Hospital study finds doctors preferred AI-written discharge summaries in most cases
Doctors reviewing complex hospital discharge documents preferred drafts generated with artificial intelligence in 95% of the cases assessed in a Stony Brook University study announced on 2 November. The finding points to a possible use for language models in clinical paperwork, but the researchers warned that human verification remains necessary.
Photo credit and license
Photo: Oli1287 / Wikimedia Commons / CC BY-SA 4.0; original photo, no modifications. · Image source · License: CC BY-SA 4.0Doctors reviewing complex hospital discharge documents preferred drafts generated with artificial intelligence in 95% of the cases assessed in a Stony Brook University study announced on 2 November. The finding points to a possible use for language models in clinical paperwork, but the researchers warned that human verification remains necessary.
The team examined 60 internal-medicine hospitalizations lasting between seven and 21 days. These longer stays produce extensive records, making it harder to prepare a concise account of treatments, medications, unresolved issues and follow-up care. Hospital doctors and primary-care physicians evaluated AI-generated and human-written summaries.
According to the university, reviewers generally gave the AI drafts higher marks for completeness, readability, accuracy and overall quality. They also reported fewer missing clinical details. Primary-care doctors found the AI versions easier to explain to patients and more useful for planning follow-up care.
The 95% figure describes preferences within this particular evaluation; it does not show that AI will perform equally well across hospitals, languages or patient groups. Both human and AI summaries contained errors. The researchers reported no statistically significant difference in their estimates of the severity or likelihood of harm associated with those errors.
That qualification matters. A discharge note can affect medication use and subsequent treatment, so a polished draft is not automatically a safe final record. The authors stressed that a physician must review and verify AI-produced material before it is placed in a patient's chart.
The researchers plan further work on safety, accuracy, real-world time savings and integration into clinical workflows. A pilot in routine practice is under consideration, not a completed rollout. The present result suggests a potential aid to documentation rather than a replacement for clinical judgment.