The cause is structural. A language model is trained to continue text plausibly, and it has no internal database of facts to check against — knowledge is distributed across weights as statistical pattern, not stored as records. Confidence in the output correlates weakly with accuracy, which is why a fabricated legal citation reads exactly like a real one, and why asking the model whether it is sure is not a verification strategy.
Hallucinations have recognisable production shapes: sources and links that were never published, outdated facts presented as current, plausible API calls against methods that do not exist, and precise-sounding numbers invented where the real figure was not retrieved. Each shape has a matching control, and the common thread is that the model's output is treated as a draft claim until something outside the model confirms it.
The mitigation stack is layered. Retrieval grounding makes sources the authority and attaches citations a reader can click. Constrained output schemas eliminate malformed structure. A verification pass re-reads each source and confirms it supports the claim attributed to it, deleting what it cannot support. Empty retrieval triggers an honest "not found" rather than an invitation to improvise. And consequential outputs pass a human gate. None of this eliminates hallucination; together it makes hallucination visible and rarely shipped.
How much the problem matters depends on the surface the output reaches. A hallucinated phrase in an internal draft is an annoyance caught at review; the same error in a customer-facing answer is a trust incident, and inside an autonomous agent it becomes an action taken on a false premise. That gradient is why verification effort is allocated by consequence rather than uniformly — and the cheapest control is often deciding which surfaces may receive unverified model output at all.
Hallucination rates are measurable, and they should be measured on your own distribution rather than trusted from a vendor benchmark. A harness of questions with known answers, run against grounded and ungrounded configurations of your system, shows precisely what the retrieval and verification layers buy and where the remaining failures concentrate. Teams that skip this measurement end up arguing about hallucination from anecdote, which produces either complacency or panic and never a fix.
Codazz builds this in production — RAG Development.