The rule. Design the failure path with the same care as the success path: dismissal, correction, and recovery are core flows, not exception handling.
Why. AI inference runs under uncertainty and routinely produces false positives and false negatives, so AI-infused systems "may demonstrate unpredictable behaviors that can be disruptive, confusing, offensive, and even dangerous" (Amershi et al., 2019, p. 1). Being wrong is common enough that the 18 guidelines for human-AI interaction give it a phase of its own: guidelines 7 through 11 cover efficient invocation, dismissal, correction, scoping services under uncertainty, and explanation (Amershi et al., 2019). Microsoft's HAX guidance adds a cost ceiling: repeated correction can cost users more than doing the task themselves, at which point the right design is a way to turn the feature off, not a better apology (Microsoft, n.d.-b).
Seen in the wild. Copilot in Word attaches keep, adjust, regenerate, and delete controls to every generated draft; Microsoft's HAX library files it under guideline 9, support efficient correction (Microsoft, n.d.-a).