AI search

Hallucination

fabricated citation · unsupported generation

Definition

Hallucination is a generated claim the system presents as fact without support in its sources — or a citation that does not back the sentence. Liu and colleagues found generative search answers frequently include unsupported sentences and inaccurate citations. For PR, the ugly case is a model inventing a quote, a funding round, or a scandal.

You cannot “prompt” the public’s chats. You can make the true version easier to retrieve than the rumor.

How it works

Log false claims with date, engine, and prompt. Publish a canonical correction page. Ask outlets that originated errors for a correction so the retrieved set changes. Do not argue with the model in a screenshot thread as your only response.

Worked example

Claude says your CEO was convicted in 2019. She was not. The retrieved set includes a same-name defendant. You publish a bio with distinguishing facts (middle name, company, jurisdiction) and earn two trade profiles that include them. Recheck the prompt in two weeks.

How it differs

A grounded sentence points at a real span. A messy paraphrase can still be unfair without being a full hallucination. Training-data errors persist until retrieval overrides them.

Common errors

Issuing a press release titled “We do not hallucinate.” Threatening a model vendor over one staff prompt. Ignoring a falsehood because “it’s just AI.”

Sources