From clinician review to measurable human oversight of AI-drafted paediatric discharge documents

When a child leaves a hospital emergency room, the family receives a document that must do two difficult things at once. It must tell the next doctor exactly what happened and what treatment was given, using precise medical language. At the same time, it must explain the same events to a worried parent in plain words, so they know how to care for the child at home and when to return if things go wrong. For decades, doctors have written these summaries by hand or by filling out fixed forms, a process that can be slow and sometimes leaves out important details or uses language that is hard to understand. Recently, a new kind of computer program, known as a large language model, has emerged. These programs can read a patient's medical record and write a draft of a discharge summary in seconds. The hope is that these tools will make the documents clearer and faster to produce. But a critical question remains: if a computer writes the first draft, does the doctor actually catch the mistakes the computer makes before the family takes the paper home?

In a hospital in Montpellier, France, researchers set out to answer this question by watching how doctors interact with these new tools in real time. They did not simply ask doctors if they liked the computer's work; they watched the entire process from the moment the computer generated a draft to the moment the doctor signed the final paper. The team recruited 153 children who were being sent home from the pediatric emergency department. They split these families into two groups. One group received the standard discharge summary the hospital had always used, assembled from a fixed template. The other group received a summary drafted by a computer program running on the hospital's own secure servers, which was then reviewed and signed by the same doctor who treated the child. The researchers preserved every version of the document: the original medical notes, the computer's first draft, and the final paper handed to the family. This allowed them to see exactly what the computer changed, what the doctor changed, and what errors might have slipped through.

The results showed that the computer was very good at making the language simpler. The documents written for families in the computer-assisted group used more common words and fewer technical terms than the standard documents. However, when the researchers checked how well the families actually understood the instructions a few days later, there was no difference between the two groups. Whether the document was written by a template or by a computer, families remembered the follow-up plan and the warning signs to watch for with similar accuracy. In fact, in both groups, nearly half of the families got the details of the follow-up plan wrong, suggesting that the problem lies not just in how the words are written, but in how the information is communicated during the busy moment of discharge.

The most revealing part of the study was what happened behind the scenes when the doctor reviewed the computer's draft. The researchers found that the computer did make mistakes, but the doctors were active and careful editors. In the group where the computer wrote the draft, the doctors changed the text in almost every single case. They removed information the computer had invented that was not in the medical record, and they addressed errors where the computer had misunderstood a medical abbreviation. For example, the computer sometimes expanded a short medical code into the wrong full term, turning a description of urinary symptoms into a completely different medical condition. The doctors corrected only a small fraction of these specific errors: they fixed 6 out of 22 incorrect abbreviation expansions, meaning 16 of these mistakes remained in the final documents handed to the families. This suggests that invented facts are easy to spot because they look strange, but subtle errors in medical terms can look perfectly normal and slip past even a careful review.

Crucially, the study found that no error in the final documents handed to the doctors was judged to be serious enough to cause harm. The computer did not invent dangerous medical advice, and the doctors caught the most critical issues. However, the researchers also discovered that the computer lost different kinds of information than the standard template did. The standard template simply failed to transfer certain fields, like past allergies, in a predictable way. The computer, on the other hand, sometimes dropped specific details about who gave a medical opinion or omitted negative findings that were recorded in the notes. These losses were not always obvious in the final text, meaning a doctor might not realize something was missing unless they compared the final paper directly with the original notes.

The study concludes that while these computer tools can help make medical documents easier to read, they do not automatically make them safer or more accurate. The human doctor remains essential, but their role has changed from writing the text to carefully checking it. The researchers argue that we cannot simply assume a doctor has caught every mistake just because they signed the paper. To truly measure how well humans oversee these tools, we need to see not just the final result, but the steps in between: what the computer wrote, what the doctor saw, and what the doctor chose to fix or leave alone. Until we can measure these steps, we cannot be sure that the safety net of human review is catching everything it needs to. The work suggests that the future of medical documentation will depend not just on better computers, but on better ways for doctors to see and correct the work those computers produce.