Researchers tested 20 AI models with over 3 million queries for their susceptibility to medical misinformation.