Doximity tested nine leading AI models on 8,500 dosing questions and found serious errors in roughly one answer in fifty.