Skip to content

International AI Safety, Governance & Assurance Alliance

Back to Resources

AI Governance in Practice

What to do when the AI output is wrong

6 min read · Updated 2026-11-08

Correct the thing a person already saw

If a customer, an applicant, or a colleague received the wrong output, fix that message first. Say what was wrong in plain words and give the correct information. Do this even when the mistake feels small. A quiet edit in the tool does not reach the person who already read the answer.

Keep a copy of what went out and of the correction. You will want both if the same issue comes back, or if someone asks what you did about it.

Write down the path, not the blame

Note the tool, the task, the input in general terms, and what the output got wrong. Leave out personal data you do not need for the note. Ask whether a person was supposed to check it and whether that check happened. The note is for the next decision, not for a hunt.

Patterns matter more than one bad paragraph. Three wrong dates from the same tool is a reason to narrow that tool's job. One awkward sentence may only need a clearer instruction.

Decide what changes

Choose one change and write it down: a tighter instruction, a required human check, a topic the tool must refuse, or a pause while you talk to the vendor. Tell the people who use the tool. A lesson that stays inside the incident note will be repeated by the next person in a hurry.

This is general education, not legal advice. If the mistake affected someone's rights, health, or money, get advice from someone qualified in that area before you decide the matter is closed.

This article is general education, not legal advice. ComplianceEU Alliance is an independent private initiative, not a regulator or certification body.