News
Dartmouth Study: Fixing AI Errors Takes Doctors Longer Than Writing From Scratch
Dartmouth College researchers found that AI-generated responses to patients on a health portal contain so many errors and unnecessary details that correcting them takes doctors longer than writing the messages themselves. The findings were presented on July 7, 2026 at the annual meeting of the Association for Computational Linguistics.
Artificial intelligence was supposed to relieve doctors of the tedious task of writing patient responses. A new study by a Dartmouth College team shows that in practice the opposite often happens: fixing errors in AI-drafted messages takes longer than writing them from scratch.
The research team analyzed an AI system used in an online patient portal that automatically drafts physicians' responses to patient messages. This is the first study of this scale to examine this specific AI application in everyday clinical practice rather than in a laboratory setting.
Where AI Goes Wrong
The researchers identified recurring error patterns in the model-generated responses. The most common issues were overly long, rambling phrasing, missing follow-up questions a doctor would normally ask, and medical details that turned out to be irrelevant, inaccurate, or inconsistent with accepted clinical practice.
The effect is paradoxical. Instead of speeding up a doctor's work, the system forces them to carefully read and correct the draft, which in many cases takes more time than formulating a response from scratch. In other words, AI can sound like a doctor, but it doesn't think like one.
We find that AI can sound like a doctor, but it doesn't think like a doctor - Sarah Preum, Assistant Professor of Computer Science, Dartmouth College
Why This Matters for Healthcare
Systems that generate draft responses for doctors are increasingly being deployed at medical facilities across the United States and Europe as a tool meant to curb staff burnout and cut down time spent on documentation. The Dartmouth findings challenge some of these assumptions, at least in the technology's current form.
For hospital administrators and healthcare facilities, this means they need to more carefully measure the real impact of such tools on doctors' working time, rather than basing purchasing decisions solely on software vendors' claims.
What It Means for Poland
In Poland, AI tools supporting medical documentation and patient communication are only just gaining traction, including through pilot programs at larger hospitals and clinic networks. The Dartmouth findings serve as a warning sign against deploying such systems without first checking how much time they actually save, versus how much they cost at the verification stage.
The authors stress that the problem does not lie in the idea of automating patient communication itself, but in the quality and reliability of the generated content. Until the models learn to recognize which information is clinically relevant in a given context, the benefits of using them may remain illusory.
What Comes Next
The Dartmouth team says it plans further work on evaluation methods that would measure not only the linguistic correctness of AI responses but also the actual time doctors need to verify and correct them. The goal is to help medical facilities make more informed decisions when choosing and deploying such systems.
Sources: AI mistakes can cost doctors time when writing to patients (eurekalert.org), Study shows how flawed AI responses increase physician workloads (news-medical.net), AI Mistakes Can Cost Doctors Time When Writing to Patients (home.dartmouth.edu)
