Direct answer
AI should say it does not know when the source is missing, stale, ambiguous, contradicted or too weak to support the requested claim. A useful refusal is not failure; it protects trust by separating known facts from guesses and next-step research.
Why this matters
This turns a safety behaviour into a positive operating rule. It also helps shape prompts and review checklists for client workflows.
How to use it
- Check whether the source directly supports the answer.
- Check whether the fact may have changed.
- Check whether the domain is high-risk.
- Answer with confidence only where evidence is adequate.
- Label the missing piece and state what would verify it.
Example
If a client asks whether a Google schema feature still produces rich results, the AI should verify current Google documentation instead of relying on old SEO advice.
Caveats and limits
Saying “I don’t know” should not become laziness. The useful pattern is: what is known, what is missing, why it matters, and what would confirm it.