Fabricated Citations Are Now One Paper in 277 — How to Make Sure Yours Are Real

A reviewer at a mid-tier journal started doing something odd in early 2026. Before reading a submission properly, she would pick three references at random and try to open them. Not check the argument — just check the papers existed.

She had started doing it after a manuscript came through with eleven citations, four of which were to work that had never been written. The authors, when asked, were mortified. Nobody had invented anything on purpose. They had drafted with an AI assistant, and it had produced references that looked exactly like references.

She is not being paranoid. The numbers say she is being sensible.

How fast this has moved

An audit of nearly 2.5 million papers tracked how often published work contains at least one fabricated reference. The trend is steep:

• 2023 — roughly one paper in 2,828
• 2025 — one in 458
• First seven weeks of 2026 — one in 277

A twelve-fold rise in two years, with a clear inflection from the middle of 2024 — the point at which generative writing assistants stopped being something people experimented with and became part of ordinary drafting.

And a figure that should worry anyone relying on the literature staying clean: of the articles found to contain fabricated references, more than 98% had seen no publisher action at the time of the audit. The bad citations are still sitting there, being cited in turn.

Why it happens, and why it is convincing

A language model does not have a database of papers it looks things up in. It produces text that fits the pattern of the text around it. A citation has a very strong pattern: plausible author names, a plausible journal, a year, a volume, page numbers, sometimes a DOI that is correctly formatted and points at nothing.

That is why these are hard to spot. A fabricated reference is not garbled. It is often more tidy than a real one — the right sort of authors publishing the right sort of finding in exactly the journal you would expect. It reads as correct because being plausible is precisely what it was generated to be.

There is a second, subtler failure that the citation counts do not capture: the reference is real, the paper exists, and it does not say what it is being cited as saying. That one is invisible to any automated existence check, and it is probably more common.

A verification routine that takes minutes

You do not need a system. You need a habit, applied to every reference before submission.

1. Open every DOI

Not the ones you are unsure about — all of them. A DOI either resolves to the paper you think it is, or it does not. This single step catches most fabrications and takes seconds each.

2. Match the title word for word

Fabricated references frequently attach a real author to an invented title, or shift a real title by a few words. Search the exact title in Semantic Scholar or Google Scholar. Near-matches are a warning, not a rounding error.

3. Check the claim, not just the paper

Open the source and find the sentence you are relying on. This is the step everyone skips and the one that catches the real-paper-wrong-claim problem. If you cannot point to the passage, you cannot cite it.

4. Check it has not been retracted

A separate problem, and growing. A retracted paper still resolves, still has a DOI, and still looks entirely normal in a reference list.

5. Be most suspicious of the citation that fits perfectly

The reference that supports your point exactly, from a journal you half-recognise, that you cannot quite remember finding — that is the one to check first. Real literature is messier than the argument you are building from it.

Using AI in writing without this happening

None of this is an argument against the tools. It is an argument about which job you hand them.

Reasonable: summarising a paper you have already read and have open. Tightening your own prose. Suggesting search terms. Explaining an unfamiliar method. Drafting structure.

Not reasonable: asking for citations. If a model produces a reference you did not give it, treat it as a suggestion of something that might exist, and go and find out.

The reliable working pattern is to gather sources with search tools, and write with the sources in front of you. Search tools retrieve real records. Writing assistants generate plausible text. Trouble arrives when the second is asked to do the first.

What this means if you publish

Journals have started requiring disclosure of AI use, and some now require complete citation verification where it has been involved. Expect that to become standard rather than exceptional.

The practical consequence for anyone submitting work: your reference list is now something a reviewer may test directly, and a single fabricated entry does more damage than a weak paragraph. It suggests the whole manuscript was assembled rather than researched — which is unfair when a tool did it silently, and entirely understandable as a reaction.

Keep the paper trail. Note where each source came from and when you checked it. Fifteen minutes of verification is cheap next to a correction notice with your name on it.

For the tools that retrieve real records rather than generating plausible ones, see our comparison of academic search engines and research tools. And if you are running a review, our literature review method builds source-checking into the extraction stage, which is the cheapest place to catch this.

Frequently asked questions

How common are fabricated citations in published research?

An audit of nearly 2.5 million papers found roughly one paper in 277 contained at least one fabricated reference in the first seven weeks of 2026, up from one in 458 in 2025 and one in 2,828 in 2023 — a twelve-fold rise in two years.

Why do AI tools invent references?

Language models generate text that fits a pattern rather than retrieving records. A citation has a very predictable pattern, so a fabricated one tends to look tidy and entirely plausible, sometimes including a correctly formatted DOI that resolves to nothing.

How do I check whether a citation is real?

Open every DOI, match the title word for word in a real database, find the specific passage supporting your claim, and check the paper has not been retracted. Be most suspicious of the reference that fits your argument perfectly.

Can a citation be real and still wrong?

Yes, and this is probably more common than outright fabrication. The paper exists but does not say what it is cited as saying. Only reading the source catches it.

Are publishers fixing the papers already affected?

Largely not so far. More than 98% of articles found to contain fabricated references had seen no publisher action at the time of the audit.

Is it safe to use AI when writing a paper?

For summarising sources you have open, improving your own prose, suggesting search terms and explaining methods, yes. For producing citations, no. Gather with search tools; write with the sources in front of you.

Figures are from published 2026 audits of the biomedical literature reported by Retraction Watch, Nature and others. Journal policies on AI use vary — check the requirements of your target publication.