The conversation about AI in clinical settings has focused on accuracy — whether models diagnose correctly, whether they surface the right drug interactions. The harder question is whether practitioners can tell what the AI is doing at all. When a configured-off meeting recorder produces a post-session summary [2], the failure is not a bug in the ordinary sense. It is an illustration that the practitioner's control state and the tool's actual operational state have come apart. That gap does not require malice or negligence — it requires only an app…