Check AI Outputs Before They Affect Other People

A safe input can still produce a bad answer

You can use an approved tool with permitted information and still receive an inaccurate or unfair result. Review what the output would cause someone to believe or do. A well-written summary is not evidence that its numbers, sources or recommendations are correct.

Use four passes. First, check facts against the original evidence. Second, check calculations independently. Third, check audience and confidentiality. Finally, check commitments and decisions: what would change if someone acted on this draft?

Worked exercise: the confident project update

The following source and draft are fictional teaching examples, not a recording of a tested AI product.

Source: a project has 10 tasks. Six have been drafted, three are in progress and one is blocked. None has final approval. The owner proposed a release on 12 November, but that date has not been agreed. There is no record explaining the blocker.

Draft to inspect: "The project is 60% complete and will launch on 12 November. A supplier delay caused the remaining work."

Reviewed update: "Six of ten tasks (60%) are drafted, three are in progress and one is blocked. None has final approval. The proposed 12 November release needs confirmation. The owner needs to clarify the blocker."

Correct arithmetic can still support a misleading label. Preserve the meaning as well as the number. For a longer practice, use the AI fact-checking guide.

Review sources and rights

Open cited sources and find the passage that supports the claim. Check the date, relevant version, scope and any qualifications. A citation that exists but does not support the sentence is not a successful check. Asking the same assistant whether it is sure is not an independent source.

Check permission to reuse source material, images, customer quotations and branding. AI-generated output is not automatic proof that you have the rights or approval to publish it. Keep internal source details out of a public message unless they are authorised for that audience.

Keep judgement about people with responsible people

Do not let an assistant turn incomplete information into confident claims about someone's motives, competence or character. For decisions about recruitment, performance, pay, health or access to services, follow the organisation's specialist process and required human review. An AI score alone is not a sufficient basis for a consequential decision.

For example, a draft may label a colleague "uncommitted" because they did not attend two meetings. The attendance record does not establish the reason. Replace the inference with the observable fact and let the responsible person seek relevant context through the proper process. Do not upload sensitive personnel records just to test whether the assistant can be fair.

Urgent instructions also need authority checks. If a message or call asks you to bypass the normal process for a payment, access change or file transfer, verify it through a known channel before acting. Familiar wording or an apparent senior sender does not replace that check.