A hallucination is an AI model stating something false with complete confidence — an invented figure, a nonexistent reference, a plausible but wrong answer. It happens because language models produce likely text, not verified facts. Hallucinations cannot be fully eliminated, but they can be engineered down to an acceptable, measurable rate.
The standard defences are grounding the model in your documents, constraining output formats, and human approval on consequential actions. It is why a production AI system is an engineering project, not a subscription — the heart of every AI integration we ship.