Cardiovascular Disease Prevention Recommendations From an Online Chat-Based AI Model—Reply

In Reply In response to the Letter by Drs Gellert and Jaszczak about our Research Letter, we agree that current application of language learning models in medicine will require close clinician and regulatory oversight. Indeed, our study findings highlight key limitations of language learning models. For example, we noted that the model provided inaccurate information about the availability of inclisiran for lipid lowering and inappropriate information about the interpretation of a low-density lipoprotein cholesterol level greater than 200 mg/dL. A false response, sometimes referred to as a “hallucination,” is a potential limitation of current language learning models as well. Such inaccuracies can have critical consequences in clinical settings, and our study points to the importance of strong guardrails for AI-assisted clinical workflows in medicine. Looking ahead to future developments beyond the current generation of language learning models, more sophisticated, theoretical, and multimodal AI models that are flexible across medical tasks have been proposed.