How do I know if ChatGPT made up a source?

The short answer

You open the links. A fabricated citation is fluent, correctly formatted, and frequently almost real — the right journal, a plausible title, an article number that does not exist. It is not detectable by reading; it is only detectable by fetching. In 26 AI answers we checked, 4 contained a citation that does not hold up — 15%. None of them looked wrong on the page.

What a made-up source actually looks like

People expect an invented citation to look invented. They do not. These are real examples, from real answers, with public receipts:

https://www.cell.com/cell-metabolism/fulltext/S1550-4131
✕ Truncated
A real journal, a real URL pattern, and no article identifier — the citation stops before the part that would make it a citation. Two different assistants produced this same stub on two unrelated questions.
alzdiscovery.org/uploads/cognitive_vitality_media/…
✕ 404
A real organisation, a plausible path, and the site itself says the page does not exist.
https://www.amjmed.com/article/S0002-9343
✕ Truncated
The American Journal of Medicine, cut off at the prefix.

Why "it looked fine" is not evidence

The failure mode is specific: a model reproduces the shape of a citation it has seen thousands of times, without the identifier that points at a particular paper. The result passes every test a reader applies — right domain, right format, right subject — and fails the only test that matters, which is whether the page is there.

What a paywall is not

This is where most checking goes wrong. ScienceDirect returns 403 to automated requests. PubMed answers 203. A checker that counts every non-200 as a fake citation will tell you most of a sound bibliography is invented. Of 429 citations we opened, 82 were paywalled or bot-blocked — nearly one in five — and every one of them was real. A checker you cannot trust is worse than no checker, because you stop reading it.

How to check one yourself

  1. Open every link. Not the first one. The fabricated citation is rarely first.
  2. Look for a truncated identifier — a URL that ends at the journal prefix.
  3. Read the page, not the title. A live page that does not say what was claimed is a different failure, and a more common one.
  4. Treat 401/403/429 as "could not open", never as "fake".

Or have it checked

Whysaid opens every source an answer cites, classifies each one — real, missing, invented, truncated, or merely paywalled — and issues a public receipt you can link to. The receipt is the point: a verdict nobody can open proves nothing.

Check an answer — from $9/month Agents: verify per call