Doctors preferred AI-written hospital discharge summaries 95% of the time in a Stony Brook study

A Stony Brook University study published online in npj Digital Medicine and announced on Monday, Nov. 2, 2026, compared discharge summaries written by a large language model with ones written by physicians for complex internal-medicine stays.

Discharge paperwork is where hospital care quietly falls apart, and it is written by exhausted people at the end of long stays. A 95% preference is a striking number, but it is 60 cases at one hospital, and "both had errors" is doing a lot of work in that sentence. The real test is the pilot: does it save doctors time without teaching them to skim?

A discharge summary is the handoff when a patient leaves the hospital. It records the treatment, medication changes, unresolved problems, and follow-up needs. Primary care doctors rely on it. Writing one for a long stay is a large documentation burden, the paperwork load on the people who have to produce it. Those lines are the study's and Stony Brook's.

The team used a large language model, in an approach designed with clinicians, to organize a patient's records and produce the core of a discharge summary within a few minutes. A large language model is the kind of AI that writes from patterns in text. A clinician, here, is a doctor or another licensed person on the care team. That description of the method is the researchers'. Measuring whether doctors actually save time is one of the next checks the team says it still has to do.

The study compared model-written and physician-written summaries for 60 internal-medicine hospitalizations of 7 to 21 days at Stony Brook University Hospital. Internal medicine is the hospital service for adult medical illness. Hospitalists and primary care physicians reviewed each pair. A hospitalist is a doctor who cares for patients inside the hospital. The summaries are from stays that had already happened. Those details are the paper's and the university's.

Reviewers preferred the AI-generated summary in 95% of the hospitalizations evaluated, and scored the AI versions higher on overall quality, readability, factual accuracy, and completeness. In the release's ratings table, on a scale of 1 to 5, the AI summaries versus the human ones averaged about 4.71 versus 3.16 for overall quality, 4.66 versus 3.39 for readability, 4.62 versus 3.73 for factual accuracy, and 4.69 versus 3.42 for completeness. A 5 is the top of that scale, so the AI averages sit close to the top and the human averages sit nearer the middle. Those scores are the release's.

The AI summaries had fewer missing clinical details. Primary care physicians found them easier to understand, easier to explain to patients and caregivers, and more useful for follow-up care. Those findings are the study's.

Both kinds of summary contained errors. Reviewers found no statistically significant difference in the estimated severity or the likelihood of harm from those errors. Statistically significant means a gap large enough that the study treats it as unlikely to be chance. The authors stress that a physician still has to review and verify a draft before it goes into a patient record. Those lines are the authors'.

The co-corresponding authors are Rachel Wong, MD, clinical associate professor of Internal Medicine and Biomedical Informatics at Stony Brook's Renaissance School of Medicine, and Fusheng Wang, professor of Biomedical Informatics and Computer Science. The lead author is Tyler Osborne, a PhD candidate in Computer Science. Co-corresponding means both are the authors a reader can treat as standing behind the paper. Biomedical informatics is the use of computing on medical information. Those roles are the release's.

Rachel Wong said: "These results provide evidence that artificial intelligence can support discharge documentation for complex hospital stays." The release paraphrases Fusheng Wang on the next steps: assessing accuracy, safety, actual time savings for physicians, and fit into clinical workflows, and says a pilot of the tool in everyday practice at Stony Brook Medicine is being considered. A clinical workflow is the way the work already moves through a hospital day. A pilot would be a first try in that real setting. That quotation is Wong's. The next-step lines are the release's paraphrase of Wang.

The picture is Stony Brook University's official portrait of Fusheng Wang, professor of Biomedical Informatics and Computer Science and co-corresponding author of the study. The credit is Stony Brook University.

In plain terms, for 60 long stays at one hospital, physician reviewers preferred the AI discharge draft 95% of the time and scored it higher on quality, readability, accuracy, and completeness. Both the AI drafts and the human ones still had errors, and the researchers say a doctor has to check every draft. A pilot in everyday practice is being considered.