Hallucination

A hallucination is a confident, fluent statement from an AI that no real source supports. The model produces the most plausible continuation of the question and presents it with the same certainty it shows when it is right. That certainty is the danger. A hallucination does not announce itself.

Why do language models hallucinate?

Language models hallucinate because they are built to predict the next most likely word, not to check a fact. Predicting fluent text is exactly what you want for a first draft or a brainstorm. It is a liability the moment the question is a binding regulation, a maintenance procedure, or a medical protocol, where a plausible guess and a correct answer are not the same thing and only one of them is safe to act on.

Ungrounded generation has no source to answer from, so it answers from a blur of everything it has seen. Grounding removes the blur by binding the answer to a verified body of knowledge. That is the subject of the grounding page.

What does a hallucination look like in a grounded system?

The obvious case is easy to catch: the model invents a citation to a document that does not exist. The dangerous case is subtler. A grounded answer can be sourced, plausible, and still wrong, because a condition that changes it was left behind.

A maintenance manual says a component is safe to handle. A separate section, pages away, adds the condition: safe to handle, provided it has not been exposed to sunlight. A system that returns the first sentence without the second gives an answer that is grounded in a real sentence and is still a lie. Nothing signals it. The source is real, the model is confident, and the exception was never in the room. A hallucination you can catch is a nuisance. A grounded-looking answer that dropped its condition is the failure that matters.

Is a grounded answer the same as a correct answer?

Not by itself. An answer is only as good as the completeness of what was retrieved to produce it. A fact retrieved without the condition that qualifies it is a confident half-truth. This is why grounding is a systems problem and not a retrieval trick: the goal is not just to answer from a real source, but to keep the fact bound to the conditions that make it true, and to report what was not included so the gap is visible instead of silent.

How many ways can retrieval fail?

More than most teams account for. Naci Akkøk published a taxonomy of sixteen distinct ways retrieval-based systems produce wrong answers, from chunking that severs a definition from the paragraphs it depends on, to systems that cite a source for a claim the source never makes. The closest published predecessor is the GroUSE benchmark (arXiv 2409.06595), which identifies seven generator-side failure modes and which the taxonomy credits as prior art.

The point of mapping the failures is not to claim a fix for each one in a sentence. It is that most teams have a patch for one or two of them and no account of the rest. The grounding service is built to remove the structural failures by construction and to make the operational ones visible rather than silent, which is the honest form of the claim and the one worth defending.

Related terms

• Grounding: the discipline that answers from fact instead of guessing.
• Ground truth: the verified knowledge a grounded answer is checked against.
• Traceability: being able to follow an answer back to its source, so a wrong one can be caught.