AI beats doctors at making 'much simpler, more concise' posthospital care plans, Stony Brook study finds

Before a patient leaves the hospital, doctors typically write a report for the posthospital care team that includes critical information about their case: why they were hospitalized, how they were treated, what condition they were in upon leaving.

Those reports, called discharge summaries, might be just a few pages or several times that for longer hospitalizations or more complicated cases. They might take 15 minutes or more than an hour to write. In an era of specialized care, the reports have become an important means of communication between inpatient and outpatient clinicians.

Now, a study by Stony Brook University researchers focused on longer hospitalizations suggests the artificial intelligence systems known as large language models generate better summaries than doctors do, making fewer errors or omissions while reducing the human workload.

In the study, published online in October in the journal Digital Medicine, physician reviewers comparing AI-generated and human-written summaries for 60 cases preferred the AI summaries 95% of the time, judging them easier to understand, communicate to patients and caregivers, and use for providing follow-up care. The researchers used GPT-5.2, an OpenAI model the company said was state-of-the-art when it was released in December 2025. Their study was retrospective, using existing clinical data, and researchers said they took special security measures to safeguard patient data.

"Because this is a transition time when [patients] are vulnerable, it’s a really big deal to get communication right, especially when it’s complicated and long and a lot of that information can get lost," said Dr. Rachel Wong, a corresponding author of the study and a clinical associate professor in the departments of internal medicine and biomedical informatics at Stony Brook’s Renaissance School of Medicine. The AI summarizes "in a much simpler, more concise, clear way that won’t miss things," she said. It also saves time: "If you have weeks or months of reports to go through, this could take hours," Wong said. "This takes you away from the patients, calling their families, communicating with them."

No significant difference

Reviewers judged the AI made fewer "clinically relevant" omissions than the doctors did, but there were no significant differences in estimated harm potential or likelihood compared with the doctor-written summaries.

Proponents of AI in healthcare have touted the technology as transformative for diagnostics, drug discovery and other areas. The study’s authors made a more modest claim: Their findings, they wrote, "support the potential for thoughtfully implemented LLM-assisted summarization to reduce cognitive burden and better support transitions of care, while maintaining the need for clinician oversight and verification."

Dr. Daniel Morgan, a professor of epidemiology and medicine at the University of Maryland School of Medicine, who has written about AI in medicine but was not involved in the Stony Brook paper, called it "a good study, rigorously done," of a timely subject, with commercial applications possible in the next six to 18 months.

Doctors now spend about two or three times as much of their time in front of a computer screen as they do talking to patients, so "this is a place where AI has great potential use," Morgan said. "This is different from AI taking over medicine from doctors. This is replacing administrative tasks that doctors don’t want to do and that we don’t do all that well."

Stony Brook computer scientist Fusheng Wang, one of the study’s corresponding authors, said hospitals would likely do more research, including pilot programs, before widely adopting AI. "The technological side is not a problem anymore," said Wang, a professor in the biomedical informatics and computer science departments. "The challenges are more on the implementation side": developing protocols for how the models are used and ensuring that the models are maintained and evaluated over time, he said.

Still, humans, not machines, will be responsible for diagnoses, Wang said. "They are in charge. ... Human doctors have to be in the loop."

By Nicholas Spangler

nicholas.spangler@newsday.comspanglernewsday

Nicholas Spangler is a general assignment reporter and has worked at Newsday since 2010.