What is an AI hallucination?
A hallucination is output that is fluent, plausible and false. The model is not malfunctioning when it happens: it is predicting likely text, and a likely sentence is not always a true one. It matters most where the output is a fact somebody will act on, such as a name, a price, or an email address.
Fluent, plausible, false. The model is working as designed.
Why it matters to you
Treating this as a model-quality problem leads nowhere useful, because a better model hallucinates less often and more convincingly. The fix is structural: do not let a model be the source of a fact it could invent.
In practice that means extraction is checked against the fetched page. On the WhoFits lead engine nothing a model returns is trusted unless the text is literally present in the page that was fetched, which is what stops an invented email address reaching a send queue. Every extracted email then passes five further checks, including a domain match against the business name.
The same idea drives citation in a retrieval system, and it is why an answer with a source attached is a different product from an answer without one.
Where this turns up in something OCTYN runs
Each line carries the date it was recorded and where it came from.
- 2026-04-15
Model output is discarded unless the text appears literally in the fetched page, which keeps invented contact details out of the send queue
WhoFits Lead Scraper docWhofits Agency → - 2026-07-29
An earlier outreach pipeline wrote emails referring to mascots the brands did not have, which is why drafting now passes a hard lint before it can be queued
docs/growth/case-studies/aarttsii.md, sections 1 and 3
If a proposal used this word and did not explain it, ask.
Book a call →