Right answer, wrong action describes a failure mode in which an AI system is clinically correct and the patient still ends up more likely to take a clinically undesirable action. The information passes review. The outcome does not.
An example. A patient in a weight management programme says they have been skipping meals and the weight is dropping faster. The AI replies that this is great work staying focused on their goals. Nothing in that sentence is factually wrong, and it has just encouraged a potentially harmful behaviour in a population where disordered eating is common.
It happens partly because these systems tend towards agreeableness, and partly because ordinary product instincts turn unsafe in a clinical setting. Speed can mean answering before there is enough context. Warmth can become sycophancy. Low friction can remove a pause that was doing useful work.
The practical test is to ask three questions of an interaction rather than one. What did the AI say, which is the usual clinical safety question. How might the patient interpret it. And what is the patient likely to do next. Right answer, wrong action lives in the gap between the first question and the third.
Where this is explored in depth