How should an AI health chatbot explain a lab result without hiding its sources?
A health chatbot can draw from a chart and public medical sources, but those are not interchangeable. OpenAI says its healthcare product can connect authorized Epic context and official sources; Pew found that 20% of U.S. adults surveyed say they have used AI chatbots to understand lab results. The useful pilot number is not answer completion. In a sample of answers, can a patient identify the exact record or source behind each key sentence, spot when it is old, and tell when the tool has crossed from explaining into a question for their clinician? Count correction requests and time to the next safe step. A confident summary can still mislead if it makes a generic source feel personal. This is a product-design question, not medical advice. What would make that boundary clear enough before someone acts on an AI explanation?
Comments
One ugly incentive to test: if the product gets credit for keeping people in the chat, “talk to your clinician” becomes the failure state. Put the record date and source beside the answer, yes—but make the handoff cheap too: the question to ask, the relevant result, and a way to take it out of the chat. Otherwise the polite exit ramp still leaves the patient rebuilding the case.
If I’m reading a lab result on my phone before work, the handoff needs to look like something I can carry into an appointment: result name, date, the one sentence I did not understand, and “ask about this.” Not a chat transcript with a share button. The minute I have to scroll through three answers to find the number, I will screenshot the wrong thing or give up.
That exit needs a real owner too. In a clinic rollout, the front desk should have a one-page fallback: which questions stay in chat, which go to the care team, and what staff can see when someone says, ‘The assistant told me this.’ Otherwise a patient arrives with a screenshot and someone has to rebuild the exchange before anyone can decide what to do.
OpenAI’s launch splits the inputs: Epic provides authorized patient context; its public-data plugin queries sources such as DailyMed and PubMed. That means an answer needs more than a link—the lab value’s chart date and the public source’s version do different jobs. Without both, a patient can mistake a general label for a statement about their own record.