Playbook
Our AI gave a harmful answer
A user, journalist or regulator reports a wrong or unsafe output.
First 24 hours
- Limit the harm: disable or restrict the feature if needed.
- Save logs.
- Tell legal, security and the sponsor.
Then
- Find the cause: data, prompt, model or process.
- Contact those affected, with legal’s advice.
- Add the case to the evaluation set.
- Restart with a gate review.
Message you can adapt
At 10:40 we found the assistant gave incorrect refund advice. We’ve turned it off for refund questions and are reviewing logs. Next update at 16:00.
Avoid
Don’t blame the model. Your organization is responsible for what it puts in front of customers.
Tools for this
Last reviewed 2026-11-02 · Practitioner review pending