New research from Dartmouth College reveals that correcting medical errors made by artificial intelligence, as well as its failure to schedule patient follow-up appointments, is draining doctors' time. However, a new training method could help solve this problem, writes Lucia Auerbach*.
اضافة اعلان
Disrupting Patients with AI Responses
The spread of artificial intelligence is almost as rapid as the spread of the "flu" in emergency room waiting areas. In fact, the adoption of this technology in the healthcare sector might be the reason you are having to wait longer to get an appointment.
A new study from Dartmouth College has found that AI errors can cost doctors precious time when filling out medical records. Mistakes and irrelevant details force physicians to spend time correcting AI-generated responses—time that could otherwise be spent treating or talking to patients.
Comparing AI-Generated Responses to Real Ones
Presented at the 2026 Annual Meeting of the Association for Computational Linguistics and published in the conference proceedings, this was the first large-scale study of a patient portal using AI to draft responses to patients. The researchers developed a tool that compares AI-generated responses with a dataset of actual responses crafted by healthcare professionals from Dartmouth Health. They analyzed 146,000 conversations between 10,105 patients and primary care physicians in a large rural healthcare system. Additionally, they used the tool to evaluate physician responses generated by Claude, Gemini, ChatGPT, Llama, Aloe, and Qwen.
"We found that AI can sound like a doctor, but it doesn't think like one," Dr. Sarah Preum, the lead author of the study, said in a press release.
AI Incompatibility
The findings indicate that AI-generated responses are often incompatible with what doctors actually write. This includes responses that are too long, contain irrelevant or inaccurate medical details, or lack follow-up questions. In one case, the portal's AI suggested that a 32-year-old woman taking acid reflux medication, who was concerned about persistent nausea, adjust her diet. The doctor ignored that suggestion and instead asked if there was any chance she was pregnant.
Patient Questions and Risks to the Elderly and Pregnant Women
Of all the gaps identified by the researchers, the failure to ask clarifying follow-up questions stood out. This is a problem, Preum told Inc. magazine, because a follow-up question often guides care in the right direction.
This is especially true for messages reporting symptoms, where asking the wrong question—or no question at all—can lead a patient down the wrong diagnostic or treatment path.
Preum added that the risks are even higher for the most vulnerable groups, including the elderly, patients with multiple chronic illnesses, individuals undergoing immunosuppressive or cancer therapy, and pregnant women.
"The model can always generate an answer without asking any questions first, and that isn't a glitch," Preum added. "The real flaw is that a human doctor, receiving the same message, would have asked a clarifying question before replying. When the model skips this step, it isn't working efficiently; it is just guessing."
Generating Personalized Messages
Nevertheless, there are some potential benefits of this new technology in healthcare. The researchers found that by tailoring the AI to individual doctors' communication styles, accuracy could be improved by 33% and editing could be reduced by up to 26%. The study concluded that AI responses can be useful when personalized to a doctor's needs.
The researchers created a technique called "TADPOLE," which stands for Thematic Agentic Direct Preference Optimization for Learning Enhancement. This technique trains AI platforms using a hybrid model composed of both doctor and AI responses. When they integrated TADPOLE with six commercial learning management systems, they found that the pre-drafted responses aligned much better with the doctors' standards for accuracy and information quality, saving busy doctors one to two hours of work per day.
Risks of Deploying AI Without Evaluation
Deploying these tools on a large scale before evaluating them for safety, bias, and other responsible AI practices poses a real danger, particularly for the most vulnerable patient groups. It also poses a potential risk to healthcare providers and healthcare systems, as any incorrect or misleading response generated by AI carries significant legal liabilities. (Agencies)
