AI patient portal message may increase clinicians' cognitive burden

Though AI is increasingly used to generate patient portal messages, new research shows it tends to write responses that don't necessarily align with what a clinician would write. This increases clinicians' cognitive burden, defeating one of the reported benefits of using AI to generate these notes in the first place.

The research, conducted by Dartmouth researchers and presented at the 2026 annual meeting of the Association for Computational Linguistics, aimed to compare the alignment between LLM-generated patient portal drafts and responses written by clinicians.

For the study, the researchers created a tool to compare AI-generated notes to clinician-written ones. They collected 146,000 patient-clinician conversations from a large academic hospital in the United States. The researchers included Claude, Gemini and ChatGPT as well as three smaller commercial platforms, Llama, Aloe and Qwen, in the study. They also assessed different adaptation approaches to align LLMs with clinicians.

The study shows that LLMs that were not adapted at all tended to over-treat and over-diagnose, generating more medical assessment and treatment-planning themes than clinicians did. This could include long explanations of test results, symptoms, and potential diagnoses, as well as recommendations of various forms of treatment.

Additionally, responses from expert clinicians and adapted LLMs tended to ask the patient more follow-up questions; however, unadapted LLMs rarely asked follow-up questions, even though their responses tended to be longer.

"You don't want to integrate large language models into the workflow and just shift the bottleneck so that doctors are devoting their cognitive energy to playing AI janitor and fixing mistakes," said Sarah Preum, Ph.D., the study's co-corresponding author and an assistant professor of computer science at Dartmouth, in a press release. "But if we're not careful, that's a likely outcome."

Thus, the study concluded that adapting AI to align with clinicians' communications is necessary to "increase the reliability and responsibility of LLM use for patient message response drafting."

The study showed that adapting LLMs can improve accuracy by 33% and reduce editing by 26%. A technique developed by the Dartmouth team, called Thematic Agentic Direct Preference Optimization for Learning Enhancement, or TADPOLE, was especially effective at drafting responses that better matched physicians' standards.

The researchers noted that future research needs to assess how much time clinicians actually spend editing AI message drafts and to further evaluate TADPOLE, including examining whether and how it lightens a physician's workload. 

"This is one of the first studies that uses real patient portal messages to establish a generative AI model. In that regard, it's innovative and shows us that this is not a simple task," said Tim Burdick, M.D., a study co-author, an associate professor of community and family medicine and a family medicine physician at Dartmouth. "We're still nowhere near the point of having clinicians removed from the workflow."

Anuja Vaidya has covered the healthcare industry since 2012. She currently covers healthcare IT and innovation, including artificial intelligence, digital healthcare, EHRs and interoperability.