Why AI invents citations, and the one check that catches most of them
AI invents citations because, to a language model, a reference is just more text that should look plausible — and a plausible title, author and page number are easy to produce without any document behind them. The check that catches most of it is simple: require the exact quoted words, and match them against the source text before trusting the claim.
That is a mechanical check, so software can do it. Most tools do not, and the gap is where the embarrassing stories come from.
Why a model produces a source that does not exist
A language model writes the most likely continuation of what came before. When the answer is expected to end with a reference, a reference is what comes next: an author name that fits the field, a journal that publishes this kind of thing, a year in the right range. Nothing in that process consults a library.
The best-known example is the 2023 New York case in which lawyers filed a brief citing court decisions that a chatbot had generated and that did not exist. The cases had names, docket numbers and quoted passages. What they did not have was a court.
Giving the model your documents helps a great deal, but it does not make the problem go away. It changes its shape.
Four ways a citation goes wrong
Once the model is working from real documents, citation failures come in four kinds, and they are not equally dangerous:
- The source is invented. A document, page or passage that is not there. Rarer when the model has your files, but not impossible.
- The quote is smoothed. The source is real, but the quoted sentence has been tidied: a hedge dropped, "may" turned into "does", two sentences merged. This is the common one, and the most dangerous, because it survives a casual glance.
- The quote is real but in the wrong place. Right words, wrong page or wrong document. Annoying rather than misleading, but it makes checking expensive.
- The quote is real and does not support the claim. Accurate words attached to a conclusion they do not carry.
The check that catches the first three
Require that every claim comes with the verbatim words it rests on, then look those words up in the source. Not "find a document that supports this" — find this exact string.
- An invented passage does not match anything, so it fails.
- A smoothed quote fails too, or matches only approximately, and the difference is the thing you needed to see. "The treatment may reduce risk" and "the treatment reduces risk" are different claims.
- A misplaced quote is found where it really is, and the citation can be corrected to point there.
You can do this by hand in any tool: copy the quoted sentence, open the source, search for it. It works. It is also tedious enough that almost nobody does it more than twice — which is why it belongs in software.
The one it does not catch
The fourth failure — a real sentence attached to a conclusion it does not support — cannot be detected by matching text. Deciding whether these words justify that claim is reading, and it stays your job.
The point of the first three checks is to make the fourth one cheap. When the quote is certain to be real and one click from its page, judging it means reading one highlighted sentence, not hunting through a chapter. That is a check people actually perform.
How MindMax handles it
Every claim in an answer about your documents carries a quote, and the quote is matched against the source text before you see it. A near miss is put back into the source's own words, so the citation stays exact. A quote that cannot be found is not shown as if it were checked: an answer flags it at the end, and an AI margin note is sent back and redone. Clicking a quote opens the document at that spot — the page of a PDF, the chapter of an ebook, the page of a presentation — with the sentence highlighted.
None of that makes the model right. It makes the model's mistakes visible, which is the property you actually need when the material is something you will act on.
What is not claimed
We do not claim MindMax never misattributes anything, and we have no hallucination-rate figure to quote. The claim is that quotes are checked mechanically, that failures are shown rather than hidden, and that the one check only a person can do has been made as cheap as it can be made.
The idea in full is checked quotes; the definitions underneath are hallucination and grounding. To see it on your own files, ask a question about a document — or download MindMax and open a folder.