A fabricated reference used to be rare, and usually a sign of deliberate misconduct. Now it's common enough to measure at scale, and most of the time nobody meant to deceive anyone. Someone asked a chatbot for sources, got a tidy list, and didn't check it.
This post looks at the best current evidence on how often it happens, explains why the tools do it, and sets out a routine for checking a reference list that takes a couple of minutes per source.
- 1 in 277PubMed papers in early 2026 that cited at least one paper that doesn't exist
- 10×Increase on 2023, when the rate was 1 in 2,828 papers
- 19.9%Citations from GPT-4o literature reviews that were fabricated (Linardon et al., 2025)
- 56%Citations in the same study that were fabricated or contained errors
What the audits found
Published papers: a tenfold rise
In May 2026, a research letter in The Lancet by Maxim Topaz and colleagues reported an audit of about 2.5 million biomedical papers in PubMed Central, published between January 2023 and February 2026. The team checked roughly 97 million references against PubMed, Crossref, OpenAlex and Google Scholar, and flagged those that matched nothing in any of the four.
The share of papers citing at least one non-existent paper went from about 1 in 2,828 in 2023, to 1 in 458 in 2025, to 1 in 277 in the first seven weeks of 2026. The sharpest rise came in mid-2024, as AI writing tools spread. According to Retraction Watch, more than 98% of the affected articles had seen no action from their publishers at the time of the audit.
The audit has critics. It counts a fabricated reference in a minor background sentence the same as one supporting a central claim, and it covers biomedicine only. But these papers had all passed peer review. The checks that are supposed to catch this aren't catching it.
AI-generated reviews: one in five citations invented
A 2025 study in JMIR Mental Health by Jake Linardon and colleagues at Deakin University asked GPT-4o to write six literature reviews on mental health topics, then checked every citation by hand. Of 176 citations, 35 (19.9%) didn't exist. Of the real ones, many had wrong details such as the year, journal or DOI. Altogether, 56% were fabricated or inaccurate.
The most useful finding is where it got worse. For a well-studied topic like major depression, fabrication was rare. For less-studied topics it was several times higher. A model has seen fewer real papers on a niche topic, so it fills the gap with plausible ones, and niche topics are exactly where a researcher is least likely to spot a fake.
Why AI tools invent references
A language model writes by predicting likely text. A reference is one of the most predictable patterns in academic writing: surnames, initials, a year in brackets, a title in sentence case, a journal, a volume, a DOI. The model can produce all of that fluently without retrieving anything. The result looks exactly as convincing as a real reference, because looking convincing is what the model is good at.
Tools that search a real database first are much safer, but they fail in quieter ways:
- the paper exists, but the details in the citation are wrong;
- the paper exists, but doesn't say what the sentence claims it says;
- the paper exists, but has since been retracted.
Five kinds of bad reference
| Type | What it looks like | How you catch it |
|---|---|---|
| Invented | A complete, plausible reference to a paper that doesn't exist | No match for the title in any database |
| Hybrid | Real authors who work in the field, attached to a title they never wrote | The authors' publication lists don't include it |
| Borrowed DOI | A real DOI that resolves to a different paper | Resolve the DOI and compare the title |
| Wrong details | A real paper with the wrong year, journal, volume or pages | Compare every field against the publisher's record |
| Misattributed | A real, correct reference that doesn't support the claim | Only by reading the paper |
The first four are mechanical and quick to catch. The fifth is the hardest and the most damaging, because everything about the reference checks out except the one thing that matters.
The five-step check
Run this on every reference you didn't add yourself from the paper in front of you.
- Resolve the DOI. Paste it after
https://doi.org/. It should land on the publisher's page for a paper with the same title and authors. Landing somewhere is not enough. - Search the exact title. Use quotation marks in Google Scholar, PubMed or Crossref. A real paper normally appears in more than one index. If you can only find it in one place, look harder.
- Compare the details. Authors, year, journal, volume, issue, pages. Fix any mismatch from the publisher's record, not from the citation you were given.
- Find the sentence. Open the paper and find the passage that supports your claim. Note the page. If you can't find it, the reference doesn't support that sentence, however relevant the title sounds.
- Check for retractions. Look at the publisher's page for a retraction notice, or search the Retraction Watch database, which Crossref now makes openly available.
Steps 1 to 3 take about a minute per reference once you're practised. Step 4 takes one or two more if you know the paper. For a typical article with 50 references, that's roughly two to three hours. Set that against a correction notice with your name on it, or a misconduct inquiry.
Warning signs worth a closer look
- The title matches your argument almost word for word.
- Well-known authors appear together on a paper you've never heard of in your own field.
- The journal doesn't publish that kind of article, or that topic.
- Page numbers or volume numbers don't fit the journal's numbering.
- The DOI prefix belongs to a different publisher from the named journal.
- Several references share an oddly similar structure or come from the same few years.
A workflow that doesn't create the problem
Checking catches fake citations. A better workflow stops them appearing at all.
- Never ask a general chatbot for references. Ask it about ideas if you like, then find the sources yourself in a scholarly database.
- Use AI on papers you already have. Asking questions of a PDF you've downloaded can't invent a paper, and good tools show you the page each answer came from.
- Insert citations from records, not text. Add a reference from the database or publisher's page, so its details come from the source.
- Attach the citation when you write the claim. Leaving "[cite]" to fill in later is where plausible-but-wrong references creep in.
This is how LitraMind handles it. Research & Cite searches scholarly sources for the claim you've highlighted and inserts the citation from the real record, in your chosen style. Chat with Files answers questions from your own PDFs with page references. Neither removes step 4: you still need to read enough of the paper to know it supports your sentence. For more on keeping citations under control as you write, see how to manage citations without breaking your writing flow.
If you supervise, review or edit
- Spot-check five references at random in every manuscript or thesis chapter. It takes ten minutes and catches most problems, because fabricated references rarely come alone.
- Check the references behind key claims, not only the existence of the list.
- Treat a fabricated reference as a reason to ask questions, not proof of intent. Most come from careless tool use. All of them need fixing.
Checklist
- Every DOI resolves to the paper with the same title and authors
- Every title appears in at least one scholarly index
- Every detail matches the publisher's record
- Every cited claim can be found in the paper, with a page number in your notes
- No cited paper has been retracted
- No reference came straight from a chatbot without being checked
