Does an AI doctor know when not to order a test? That is one of the hardest things in medicine. Patients ask for an extra scan "just to be safe," and an unnecessary MRI often finds a harmless disc bulge that sets off more worry, more referrals and sometimes procedures nobody needed. A new study in Annals of Internal Medicine, from Ben-Gurion University and Clalit Health Services in Israel, tested six leading models from Google, Anthropic and OpenAI on hundreds of clinical cases built around Choosing Wisely, the global campaign against low-value care. Every case had a twin: the young man with back pain after lifting furniture who needs no imaging, and the woman with a history of breast cancer and new back pain at night who does. When care was needed, the models wrongly held it back in only 0.6% of cases. When it wasn't needed, they recommended it anyway in 5.4%. In eight of the fifteen recommendations they never suggested a single unnecessary test. Most of the mistakes landed where doctors themselves disagree, like whether to scan small thyroid nodules found by chance. The weak spot is social. As Guy Laban, the Ben-Gurion researcher who co-led the study, points out, a doctor has to stand firm in front of an anxious patient, while language models tend to go along with whoever is asking. I work on AI medicine at Certuma, and to me this is where AI can do the most good. Restraint saves money and it saves patients from harm. The best doctor is not the one who orders everything. It's the one who knows when to stop.