Summary. I am cautiously optimistic about generative AI for the patient-facing side of clinical communication — translating a discharge plan into the patient’s language, expanding a medication list into a usable take-home document, answering specific questions about a diagnosis at a sixth-grade reading level. I am cautious about generative AI on the clinician-facing side, where it is being asked to do the thinking I am paid to do. The two use cases get conflated by vendors and they should not be.
A fuller treatment will appear here, organized around specific use cases I have evaluated:
- Discharge instruction translation and reading-level adaptation. Promising; needs validation in real outpatient follow-up outcomes, not just readability scores.
- Patient-message-portal drafted replies. Mixed. The drafts are convincing-sounding and often clinically wrong. The cognitive load of editing them is comparable to writing from scratch, and the cognitive risk of accepting them uncritically is real.
- Patient-facing condition explanation. This is where I think the technology has the most upside, particularly when paired with validated visual content. See AI as teaching, not transcription.
- Clinician-facing differential generation and assessment drafting. I am not using these. The hedging is the cognitive work, and outsourcing it produces uniformly worse notes.
A more developed version of this case study is in progress.
— Jeremy Tabernero, MD · More case studies · Get in touch