ChatGPT gave you citations that don't exist. Now what?

You asked an AI for sources. It produced a tidy list: real-sounding authors, a real journal, a plausible year, even a DOI. You pasted three of them into your literature review. Then you tried to find one — and it isn't there. Not behind a paywall, not renamed, not moved. It was never published, because it was never written.

This is the most common integrity problem in academic writing right now, and the least understood. It is not a moral failure and usually not even carelessness: it is a predictable consequence of how these tools work, meeting a citation format that happens to be trivially easy to imitate. Here is why it happens, the five tells that expose an invented reference, how to verify one in about two minutes, and — the part most advice skips — what to do with the sentence it was holding up.

Why a language model invents references so convincingly

A reference is a pattern. Surname, initial, year, title-shaped phrase, journal-shaped phrase, volume, issue, page range, identifier. A model trained on millions of them learns that shape perfectly. What it does not have is a lookup table of which combinations actually exist. So when you ask for sources on a topic, it does what it always does: produces the most probable continuation. The most probable continuation of "a reference about employee motivation in the public sector" is a reference that looks exactly like that — assembled from the right parts, in the right order, describing a paper nobody wrote.

It isn't lying. It is completing a pattern, and a citation is one of the most learnable patterns in academic English.

This also explains the cruellest property of a fabricated reference: it is often more plausible than a real one. Real bibliographies are lumpy — an odd page range, a journal that renamed itself mid-decade, four authors where you expected two, a title that doesn't quite say what you need. An invented reference has none of that friction. It says precisely what your sentence requires, in flawless house style. Tidiness is a warning sign, not a reassurance.

A newer variant is worth naming, because the usual advice misses it: models with web access hallucinate less often but fail differently. They return real papers that don't support the claim, or real metadata stitched to the wrong title. The reference resolves; the attribution is still wrong. Checking that a paper exists is only half the job.

The five tells

These come from the defect classes we meet in real documents. Any one of them is enough to stop and verify.

1. The DOI that resolves to something else

The strongest single tell. An invented DOI usually 404s — but the more dangerous version resolves to a completely unrelated paper, because a plausible-looking identifier can collide with a real one. If the DOI opens a different title than your reference claims, the reference is wrong even though the link "works".

2. Real authors, real journal, unreal paper

The most convincing shape: two genuine scholars who genuinely publish on the topic, in a journal that genuinely exists, on a paper neither of them wrote. Every component checks out individually. Only the combination is fiction.

3. The title that is too perfectly on-point

If the title reads like a restatement of your own sentence — your variables, your population, your framing, all in one line — treat it as suspect. Real research is rarely that obliging.

4. Volume, issue and pages that don't line up

Journals number predictably. A 2021 article cited in a volume the journal reached in 2016, or a page range beyond the length of that issue, is a fabrication tell that survives even when authors and title are real.

5. A cluster that formats identically

Five references in a row with the same punctuation habits, the same author-count, the same page-range width. Human bibliographies, assembled over years from different sources, don't look like that. Machine-generated batches do.

Okafor, T., & Lindqvist, M. (2021). Employee motivation and
  organisational effectiveness in public institutions.
  Journal of Public Administration Review, 34(2), 118-137.
  https://doi.org/10.1016/j.jpar.2021.04.007

✓ Authors publish in this field       real
✓ Journal exists                     real
✗ Title returns 0 results, exact match  INVENTED
✗ DOI resolves — to a different paper   MISMATCH
✗ Vol. 34 was 2017 for this journal     IMPOSSIBLE

The names, journal and identifier above are invented for illustration.

How to verify one in two minutes

1. Search the exact title in quotation marks. A real paper appears in the publisher's catalogue, an indexing service, or an institutional repository. Zero exact-match results is close to a verdict on its own.
2. Click the DOI. Then read the page it lands on and confirm the title, authors and year match what you cited. A resolving link is not agreement.
3. Check the volume against the journal's own archive. Publishers list volumes by year; a mismatch is arithmetic, not opinion.
4. Look at the authors' own publication lists. An ORCID record or departmental page settles whether they wrote the paper attributed to them.
5. Distrust echo results. Aggregator sites and AI-written blog posts now republish fabricated references, so a hit that isn't a publisher, index or repository is not evidence. Follow it to a primary source or discard it.

What to do once you find one

There are exactly three honest options, and the order matters.

1. Find the real source for the claim. The idea in your sentence usually does have support — the model was pattern-matching against a real literature. Search the topic properly, read what you find, and cite that. This is the only outcome where your argument survives intact.
2. Rewrite the claim to match what you can actually cite. Often the real literature says something narrower or more hedged than the invented reference promised. Say the narrower thing.
3. Delete it. If the claim isn't load-bearing and nothing supports it, it goes. A thesis is not weakened by removing a sentence that was never evidenced.

What you must not do is keep the citation because it looks right, or swap in a real paper you haven't read because the title fits. Both convert an honest error into a fabricated one — and the second is harder to spot and worse to be caught at. If a reference stays in your document, it must resolve to something you can put in front of an examiner.

Why this matters more than it feels like it should

An invented reference is not read as a formatting slip. It is read as evidence that the writer did not consult their own sources — and examiners now check, because they know what these tools produce. A single unverifiable reference invites a second look at the entire bibliography.

Which is also the argument for finding them yourself, early and unglamorously. A hallucinated citation caught during your own pass is a two-minute correction. The same citation caught by an examiner is a conversation about your process. We built our integrity report around exactly that distinction: it is a quality instrument, not an accusation, and it exists so the writer finds the problem first.

How our own AI is built so it cannot do this

Since this post is about machines inventing references, we owe you the constraint we work under. Ref[In] AI never authors a reference. It cannot: the only records it is allowed to propose come from academic search engines — Crossref, OpenAlex, PubMed and their peers — and each candidate arrives carrying the identifier it was found by. The AI's job is judgement on candidates that already exist: is this the work your sentence cites, is this year the same paper, does this record fit the passage? A generative model writing a reference from memory is precisely the failure mode this architecture removes, and it is why every suggestion is a proposal you approve rather than an edit that happens to you. We do not train models on your document either; the work is checking, not learning.

One last thing worth doing while your bibliography is open: a reference can be entirely real and quietly out of date — a superseded edition, a preprint that has since been published properly. That is a different job from proving existence, and our sister tool covers it in how to update outdated references before submission.

The checklist

• Every AI-suggested reference verified against a publisher, index or repository — not a search-result snippet.
• Every DOI clicked, and the landing page's title and authors compared to your entry.
• Volume and year cross-checked against the journal's archive.
• Any reference that cannot be confirmed: replaced with a real source, rewritten to what the real literature supports, or removed.
• Nothing kept because it merely looks correct.