Fake Citation

Fake citations have become a distinct, named research-integrity problem separate from plagiarism and from coercive citation, driven largely by generative AI tools that produce confident, correctly-formatted references to papers that were never written. The underlying failure mode is hallucination: large language models generate plausible-sounding author names, journal titles, and DOIs because they are predicting statistically likely text, not retrieving verified bibliographic records.

Fake citation vs. coercive citation vs. plagiarism

These three problems are frequently confused because all three involve misuse of the citation record, but they are operationally distinct:

• Fake citation — the cited source does not exist, or the real source is misrepresented as saying something it does not say. The defect is in whether the reference is real and accurate.
• Coercive citation — the cited sources are entirely real, but a reviewer or editor pressures an author to add citations (often self-citations to the reviewer’s own or the journal’s own work) for no scholarly reason. The defect is in why the citation was added, not whether it is real.
• Plagiarism — presenting someone else’s words or ideas as one’s own, with or without a citation at all. The defect is in attribution of authorship, not in the citation record’s factual accuracy.

A manuscript can exhibit any one of these without the others: a paper can be full of coercive but entirely real citations, or contain fabricated references while every argument is the author’s own original work.

How fake citations enter manuscripts

Two distinct pathways produce fake citations, and they carry different institutional weight:

AI-hallucinated references

When an author asks a generative AI tool to draft a literature review, background section, or reference list, the model can invent citations wholesale — a real-sounding author, a real journal name, and a plausible year and volume, none of which resolve to an actual paper. A 2023 study in Scientific Reports tested ChatGPT-3.5 and ChatGPT-4 on 42 multidisciplinary topics and found a substantial share of the 636 generated citations were fabricated and did not exist in any database. A 2025 benchmark of eight AI chatbots found an average fabrication rate of roughly 40% across tools, with meaningful variation between models. Retraction Watch’s own analysis found fabricated-reference detections in the PubMed-indexed literature rose roughly twelve-fold over two years as generative AI use increased, and reported thousands of individual fabricated references across the biomedical literature.

Careless or dishonest human citing

Independent of AI, an author can cite a source they never actually read, misremember what a source said, cite a paper secondhand from another paper’s reference list without checking the original, or — rarely, and treated as a more serious integrity breach — knowingly invent a citation to make a claim look better-supported than it is.

Documented real-world cases

Fake citations are not a hypothetical risk; they have produced real retractions and corrections:

• Springer Nature, 2025: the publisher retracted the book Mastering Machine Learning: From Basics to Advanced after reviewers checking its citations found that a large share — roughly two-thirds of a sample checked by Retraction Watch — either did not exist or contained substantial errors, with several researchers confirming that works attributed to them were fake or misattributed. The publisher was unable to verify roughly half of the book’s total references.
• PLOS ONE, 2024: a published article was flagged on PubPeer after a commenter noticed the phrase “Regenerate response” — at the time a literal button label in the ChatGPT interface — embedded in the reference list, indicating the citations had been copy-pasted directly out of an AI chat session without review.

Both cases follow the same pattern: a fabricated or copy-pasted reference list passed initial submission and was only caught when a reader or reviewer actually tried to verify individual citations.

How journals and reviewers detect fake citations

Detection is fundamentally a verification problem, not a stylistic one — catching a fake citation means actually resolving each reference against an authoritative record rather than trusting its formatting:

• DOI and identifier resolution. A real DOI resolves to the actual publication it names; an invented or mismatched DOI is one of the fastest tells, since fabricated references frequently carry a DOI that either does not resolve at all or resolves to an unrelated paper.
• Database cross-checking. Looking up the cited author, title, and journal directly in Crossref, PubMed, or the publisher’s own index to confirm the work exists as described, rather than trusting the reference list’s formatting.
• Reading the cited passage. Confirming a real, findable source actually supports the specific claim attached to it — catching the misrepresentation case, not just outright fabrication.
• Editorial policy and disclosure. Publishers increasingly require authors to disclose generative-AI use in manuscript preparation (see generative AI disclosure statement) specifically because undisclosed AI drafting is the highest-risk pathway for fabricated references to enter a submission undetected.

Some researchers and integrity scholars argue that including a hallucinated citation in a submitted or published manuscript can meet the threshold for research misconduct under U.S. federal definitions when the citation functions as data — for example, a citation used to support a factual claim about prior findings — rather than treating it as a purely careless, non-culpable error. Whether a given case is handled as honest error, negligence, or misconduct is a matter for the journal’s or institution’s own investigation process (see how a retraction actually happens), not something the presence of a fake citation alone determines.

Worked example (hypothetical, for illustration)

A reference list entry reads: Alvarez, M. & Chen, R. (2021). “Machine Learning Approaches to Citation Verification.” Journal of Information Science, 47(3), 412-428. https://doi.org/10.1177/0165551520999999. If that DOI does not resolve, and no combination of the author names, title, and journal turns up a real matching record in Crossref or the journal’s own archive, the citation is fake regardless of how correctly it is formatted in the required citation style — correct formatting is not evidence the source exists.

Counter-example

A citation with a wrong publication year (citing a 2019 paper as 2018) or an incorrect page range, where the author, title, journal, and DOI all otherwise correctly identify a real, findable paper, is a citation error, not a fake citation — the source exists and is identifiable, the record just contains a factual mistake in one field.

References

• Walters, W.H. & Wilder, E.I. (2023) ‘Fabrication and errors in the bibliographic citations generated by ChatGPT’, Scientific Reports 13:14045.
• Retraction Watch, ‘Springer Nature book on machine learning is full of made-up citations’ (2025) and follow-up retraction coverage.
• Retraction Watch, ‘One in 277 PubMed-indexed papers in 2026 shows fabricated references, says analysis’ (2026).