The signs worth noticing
- Suspiciously perfect fit. The reference supports the sentence precisely, with a title that reads like a restatement of the claim.
- Generic titles. "A comprehensive review of…", "The effects of X on Y: evidence from a longitudinal study" — templates rather than the specific, often awkward titles real papers have.
- Mismatched parts. Real authors attached to the wrong journal, a volume or issue that does not exist for that year, page ranges that are too tidy, a DOI whose prefix belongs to a different publisher.
- No identifiers. A missing DOI, arXiv number or publisher link, especially in a list where the other entries have them.
- Quotations without a locatable page. A quoted sentence you cannot find by searching the document.
None of these prove fabrication on its own. Real references are sometimes sloppy, and invented ones are sometimes immaculate. They tell you which entries to check first.
The five-minute check
- Search the title in quotation marks in a scholarly index or a general search engine. A real paper almost always surfaces somewhere: the publisher, a repository, a preprint server, a citation in another work.
- Resolve the DOI. Put it after
https://doi.org/. If it does not resolve, the identifier is wrong, and a wrong identifier on a confident reference is a strong signal. - Check the author actually wrote it. Look at the author's publication list. Real researchers work in recognisable areas; a paper far outside them deserves scrutiny.
- Open the source and find the sentence. Existence is not support. Search the paper for the claim and read the surrounding paragraph.
- For a quotation from your own PDF, search a distinctive five-word fragment of it in the document itself. Either it is there, at the page given, or it is not.
What to do with a bad reference
Do not repair it by guessing which real paper was meant — that produces a citation you have not read attached to a claim you have not verified. Delete the claim, or replace it with a source you have actually opened. If a reference is partly wrong (right paper, wrong year or page), fix it from the paper itself, not from the answer.
Keep a line per reference while you work:
- Reference as given: authors, title, venue, year, identifier.
- Found? where you found it, or "not found" with what you searched.
- Supports the claim? the page or passage you read, or "does not support".
- Action: keep, correct, replace, or remove.
Reduce the risk in the first place
Ask for claims about the document in front of you rather than about the literature at large, and prefer answers that quote. When you need outside sources, gather them yourself from an index and then use AI on the papers you have downloaded. Working from a PDF you uploaded turns "is this real?" into a question you can answer by looking at a page.
Related: Find the page a claim came from and How to check an AI explanation of a research paper.