Review is not a sign that the AI failed. Review is how human judgement remains connected to the work.
Separate three kinds of statement
| Kind | Meaning | Example |
|---|---|---|
| Fact | A claim that should be supported by evidence | "Revenue fell 8% in Q2" |
| Inference | A conclusion drawn from one or more facts | "The decline is concentrated in one segment" |
| Recommendation | A proposed action based on facts and judgement | "Interview ten churned customers" |
The further a statement moves from fact to recommendation, the more human judgement it needs.
A three-step review
01
Check shape
Did it follow the task?
02
Check evidence
Do important claims match sources?
03
Check consequence
What happens if we use it?
- Check shape. Did the AI answer the actual question and use the requested format?
- Check evidence. Do important claims match the cited or supplied source?
- Check consequence. What could happen if the output is wrong?
Low-impact drafts need a light review. Customer messages, financial decisions and system changes need a stronger review.
Hallucination in plain language
A hallucination is information the AI presents as if it were supported, even though the available evidence does not support it. It can be a fabricated fact, an incorrect citation or a detail filled into a gap.
Approve, correct or escalate
- Approve when the output follows the task, important claims are supported and the consequence is acceptable.
- Correct when the task is sound but the output needs a contained change.
- Escalate when evidence conflicts, permissions are unclear or the consequence exceeds your authority.
Recap
- Distinguish facts, inferences and recommendations.
- Check the output shape, evidence and consequence.
- Treat unsupported information as a signal to investigate.
- Escalation is the correct outcome when authority or evidence is unclear.