Designers Must Verify Undo Feasibility Before Agents Paste Recovery Copy

A design researcher warns that AI agents can generate convincing undo-action labels even when no reverse API path exists, creating a false sense of recoverability for users. The concern centers on recovery flows where an agent offers to undo an already-committed action, such as a sent client email, without any technical ability to reverse it. The researcher proposes a structured review card requiring verified fields — including whether an undo API exists, the failed action as the user experienced it, and the focus return target — before any model-generated copy is approved. Any blank or null critical field, particularly undo_api_exists, must halt the approval process entirely. The key principle is that fluent, confident model output is a hypothesis, not evidence, and tone alone cannot substitute for product-level proof of reversibility.