What is Hallucination?
A hallucination is output from a language model that is fluent, confident and false — an invented statistic, a misattributed quote, a citation to a source that does not exist, or a claim about a company that is simply untrue.
The cause is structural rather than a bug. A language model produces text that is likely given what came before, and a plausible-sounding falsehood can be more likely than an admission of ignorance. Nothing in the generation step checks a claim against the world.
Retrieval reduces the rate substantially by putting real documents in front of the model, but does not eliminate it. A model can still misread a retrieved document, blend two sources into a claim neither made, or cite a retrieved page for a fact that appears nowhere in it.
For a brand the risk runs in two directions. An assistant may state something false about the company — wrong pricing, a discontinued feature, an invented limitation — to a buyer who will never verify it. And a tool that generates content can hallucinate a statistic into an article that then gets published under the brand's own name, which is how invented figures enter the record as if sourced.
The second risk is the one a content tool controls, and the honest response is to refuse rather than guess: leave a marked gap where a real figure belongs and block publishing until a human supplies it.
Why Hallucination matters
An assistant stating something false about a brand reaches a buyer who will not check it, and a generated article that invents a statistic publishes that invention under the brand's byline.
Fluency is not accuracy
Confident phrasing carries no information about truth. The most fluent passage in an answer can be the fabricated one.
Retrieval reduces but does not remove it
Grounded answers still misread sources and still cite retrieved pages for claims those pages never made.
Fabricated citations are common
Plausible author, title and date combinations pointing at nothing are a recognised failure mode. Every reference needs checking.
Refusing beats guessing in generated content
A marked placeholder that blocks publishing is recoverable. An invented statistic that ships is not.
Common questions
Fix the source material first: publish a clear, current, easily retrieved page stating the correct fact, since grounded answers draw from what they retrieve. Then monitor whether the correction propagates — assistants update at different rates, and some cache answers. There is no direct edit mechanism for most of them.