Hallucination (AI)
When a model generates plausible-sounding content that's false or unsupported by any real source.
Hallucination is when a model generates content that reads as fluent and plausible but isn't actually supported by a real source: a fabricated statistic, a citation that doesn't exist, a clinical-sounding claim with nothing behind it. It happens because a language model is optimized to produce likely-sounding text, not verified text, and from the outside those two things can look identical.
In most AI use cases, a hallucination is an annoying error someone catches and fixes. In medical and promotional content it's a different category of failure, because the entire point of the output is to be an accurate clinical or regulatory statement, and a fabricated one is indistinguishable from a real one right up until someone checks it against the actual label. That's exactly the gap a general-purpose model can't close on its own, since it has no built-in concept of what a specific product is actually approved to claim.