The rule. Audit every AI feature twice: once against classic usability heuristics, once against AI-specific guidelines. Passing one test says nothing about passing the other.
Why. When Microsoft synthesized two decades of HCI research into 18 guidelines for human-AI interaction, the finding was not that the classics had expired. They hold, but properties unique to AI (accuracy that varies, failure modes that surprise, behavior that shifts as systems learn) raise problems the classics never had to answer (Amershi et al., 2019). Yang et al. (2020) trace the difficulty to two roots: uncertainty about what the system can actually do, and the open-ended complexity of what it produces. Nielsen (2023) argues the interaction model itself flipped: users now state the outcome they want rather than the steps to reach it, which unsettles assumptions as basic as visibility of system status and direct manipulation.
Seen in the wild. Perplexity adapts a sixty-year-old heuristic to a probabilistic system by replacing the spinner with staged progress (searching, reading sources, writing answer). Old rule, new material.