The 2026 Citation Integrity Crisis: What the Numbers Tell Researchers
A citation that resolves to a real DOI used to be a reasonable proxy for manuscript trustworthiness. That proxy is breaking down. Citation integrity in 2026 is not a theoretical worry about AI-assisted writing anymore — it is a measured, accelerating pattern in the published biomedical literature, tracked across nearly 100 million references.
The scale, in numbers
A Lancet-published audit led by Maxim Topaz at Columbia University's Data Science Institute checked 97.1 million references across nearly 2.5 million PubMed-indexed papers. In 2023, roughly 1 in 2,828 papers contained a fabricated reference. By 2025, that rate reached 1 in 458. Across the first seven weeks of 2026, it hit 1 in 277 — a sixfold jump from 2023 to 2025 alone, as Retraction Watch reported. Review articles fared worse than other paper types, with a fabrication rate 57% higher — a specific concern given how often reviews feed into clinical guidelines.
Why the line moved in mid-2024
Topaz's team located the sharpest jump in hallucinated references in mid-2024, which coincided with the wider adoption of AI writing tools in manuscript preparation. The audit does not claim to prove causation from timing alone — coincidence is not the same as mechanism. But the shape of the curve, combined with what is already known about how large language models generate references (plausible author names, plausible journal titles, no actual retrieval step), makes the mechanism behind citation hallucination a reasonable explanation, not a leap.
What publishers are actually catching
A separate analysis, run by Grounded AI in collaboration with Nature, sampled 4,000-plus 2025 publications across five major publishers — Elsevier, Sage, Springer Nature, Taylor & Francis, and Wiley — and estimated that 1.65% of 2025 publications, over 110,000 papers, contain at least one fabricated citation. Manual review confirmed at least one invalid reference in 65 of the 100 highest-risk publications the model flagged. The same analysis found that pre-publication screening currently catches only 5-10% of the fabrication rate that later surfaces in the published literature — screening exists, but it is not close to comprehensive yet.
Publisher policy is catching up faster than publisher tooling. Elsevier's generative-AI policy, updated June 2026, now states plainly that AI-generated references "can be incorrect, hallucinated, or fabricated," puts the verification burden on the author, and warns that fabricated references can be grounds for manuscript rejection. That is a real deterrent. It is not a technical safeguard.
The gap most checks don't cover
Most reference-verification tools, automated or manual, operate at the metadata layer: does this citation resolve to a real paper with a matching title, author list, and journal. That layer catches fabricated DOIs and invented papers, which is most of what the audits above measured. It does not catch a real, correctly cited paper that simply doesn't support the sentence attached to it — a distinct failure the CIVER framework separates from reference existence as its own tier. CiteCheck, built as the reference implementation of CIVER's citation-fidelity axis, checks the existence layer against CrossRef, PubMed, Semantic Scholar, and OpenAlex; it does not by itself confirm claim support, and no automated existence-check currently does either.
Where this leaves individual researchers
Submission-side screening is real but still catches a minority of what eventually surfaces. Until that changes, the more reliable gate sits before submission, in the author's own workflow, not after it in the editor's. The Check Citations tool inside AI for Academic runs the same CrossRef/PubMed/Semantic Scholar/OpenAlex cross-check CiteCheck implements, free to start, and it takes a manuscript upload rather than a citation-by-citation manual lookup. Given where the fabrication rate is trending, that check belongs earlier in the writing process than most researchers currently place it — before the cover letter, not after a reviewer flags it.