A citation is supposed to be a promise: this claim rests on something real, and you can go check it yourself. That promise is breaking down faster than most researchers realize. A new Lancet-published audit, along with separate findings from AI-conference peer review, shows fabricated citations in published science climbing sharply since 2023 — and the fakes are good enough that trained reviewers are missing them.
The Numbers Behind the Surge
Researchers led by Columbia University’s Daniel Topaz audited nearly 2.5 million biomedical papers and roughly 97 million citations indexed in PubMed Central. The trend line is stark: fabricated references were rare in 2023, uncommon but rising by 2025, and now appear in more than one out of every 300 papers published in early 2026. That’s roughly a twelve-fold increase in three years — and the timing lines up almost exactly with when generative AI writing tools became common in academic workflows.
Why Trained Reviewers Keep Missing Them
The unsettling part isn’t just the volume — it’s how convincing the fakes are. The fabricated references Topaz’s team found weren’t sloppy or obviously broken. They were topically on-point, properly formatted, attributed to real researchers, and dated plausibly. In other words, they looked exactly like real citations unless someone actually tried to pull up the source.
That’s borne out at the top of the field too. GPTZero’s hallucination-detection tooling surfaced more than 50 previously unreported fabricated citations in ICLR 2026 submissions — one of the most competitive AI research venues in the world, where every paper already goes through three to five expert reviewers. A related audit of NeurIPS 2025 found roughly 100 confirmed hallucinated citations across about 1% of accepted papers, despite the same multi-reviewer process.
Perhaps the most telling data point: at the time of the Lancet audit, 98.4% of the flagged papers with fake references had not been retracted or corrected by their publishers. The problem isn’t just that fabricated citations get created — it’s that almost nothing currently catches them after publication.
How This Actually Happens
Large language models generate citations the same way they generate any other text: by predicting what a plausible-looking reference would contain, based on patterns learned from real bibliographies. When a model is asked to “find a source” or “cite evidence for” a claim, it isn’t running a database search — it’s pattern-matching against the general shape of academic citations, which is exactly why the fabrications look so legitimate on the surface.
A related audit of NeurIPS submissions broke the fabrications down by type and found most were what researchers classified as total fabrications — citations invented wholesale rather than a real source with a corrupted detail. Every hallucination examined also combined more than one deception mechanism at once, which is part of why simple spot-checks tend to miss them.
What Researchers Can Actually Do About It
None of this means avoiding AI tools in research. It means treating anything an AI tool cites as a lead to verify, not a fact to trust. The practical workflow researchers and research-integrity guides now recommend is straightforward:
Cross-check every AI-suggested citation
Before a reference goes anywhere near a bibliography, confirm it actually resolves in Crossref, PubMed, Scopus, or Google Scholar. If the DOI doesn’t resolve, or the title doesn’t match a real indexed paper, that’s disqualifying on its own.
Verify the details, not just the existence
A citation can point to a real paper and still misrepresent it — wrong authors, wrong year, or a claim the source doesn’t actually support. Confirming a source exists is step one, not the whole job.
Treat AI output as a lead, not an authority
Research-integrity guidance from multiple university libraries now frames AI-suggested sources explicitly as starting points for a real search, not results to cite directly. That framing alone would have caught the large majority of fabrications found in both the Lancet and NeurIPS audits.
The Bigger Picture
This is precisely the gap Cite Forward exists to talk about honestly. AI tools are genuinely useful for speeding up discovery, screening, and early drafting — but citation accuracy still requires a human doing real verification. The moment that step gets skipped, fabricated references don’t just risk one paper’s credibility. They quietly become the citation record other researchers build on, which is exactly how a small number of fabrications compounds into a much bigger integrity problem over time.
Sources & Further Reading
- STAT News — Fraudulent citations, blamed on AI hallucinations, are becoming more common in research papers — statnews.com
- Fortune — AI hallucinations are slipping past experts into papers and books to enter the permanent record — fortune.com
- Forbes — AI Blamed For Rise In Fabricated Citations Found In Recent Research Papers — forbes.com
- GPTZero — GPTZero uncovers 50+ Hallucinations in ICLR 2026 — gptzero.me
- ScienceDirect — Hallucinations in generative AI: A threat to scholarly integrity — sciencedirect.com
- arXiv — Compound Deception in Elite Peer Review: A Failure Mode Taxonomy of 100 Fabricated Citations at NeurIPS 2025 — arxiv.org
- Enago Academy — AI Hallucinations in Research: Why AI Citations Go Wrong — enago.com
- Indiana University Indianapolis Libraries — Verifying AI Outputs research guide — iu.libguides.com