An LLM hallucinated a court case that didn't exist.

A lawyer cited it in a filing. The judge was not amused.

This actually happened. And it's the clearest illustration of why hallucination mitigation isn't optional — it's critical infrastructure.

After building dozens of LLM applications, here's what actually reduces hallucinations in practice:

RAG with strict source attribution. The model can ONLY use information from retrieved documents. If it can't find it in the context, it says "I don't know."

Output validation against known facts. For structured outputs, validate every field against your database. For factual claims, cross-reference with a knowledge base.

Confidence calibration. Ask the model to rate its confidence. Low confidence responses get routed to human review.

Domain-specific guardrails. In legal applications, never generate case citations — only retrieve real ones. In medical applications, always include the source study.

And the simplest, most effective technique: just tell the model "if you're not sure, say you're not sure." It works better than you'd expect.

Zero hallucination is probably impossible. But reducing hallucination rate from 15% to under 1% is achievable engineering. And that difference is the line between a toy and a product.