Hallucination (AI)
When an AI states something false with complete confidence, such as an invented fact, number, feature or source, because it produces plausible text rather than checked truth.
Updated October 2026 · with examples from our own builds
What does hallucination mean in practice?
Models generate the most likely next words, and a confident falsehood can be very likely. They invent statistics, product features, certifications and citations, and nothing in the output marks which parts are made up.
The most damaging cases are plausible claims about your own business, published under your name: a certification you don't hold, an uptime you never promised, a care instruction with no source. Style rules won't catch these; only a check against what you actually know will.
Products reduce the risk by keeping facts in data the model can't overwrite, checking generated claims against sources in code, and holding anything unsupported for a person.
Hallucination at a glance
- Looks like
- Confident, specific and wrong
- Costliest when
- It concerns your own business, published as fact
- Reduce with
- Facts in data, claims checked in code
Where has hallucination come up in our builds?
Run on a real Storafleet post, the support gate caught both planted inventions, an ISO 27001 certification and a 99.99% uptime SLA, while 'over fifty storage providers' passed because Storafleet says '50+' on four of its own pages. A held post has no publish-anyway button.
Pet toxicity is only ever imported from published records, never written by a model, and reads harmful, cleared or not assessed; an unassessed animal is shown as unknown.
Early AI-built forms looked right but broke: on a priced-menu test, 18 to 25 operations per run failed because totals pointed at price fields that didn't exist. Once code wrote every reference from the model's plan, the same model went to none.
Common questions about hallucination
Can AI-written content be published without review?
Only content whose claims have been checked against your own sources. Anything stating facts about your business, figures or certifications should either be verified in code or held for a person.
Will a newer model stop making things up?
Newer models make fewer mistakes, but none of them is a fact-checker. Design as if any specific claim might be invented, and check the ones that matter.
Related terms
Sources
Planning a product and not sure where hallucination fits? Ask us, or browse all 116 terms in the glossary.