Accountability

When an AI system harms someone, "the model did it" is an evasion. (If you think that's a joke, review the Air Canada case, where a chatbot gave a user a discount that didn't exist. When the user complained that it did not work and made it a legal battle, Air Canada's response was that the chatbot is responsible for its own decisions! The judge did not buy that argument, and neither should you!).

A model cannot approve its own purpose, choose its deployment context, design the interface, or decide what happens after a complaint. People and organizations make those choices.

Accountability asks a practical set of questions. Who owns the decision? Who can inspect it? What evidence is preserved? Who can override it? How does a person appeal? What changes after a failure?

Consider a customer-service chatbot that invents a discount policy. The failure may involve an outdated source document, weak retrieval, a model that fills gaps, an interface that presents uncertain text as policy, and an escalation path that is difficult to find. The final sentence came from the model. The incident came from the system.

Explainability supports accountability when it produces evidence tied to action. A feature-attribution chart may help an investigator see why a claim was flagged. It does not identify who approved the threshold, whether the user saw an appeal option, or which model version was running. Those answers require transparency and governance.

Good accountability changes future behavior. The team corrects the source, tests similar edge cases, changes the interface, updates monitoring, and records who owns each fix. An incident report that ends with "AI error" has explained very little.