A note before the recovery plan. This behaviour is structural rather than a bug in any particular version. That is bad news for anyone hoping it gets fixed, and good news for the habit below, which will keep working indefinitely.
There are two versions of the moment. In the better one, you went to check a reference before using it, searched for the paper or the case, and found nothing, then searched again with a cold feeling and confirmed that the thing ChatGPT cited so precisely has never existed. In the worse version, someone else found out first, after the document went out with your name on it.
Either way, nothing unusual happened to you. This is one of the most common and best-documented AI failures there is, from lawyers sanctioned in court for filing briefs built on nonexistent cases to a newspaper printing a reading list that paired real authors with books they never wrote. The failure has excellent production values, which is exactly the problem.
The short answer: ChatGPT, and every tool like it, invents citations because it generates text that looks right rather than retrieving records that are right, and a citation is the easiest thing in the world to fake convincingly, since the format carries all the authority and the format is trivial to reproduce. If a fake reference already went out, correct it fast and plainly, then check every other citation in the same document, because they cluster. Going forward, one rule prevents the whole category: no citation gets used until you have opened it and confirmed it says what is claimed.
Why the fakes look so real
The tool writes citations the same way it writes everything, by predicting plausible text. The tragedy is that a citation's plausibility lives entirely in its formatting, so real-sounding author names, a sensible year, a title that fits the field, a journal that exists and a page number are the whole apparatus of credibility, and all of it is surface-level.
What the tool cannot do is guarantee the underlying record exists, because it is not consulting records. It presents the invented reference with precisely the confidence of a real one, since from the inside there is no difference. The polish is not evidence of accuracy. The polish is the disguise.
If it already went out
Speed and plainness beat elaborate explanation. Tell the recipient directly that a reference in the document does not check out and is withdrawn, supply the corrected version or the honest "no source supports this" if that is the truth, and resist the urge to explain the technology at length. One clean correction reads as competence, while a paragraph about AI reads as excuse.
Then, before anything else, audit every other citation in the same document. Invented references rarely travel alone, since the same generation that produced one produced its neighbours, and the check per citation takes seconds.
The check, and the habit
| Ask of every citation | How |
|---|---|
| Does it exist | Search the exact title, or the case or article name, and find the actual source |
| Does it say what is claimed | Open it and find the claim, since real sources get misattributed too, which is the subtler failure |
| Is it current enough to use | Check the date against what the claim needs. An outdated real source can be as damaging as a fake one |
The permanent habit is one sentence long. No citation gets used until it has been opened, with no exception for references that look impeccable, because looking impeccable is the one thing fakes are reliably good at.
Two smaller habits help around it:
- Ask for links, and for anything the tool is uncertain about. That surfaces some of the weak spots at generation time rather than at the worst possible moment.
- Do citation-dependent research in a tool built to cite. Perplexity searches the live web and shows its sources, which makes verification a click rather than an investigation. Its sources want opening too, since a citing tool can still cite loosely, and there is more on that trade-off in ChatGPT vs Perplexity for research.
The wider lesson, briefly
An invented citation is the sharpest instance of the general behaviour, which is confident text with nothing underneath. The fuller guide to spotting that behaviour across everything these tools produce is how to tell when ChatGPT is making it up, and the wider routine sits in how to fact-check AI output.
For the specific wound that brought you here, the treatment is complete. Correct fast, audit the neighbours, and open every reference forever after. The habit costs seconds, and it retires this entire category of very bad afternoon.
Not sure where to start?
Research needs a tool that shows its sources.
The free finder tells you which AI to open for the task in front of you, in about 60 seconds, and the full "AI, sorted." reference lands in your inbox.
Try the free finder →One question in, one answer out.
Clair helps non-technical professionals know when to trust their AI, when to check it, and when to skip it.