AI hallucination is usually discussed as a technical curiosity. In business workflows it is something else entirely: a compliance risk.

I've been writing about practical AI adoption for office teams, and the most common failure I see isn't the model being "dumb" — it's people trusting outputs that look authoritative but were never grounded in anything. The fabricated legal citation with a plausible article number. The "2024 market data" that doesn't exist. A confidence level that never drops.

Here's the core insight: hallucination is a property of probabilistic generation, not a bug that the next model update will fix. The model's objective is "produce the most plausible next token," not "produce the true one." Once you accept that, the right question changes from "how do I stop it" to "how do I catch it."

So instead of treating this theoretically, here are five defenses you can actually put into a prompt template today, roughly ordered by return on effort.

1. Restrict the source (highest ROI, one sentence)

The single most effective instruction you can add:

Answer only using the following material: """[paste material]"""
If the information is not in the material, reply "not covered in the material." Do not speculate, do not supplement, do not use your existing knowledge.

This trades "intelligence" for traceability. In enterprise contexts that trade is almost always worth it — a boring, verifiable answer beats a brilliant, fabricated one.

2. Require inline citations

Force every conclusion to carry its source:

After each conclusion, add a source note in parentheses, e.g. "(source: page 3 of the attached table)".

Hallucinations hate receipts. When the model has to name where each claim came from, fabricated claims either get attributed to nothing (and become visible) or get attributed to a page that doesn't say that (and become checkable).

3. Cross-verify on the dimensions that hurt

You don't need to verify everything — you need to verify the claims whose failure costs money:

  • Numbers: re-ask the model to extract them independently, compare the two runs. Mismatch = human check.
  • Legal citations / standards / regulations: verify in the original database, always. This is non-negotiable.
  • Named sources (people, companies, papers): search for at least one independent confirmation.

4. Constrain the output structure

Open-ended answers give hallucination room to hide. Structured output doesn't fix accuracy by itself, but it makes anomalies visible: a template with "basis / confidence / missing info" fields makes it obvious when a confident claim has no basis behind it. Asking the model to explicitly list what information is missing also reduces the pressure to invent — you've given it an escape hatch other than fabrication.

5. Keep a human review checklist

The last line of defense is procedural, not technical. For any output that will be sent outside the company: every number traced to a source, every citation opened once, every named entity spot-checked. It takes two minutes and it's the difference between "AI-assisted" and "AI-responsible" when something goes wrong.

Putting it together

None of these defenses are clever. That's the point — the fix for hallucination isn't a better model, it's a better pipeline: constrain sources, demand receipts, verify what's expensive, structure the output, keep a human checklist.

I wrote a longer Chinese-language version of this playbook (with copy-paste prompt templates and a five-step flow diagram) on my site: AI幻觉识别与规避:五道防线让输出可信 — it's part of an ongoing series on practical AI adoption for office teams.

If you have a hallucination war story from a real workflow — especially one where a fabricated detail almost made it into production — I'd genuinely like to read it in the comments.