Hallucination
A hallucination is text a language model produces that reads as confident and well-formed but is not supported by any source — an invented citation, a statistic that does not exist, a quote nobody wrote.
- Also called
- Confabulation, fabrication
- Better models
- Rarer, not gone
- Practical fix
- Make checking cheap
What it means when the material is your own documents
The general worry about hallucination is about facts in the world. When you are working through documents you own, it takes a narrower and more annoying form: the model summarises a paper you have not read, and there is no cheap way to tell a good summary from a plausible one. Everything sounds equally fluent. That is the problem — not that the model is wrong often, but that being wrong and being right look identical from the outside.
Three shapes, three different defences. Only the first can be caught by the software alone.
- A quote that was never written. Close to something in the text, smoothed into better prose.
- A real quote attached to the wrong conclusion. The words exist; they do not support the claim they are cited for.
- A confident answer about a passage that was never retrieved. With chunk-based retrieval, if the relevant paragraph was not fetched, the model still answers — from the surrounding material and its priors.
Better models make all three rarer. None of them make checking unnecessary, because the failure is invisible at exactly the moment you would need to catch it: you cannot tell which answers to double-check without already knowing the material.
How MindMax handles it
Rather than promising to eliminate hallucination, MindMax makes it cheap to catch. Every claim in an answer carries the verbatim quote it rests on, and the quote is matched against the source text before you see it. A reworded quote is replaced with the source’s own words; one that cannot be found is flagged at the end of the answer rather than shown as checked. Clicking a quote takes you to the passage it came from.
That handles the first shape mechanically. The second — a real quote, the wrong conclusion — is not something matching can detect, but it is now one sentence away from you. The third is avoided at the source: a document that fits the reading budget is read whole instead of retrieved in fragments.
The result is not a guarantee of correctness — no tool can give you that. It is a shift in cost: checking a claim goes from “re-read the chapter” to “click and read one sentence”, which is the difference between a check you will perform and one you will not.
A claim about a novel, with its quote opened in place: the highlighted words are the novel’s own, shown in their surrounding text.
Common questions
Why do AI models hallucinate?
A language model produces the most plausible continuation of text, and plausible is not the same as supported. Without a source to anchor a statement — or when the relevant passage was never shown to it — the model fills the gap with something that reads right.
How do I catch a hallucination in an AI summary of my document?
Check the quotes, not the prose. Ask for every claim to carry the exact words it rests on, and check that those words are really in the document and really say what the summary claims. MindMax matches each quote against the source before showing it, so what is left for you is the one-sentence judgement.
Will a better model stop hallucinating?
It will hallucinate less often, but not never — and because a wrong answer looks exactly like a right one, a lower rate does not tell you which answers to trust. Cheap verification is what changes that.