LLMs in the Clinic: Should Doctors Use AI to Treat Patients?

Every time a physician sees a patient, they spend time writing notes. Not the interesting clinical reasoning part. The documentation part. Recording what the patient said, what the exam showed, what was ordered, and why. Studies consistently find that physicians spend between one and two hours on documentation for every hour they spend with patients. That ratio has gotten worse as electronic health records have become more complex.

Large language models are, at their core, very good at turning spoken language into structured text. The connection to medical documentation was obvious from the moment GPT-3 was released, and several companies have built products around it. Ambient clinical intelligence systems use microphones placed in exam rooms to capture the patient encounter and automatically generate clinical notes. Microsoft, through its Nuance subsidiary, offers one. Suki AI offers another. Epic, the dominant electronic health record company, has integrated its own AI-assisted note generation tools.

Physicians who use these systems report meaningful time savings. A family medicine doctor who previously spent ninety minutes on notes after a full day of patients might spend thirty minutes reviewing and approving AI-generated drafts instead. That time goes back to patients, to thinking, or to rest.

But the use of large language models in medicine goes further than documentation, and that is where the stakes get considerably higher.

Med-PaLM 2, developed by Google, was designed specifically for medical question answering. When tested on the United States Medical Licensing Examination, it scored at or above the passing threshold. When evaluated by physicians on the quality and safety of its responses to medical questions, it rated at expert level on several benchmarks. OpenAI has reported similar results for GPT-4 on medical licensing exams.

What this has led to is a genuine debate about whether AI should be used for clinical decision support, helping physicians make diagnoses or treatment decisions rather than just writing their notes. Some hospitals have begun piloting tools that surface relevant clinical evidence during a patient encounter, flagging potential drug interactions or suggesting differential diagnoses the physician might not have considered.

The risks are real. Language models can be confidently wrong. They can produce responses that are fluent, plausible, and incorrect in ways that a physician without deep expertise in a specific area might not catch. The term hallucination describes when these models produce fabricated information presented as fact, and that is a particularly concerning failure mode in a medical context.

I do not think the serious question is "AI or no AI" anymore. Doctors are already using these tools in some form, even if the hospital policy has not caught up. The real question is where the line gets drawn: what can the model draft, what can it suggest, and what still has to clearly belong to the physician.