What Happens If a Reviewer Finds a Fabricated Citation

An editor checking the reference list of an accepted manuscript pastes its twenty-third citation's DOI into a browser and gets a page that says the identifier does not exist. The reference looks real: a plausible journal, a plausible year, three plausible author surnames. It is not attached to anything any database has ever indexed. The manuscript is two weeks from print.

A DOI that will not resolve is usually the first sign of what an editor will eventually classify as fabricated citations academic misconduct, not a formatting slip to fix quietly in proofs. A reference to a source that does not exist sits with fabricated data and falsified results on the severity scale journals actually use, well above a wrong page number or a missing initial, and reviewers now check for exactly this before a manuscript ever reaches that stage.

This piece follows that path from the reviewer's side: how a fabricated reference gets into a manuscript in the first place, what an editor does once a DOI will not resolve, the process the Committee on Publication Ethics sets out, and what actually happens to the paper afterward. TextPulse's guide to how universities and journals set AI policy covers the disclosure rules meant to stop this before submission. This piece picks up after those rules have already failed.

How Do Fabricated Citations Get Into a Manuscript?

Almost never through deliberate invention from scratch. An author drafting a literature review with an AI tool asks it to summarize the work on a topic, and the tool returns a paragraph with citations attached, because a plausible-sounding reference is part of what the pattern of academic writing looks like to a model predicting the next likely word. The reference has the shape of a real one: a named author, a real journal title, a year inside a sensible range. It just does not point at anything that was ever published.

The author copies the paragraph in, trusts the citation because the surrounding text reads competently, and never opens the source to check it exists. Time pressure explains most of what happens next: a reference list is usually the last thing checked before submission, not the first, and a fabricated entry sitting among ninety real ones is easy to miss on a read-through aimed at argument and structure rather than at each footnote. Running a finished reference list through an AI citation checker before submission catches a good number of these before a reviewer ever has to paste a DOI into a browser.

How and why a model invents a plausible-looking source in the first place is its own question, with its own mechanics. TextPulse's explainer on whether ChatGPT makes up references walks through that part in more depth than this piece needs.

Are Fabricated Citations Academic Misconduct?

Yes, and journals do not treat it as a style issue. Research misconduct is conventionally split into fabrication, falsification and plagiarism, and a reference to a source that does not exist falls into the first of those: the citation itself is invented, and its presence misrepresents what evidence actually supports the argument sitting above it in the text. That puts a fabricated citation in the same category as invented data, not beside a missing comma or an inconsistent heading style.

Elsevier's own policy states the point plainly for anyone drafting with AI assistance: generative tools may help with structure and language, but fabricating or altering data or references is listed as a prohibited use, with no lighter category set aside for a reference that turns out to be invented rather than altered. Wiley, Springer Nature and Taylor and Francis all draw the same line in their own author guidelines. Reviewers check for exactly this now, and an unresolved DOI is usually the first thing that gives a fabricated reference away.

What Happens When a Reviewer Cannot Resolve a DOI?

This mechanical check is simple enough to run in under a minute: paste the DOI into doi.org and see where it leads. When the DOI points to a not-found page or to a completely different article than the one cited, it's a strong signal that the citation was never real, especially if it lives inside a reference list an AI tool had a hand in drafting.

A single broken DOI does not end a manuscript's chances by itself, since a real identifier does occasionally get mistyped. What changes the response is scale and pattern. One broken reference usually gets a query to the author, asking for a corrected DOI or the original source. Several broken references in the same reference list, especially ones clustered in a literature review section written quickly, get escalated past a routine query and into the kind of citation integrity check an editor runs before deciding anything further.

The COPE Process an Editor Follows

COPE, the Committee on Publication Ethics, gives editors a structured way to weigh a citation problem rather than a single fixed rule. Its guidance on citation integrity sets out four factors an editor is expected to weigh before deciding what happens next.

• Nature: whether the reference is partially genuine or entirely fabricated
• Location: where it sits in the paper, for instance inside a meta-analysis or a single discussion paragraph
• Category: whether it is an unfindable reference or a genuine one cited for something it does not actually say
• Scale: whether it is one or two affected references or a large share of the whole list

How an editor weighs those four factors decides which response fits: a straightforward correction, a published expression of concern while an investigation runs, or, for the most serious pattern, contacting the author's institution directly. The reasoning COPE gives for why an AI tool cannot take the blame is the same reasoning it applies to authorship generally: responsibility for a submitted manuscript sits with the named authors, who accepted that responsibility the moment they submitted it, not with whatever tool helped draft a paragraph along the way.

"The model made it up" explains how a fabricated reference entered the manuscript. It does not excuse it. Every publisher's submission process requires the corresponding author to confirm the manuscript is accurate and take responsibility for its content, which is precisely the step that turns an AI tool's error into the author's own.

TextPulse's own guide to the ethical line for AI-assisted academic writing sets out the same responsibility principle: disclosure and verification happen before submission, not after a reviewer finds the gap.

From Correction to Retraction: What Actually Happens to the Paper

A correction is the lightest outcome. The publisher issues a note fixing the specific reference, the paper itself stays in place, and the published record shows exactly what changed and why. That is usually where a single fabricated citation lands when it turns up in a section that does not affect the paper's actual findings, and the author cooperates and supplies a real source or removes the claim it was propping up.

At the other end of the spectrum lies retraction. The very first set of guidelines on retractions, from COPE, also lists fabrication and falsification (again, the categories that a fabricated reference falls into) as one of its clear reasons for retracting a paper, along with plagiarism and the discovery of untrustworthy results. The process does not disappear quietly. A retracted paper remains visible, with the journal itself and in indexes used by other researchers tagged as retracted. This is why a made reference is so costly: the correction is permanent and public, attached to the author's name for as long as the paper is findable at all.

Disclosure is the practical defence, and how to disclose AI use in a paper is set out separately, along with a disclosure statement template you can adapt rather than draft.

Reference-checking used to be the least interesting part of a reviewer's job, a formality before the real critique started. A DOI that will not resolve has turned it into the opposite: often the fastest, cleanest signal that a manuscript needs a far closer read than the rest of it got.