Half of ChatGPT's citations are wrong. One half is dangerous.

A Deakin University study tested GPT-4o''s citations. 56% were fabricated or contained errors. For less-researched topics, the fabrication rate approached 30%.

There are two kinds of wrong.

The paper does not exist. Easy to catch. Free tools flag it in seconds.

The paper exists, but does not say what you claimed. No tool catches this. A reviewer does. So does a retraction.

The second kind is dangerous because everything looks correct: real author, real DOI, correct year. Only the claim is wrong.

How to protect yourself today:

1. Run your reference list through a free checker — CiteTrue, SwanRef, Citely.
2. Open every source behind a key claim and find the exact line that supports it.
3. No line, no citation.

I am building an audit for the second problem — it reads the source and flags where it does not support your claim. Early access, first 100 get a free audit of one paper:

https://least-visited-lottery-fitness.trycloudflare.com

Open your sources. AI assists; you stay accountable.