The Hallucination Checklist

Six questions to ask before you trust an AI output.

Purpose

Six questions to ask before you trust an AI output.

AI does not know what it does not know. It generates text based on patterns — not on verified facts — and it does so with the same confident tone whether the answer is accurate or entirely fabricated. That is what hallucination means in practice: not fiction presented as fiction, but fiction presented as fact.

This checklist is a six-question filter. Run it before you use any AI output for something that matters.

How to Do It

The Checklist
If yes — verify at least two independently. If no — treat the output as unverified, regardless of how confident it sounds.
Names attached to quotes, roles or events are among the most frequently hallucinated elements. Check each one against an independent source.
Statistics and dates are the highest-risk hallucination zone. A number that sounds precise is not necessarily correct. Always verify.
If yes: does it match what you know? If no: you are not in a position to spot errors — which means external verification is essential, not optional.
Confident tone is not a signal of accuracy. If the answer feels too clean or too complete, ask the model directly: ‘How confident are you in this, and why?’
If yes: do not use AI output without independent verification. Full stop. The stakes determine the standard of verification, not the quality of the output.
KEY INSIGHT
The model doesn’t know what it doesn’t know. You have to be the one who checks.
When to Use This Checklist
Before including AI-generated content in a report, presentation or publication
When AI provides specific facts, figures or quotes you are about to pass on
Before making a decision based on AI-generated analysis
When verifying AI output in a field you are not expert in
Any time the consequences of a wrong answer are more than trivial

Reflection Questions

Which of the six checks did you almost skip — and why?
Was the output accurate? If not, which check would have caught the error?
How does this change the way you use AI output in your work?

Download

Download resource