A study found that even when prompted for accuracy, most AI models routinely oversimplified the findings of research and medicine.