Why it matters
The danger isn't that models are often wrong: it's that wrong and right sound identical. Unreviewed AI output in proposals, published content or customer answers is how businesses confidently ship fiction. (It's also why AI answers about your brand need auditing.)
How it works
Language models generate what's plausible given patterns, not what's verified against a database. When the pattern outruns the knowledge, they fill gaps smoothly. Grounding in real documents (RAG), asking for sources, and constraining scope reduce it substantially; nothing eliminates it.
What to do about it
Adopt one rule everywhere AI drafts: facts, numbers, names and quotes get verified by a human before anything ships. Design workflows assuming some hallucination, and it stops being dangerous.
