliberate fabrication — the data cannot separate these mechanisms.
• Whether India-specific or other national publication ecosystems show a comparable trajectory — not examined in the published work reviewed here; an open question, not an assumption.

Bottom line

A citation is a claim that evidence exists. A 2.5-million-paper audit, independently corroborated across The Lancet, Nature News, Columbia University, STAT News, CIDRAP and Retraction Watch, shows that claim is being violated at a rate that has grown more than twelvefold in three years — and that fewer than 2% of the affected papers have so far seen any publisher correction. The evidence supports a narrower conclusion than "AI is destroying science": generative AI is a plausible contributor to a real, measured rise in fabricated citations, alongside paper mills and deliberate misconduct — and reference verification is no longer optional at current publication volumes.