AI-generated Replies Increase Physician Editing Workload
For practitioners, deploying LLM assistants into clinical messaging can trade drafting speed for increased human editing overhead, changing the operational cost of human-in-the-loop workflows. According to a Dartmouth study presented at the 2026 Annual Meeting of the Association for Computational Linguistics and reported by News-Medical, researchers developed a comparison tool and analyzed 146,000 conversations between 10,105 patients and their primary care physicians at a large rural health system. The team evaluated physician replies drafted by Claude, Gemini, ChatGPT, and smaller platforms Llama, Aloe, and Qwen, and found AI-generated answers frequently misalign with clinician-written responses, often being too long, omitting follow-up questions, or including irrelevant or inaccurate medical details. The researchers report that physicians may spend more time editing AI drafts than composing replies from scratch. "We find that AI can sound like a doctor but not think like one," said Sarah Preum, a corresponding author on the study.
Editorial analysis
For practitioners, the Dartmouth study underscores a recurring operational trade-off when adding LLM drafts to clinical workflows: models can accelerate first-draft production while increasing downstream verification and editing costs. Companies and hospitals evaluating LLMs for patient messaging should treat model output quality as an efficiency variable, not just a performance metric.
What happened - Reported facts: Per News-Medical coverage of the Dartmouth study presented at the 2026 Annual Meeting of the Association for Computational Linguistics, the research team built a tool that compares AI-generated replies to a dataset of clinician-written responses created with Dartmouth Health professionals. The team analyzed 146,000 conversations involving 10,105 patients and their primary care physicians; the study received Dartmouth Health Institutional Review Board approval and used anonymization to protect patient privacy. The researchers tested physician replies drafted by Claude, Gemini, and ChatGPT, along with smaller commercial systems Llama, Aloe, and Qwen and report frequent misalignments: drafts that are overly verbose, omit needed follow-up questions, or introduce irrelevant or inaccurate medical details. The News-Medical article quotes the study's corresponding author, Sarah Preum: "We find that AI can sound like a doctor but not think like one."
Industry-pattern observations suggest these failure modes are familiar to practitioners who have deployed LLMs in regulated or safety-critical domains. Hallucinations, verbosity, and missing clarifying questions raise verification burden and can reverse expected time savings unless systems include targeted guardrails, prompt engineering, selective retrieval, or lightweight human-review workflows.
For practitioners
the immediate indicators to monitor are edit-time per message, frequency of clinically significant inaccuracies, and rates of missing follow-ups. Where privacy permits, logging model outputs alongside final clinician edits-then measuring edit distance and error types-will make ROI calculations for assisted messaging empirical rather than anecdotal.
Key Points
- 1Large-scale evaluation found AI drafts often add editing burden, offsetting drafting speed gains for clinicians.
- 2Common failure modes are verbosity, missing follow-up questions, and insertion of irrelevant or inaccurate medical details.
- 3Practitioners should instrument edit-time, error rates, and missing-follow-up metrics before scaling LLM assistants.
Scoring Rationale
The study is notable for scale and direct measurement of clinician workload implications, making it relevant for teams deploying LLMs in healthcare. It is not a paradigm-shifting paper but provides actionable evidence that quality issues can negate expected efficiency gains.
Sources
Public references used for this report.
Practice with real Health & Insurance data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Health & Insurance problems