A hallucination is when a model states something false with the same fluent confidence it uses for correct answers. It might invent a citation, a legal case, an API method, or a statistic that sounds plausible and simply does not exist. The unsettling part is that nothing in the tone signals a problem, so for anyone shipping a product this is the failure mode to plan around first.
This happens because a language model predicts likely-sounding text; it does not look facts up in a database of truth. When it lacks the right information, it fills the gap with whatever pattern seems most probable, which is often close enough to fool a quick reader. It is guessing, and sometimes the guess is wrong.
For a product, the risk scales with the stakes. A hallucinated brainstorm idea costs nothing, but a wrong dosage, price, or contract clause can cause real harm. Mapping where confident errors would actually hurt tells you where to add checks and where a relaxed approach is fine.
You cannot eliminate hallucination, but you can shrink it. Ground answers in retrieved source documents, ask the model to cite where each claim came from, keep a human in the loop for high-stakes output, and design interfaces that invite users to verify rather than blindly trust. Treating the model as a fast draft rather than a final authority is the safest default.