Why AI writing tools invent citations, and how to catch them

How common is it?

Fake or wrong references from AI chatbots are well documented:

• A 2025 study of eight chatbots found only 26.5% of the references they produced were fully correct, and 39.8% were erroneous or fabricated (arXiv 2505.18059).
• A Lancet analysis reported by STAT found the share of published papers with a fabricated reference rising from 1 in 2,828 (2023) to 1 in 458 (2025) (STAT, May 2026).

Why it happens

A language model writes text that looks like a reference: plausible authors, a journal that exists, a year, even a DOI-shaped number. Unless the tool looks the reference up in a real index before showing it to you, nothing guarantees the paper exists.

Check a reference in two minutes

• Has a DOI? Open https://doi.org/ followed by the DOI. If it doesn't resolve, or resolves to a different title, the reference is wrong.
• No DOI? Search the exact title in Crossref or OpenAlex. A journal article from the last 20 years that appears in neither is a warning sign.
• Check the details. Authors, year and journal should match the record. AI-written references often mix a real title with the wrong year or a DOI from another paper.
• Books, theses and reports are often missing from these indexes, so check those in a library catalogue.

Or paste the whole list into our free reference checker: it does these checks for every entry and links each record. No sign-in.

Stop fake citations at the source

In CitedWriter the AI never writes citations. You pick papers from real indexes, and each one is checked against Crossref or OpenAlex before it can be inserted. A citation that doesn't pass simply can't go in. Other tools, such as Jenni AI's Claim Confidence, review claims after you write; checking before insert means a wrong reference never reaches your draft.