Incident Response
An AI incident is an event where the system causes or could cause harm, violates policy, exposes data, behaves outside its intended scope, or loses critical reliability. Incident response is the organized process for detecting, containing, communicating, fixing, and learning from that event.
Transparent AI makes incident response faster because responders can see what changed, who was affected, which decisions were made, and which controls failed.
Respond to an Unsafe Chatbot Incident
Step through detect, contain, communicate, and learn. At each stage, choose the response — then see its consequence.
A monitoring alert shows a spike in harmful-output reports from the support chatbot.
Sound response choices
0 / 0 answered
A response plan should define severity levels, escalation contacts, containment options, communication templates, evidence preservation, root-cause analysis, remediation owners, and post-incident tests. For high-impact systems, teams should rehearse these steps before launch.
Containment might mean disabling a feature, rolling back a model, changing a threshold, removing a retrieval source, requiring human review, or pausing use in a risky workflow.
Do Not Wait for Certainty
Early incident response often starts with incomplete facts. The first goal is to reduce harm while preserving evidence, then refine the diagnosis.
What should a postmortem add to the system after an AI incident?