AI Citation Hallucinations
and How to Catch Them

A hallucinated citation is a reference generated by an AI system that does not correspond to any real, verifiable publication. It may be an entirely invented paper, a genuine title attached to a fabricated DOI, or a real article credited to authors who never wrote it. The output looks credible — real journal names, plausible titles, properly formatted DOIs — which is exactly why it slips past a quick visual check.

Large language models such as ChatGPT, Claude and Gemini produce these because they are built to generate statistically plausible text, not to perform database lookups. Asked for "three papers on a topic," a model assembles references that look like what such papers would look like. It is performing text completion, not retrieval.

The risk is not limited to cases where you asked an AI to find sources. A model can also insert a citation while doing something else entirely — polishing an abstract, rewording a methods section — because adding a reference makes the text read as more authoritative. You did not ask for it, and unless you re-check the bibliography against the body text, you may not notice it was added.

This is measured behaviour, not anecdote. Several peer-reviewed studies now quantify it:

• Walters & Wilder (2023), in Scientific Reports, found that a substantial share of ChatGPT-generated bibliographic citations were fabricated or contained errors significant enough to mislead a reader who did not check them (doi.org/10.1038/s41598-023-41032-5).
• Topaz and colleagues (2026) reported a reference-integrity audit spanning 2.5 million biomedical papers in The Lancet, identifying fabricated citations across the literature and a rising trend over time (doi.org/10.1016/S0140-6736(26)00603-3).
• Resnik & Hosseini (2026), in Accountability in Research, argue that hallucinated citations may constitute research misconduct when they function as data — for example in systematic reviews — and the author fails to verify the AI's output (doi.org/10.1080/08989621.2026.2645390). Whether such a lapse rises to misconduct turns on the legal standard of recklessness — distinct from mere negligence — which Caron et al. (2025) examine in detail (doi.org/10.1080/08989621.2023.2256650).

Beyond academia, the same failure mode is visible in the courts: a public database maintained by legal researcher Damien Charlotin documents more than 1,500 legal decisions in which generative AI produced fabricated citations (Charlotin, 2026). Those are court filings — a domain outside the academic databases used here — but the pattern is identical.

Every peer-reviewed study cited above was checked through AiCitationChecker before publishing this page — fitting, for an article about unverified citations.

Scientific platforms have started treating careless AI use as an actionable problem. In 2026, the preprint server arXiv announced stricter enforcement holding authors responsible for unchecked AI output — including hallucinated references — a shift widely reported across the scientific and technology press. The policy is not anti-AI; it encodes a principle that was always implicit: authors are responsible for everything in their paper, however it was generated.

The practical consequence is that a single fabricated reference is no longer a harmless slip. It can trigger a desk rejection, an integrity review, or a correction — and, because citations propagate, a fabricated reference cited once can seep into later papers and reviews. Verifying your bibliography before submission has moved from good hygiene to basic risk management.

No verification tool can cover every discipline equally. Checking a reference means matching it against scholarly databases such as CrossRef and OpenAlex, and their coverage varies by field — strong for journal articles with DOIs, thinner for some books, chapters, and niche venues. Results can therefore differ depending on what you work on.

That is exactly why every AiCitationChecker account includes free daily checks: run your own references, in your own field, and see how the tool performs before you rely on it — or pay for anything. We would rather you confirm it fits your work than take our word for it.

• Treat every AI-supplied citation as unverified until proven otherwise — a draft hypothesis, not a fact.
• Check every DOI: confirm it resolves to a paper whose title, authors and year actually match the citation.
• Cross-check against authoritative databases — a genuine paper appears in at least one of CrossRef, OpenAlex, or PubMed.
• Verify the whole reference list before submission, not just the references you knowingly asked an AI to find.
• Automate the first pass so checking is fast enough that you never skip it.

Full references for the sources cited in this article:

1. Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13(1), 14045. https://doi.org/10.1038/s41598-023-41032-5
2. Topaz, M., Roguin, N., Gupta, P., Zhang, Z., & Peltonen, L.-M. (2026). Fabricated citations: an audit across 2.5 million biomedical papers. The Lancet, 407(10541), 1779–1781. https://doi.org/10.1016/S0140-6736(26)00603-3
3. Resnik, D. B., & Hosseini, M. (2026). Hallucinated citations produced by generative artificial intelligence may constitute research misconduct when citations function as data in scholarly papers. Accountability in Research. https://doi.org/10.1080/08989621.2026.2645390
4. Caron, M. M., Dohan, S. B., Barnes, M., & Bierer, B. E. (2025). Defining “recklessness” in research misconduct proceedings. Accountability in Research, 32(2), 120–142. https://doi.org/10.1080/08989621.2023.2256650
5. Charlotin, D. (2026). AI Hallucination Cases [database]. https://www.damiencharlotin.com/hallucinations/