For family physicians like Tim Burdick, who practices at Dartmouth Health in Lebanon, the work day is long. He gets to work at 7 a.m., sees around 15 patients from 9 to 5 p.m., and leaves around 6 p.m. But his day doesn’t end there. He spends time after work responding to messages from his patients.
And there’s a lot of them, thanks to the availability of online patient portal messaging.
“That's nice because it increases the availability of doctors, family physicians, and other health care teams, but it also creates an opportunity for a lot of high volume of messages to be coming in around the clock,” Burdick said.
However, at the end of the week, that can add up to roughly 500 messages and leaves physicians like him working until 9:30 p.m.
“That's leading to burnout among family physicians and other clinicians,” Burdick said. “We need to figure out some ways to use new technology to make that more effective, more efficient for them.”
So Burdick wanted to know if an AI large language model could help. Based on a study he co-authored, the answer to that question is: maybe, with caveats.
Burdick co-authored the study with Dr. Sarah Preum. She’s the technical associate director at Dartmouth's Center for Precision Health and Artificial Intelligence.
They wanted to see if LLMs could read a patient’s anonymized messages, and using the patient’s medical history, create a rough draft of a response that a physician would then review.
They trained six LLMs on over 146,000 patient portal conversations between patients and clinicians, and then had the models draft responses. There were large commercial models, including a version of Open AI’s ChatGPT, Anthropic’s Claude 4.5 Sonnet, and Google’s Gemini 2.5 Pro. The researchers also tested three open-source models that used Llama and Qwen AI.
Ultimately, the study found LLMs still have a ways to go before they could be used to interact with real patients. When responding to real, anonymized patient messages, the six LLM models tested introduced errors and over-diagnosed to the point where some participating physicians told the researchers they didn’t want to use AI. The LLMs didn’t ask key follow-up questions, as physicians often do. However, the models used empathy. AI showed other signs of potential, too.
AI over-diagnosed and overtreated
“Our study shows that when you look at a very granular level, like looking at what the AI wrote versus what the actual clinician wrote, there is a gap,” Preum said.
When working with patients, Preum said clinicians will usually ask questions to narrow down a diagnosis, such as asking when pain started, and if a patient can rate the pain from 1-10. LLMs, Preum said, tend to go straight to diagnosing the pain’s cause.
She also said LLMs tend to overtreat as well.
“Like, ‘Oh, you should take this medication’ or like prescribe, generate a solution where clinicians are much more cautious, I’d say, before diagnosing and before treatment, suggesting a treatment,” Preum said.
Previous studies have also shown that AI has the potential to slow doctors down. Physicians may spend more time filling in gaps and editing the messages.
AI has a good bedside manner
Preum said doctors are often pressed for time, and may message patients with a focus on solving their problem rather than acknowledging that they are heard. She said unlike people, AI doesn’t get fatigued.
“I'm not like saying this to undermine clinicians because all the clinicians we interviewed, everybody agreed, ‘Yeah, I want to write something which is empathetic,’” Preum said.
And the results of the study found that AI messages can be more compassionate and empathetic than doctors.
“[Empathy’s] actually a low hanging fruit for AI or large language models,” Preum said.
The LLMs were also pretty good with helping with tasks like scheduling and refilling prescriptions. Preum said she could see them used in that capacity.
“I feel like that's an area for AI integration where AI can really help, like the workflow,” she said. “Those are the cases where something like an AI assisted tool can actually help clinicians, probably more reliably.”
What’s next
The researchers concluded that there’s still some work to be done before AI models are ready to be used in this way.
Preum has built a framework that gets LLMs to ask questions that are more personalized to patients and diagnose less.
“I would say that's actually not an easy problem to teach large language models or AI to ask good follow up questions,” Preum said. “It can definitely ask questions but it's not probably as good as clinicians, but we are currently working on addressing that problem. Can we teach it in a better way to mimic clinicians' reasoning?”
The next step for Preum and Burdick is quantifying how much time doctors spend on editing responses that AI has drafted. They’re also looking into whether modified LLMs that ask more questions can help physicians reduce their overall time spent responding to patient questions.
