,

AI Takes Notes: How Language Models Are Changing Clinical Letters and Documentation

Abstrakte Grafik: digitales Dokument mit Textzeilen und KI-Markierung auf hellblauem Grund

Documentation is one of the biggest burdens in doctors’ everyday work. Large language models promise relief: they listen in on the consultation, produce a structured draft note and suggest clinical letters. The first randomised trials show what these “ambient scribes” actually deliver.

How an ambient scribe works

With the patient’s consent, an app records the conversation. Speech recognition converts it into text, and a language model extracts the medically relevant information and arranges it into the usual structure: history, findings, assessment, plan. The doctor reviews the draft, corrects it and transfers it to the record. Similar tools summarise long medical histories or pre-draft discharge letters.

What a randomised trial shows

At the University of California, Los Angeles (UCLA), 238 doctors from 14 specialties were randomly assigned to one of two AI applications or to usual documentation. Around 72,000 patient encounters over roughly two months were included in the analysis. The results were published in NEJM AI in 2025:

  • With one of the two programs, the time per note fell by an average of about 41 seconds, compared with 18 seconds in the control group. With the second program, the saving was smaller and not statistically significant.
  • Scores for burnout, exhaustion and cognitive load improved moderately, with burnout about seven per cent lower than in the control group.
  • The doctors reported that the AI notes occasionally contained clinically relevant errors – mostly missing information or wrong pronouns (for example “he” instead of “she”). One minor safety event was reported.

The authors stress that the study took place at a single centre and over a short period. The time saving is real, but smaller than is often advertised.

Opportunities

  • More eye contact: doctors who are not typing at the same time can focus more on the person in front of them.
  • Less evening work: some of the documentation that would otherwise be done after clinic hours is no longer needed.
  • Better readability: structured notes and clear summaries can also help patients, for example as a letter in plain language.

Risks and obligations

  • Errors and omissions: language models can leave out statements, attribute them wrongly or add things that were never said. A doctor’s review of every draft is therefore mandatory – and takes time.
  • Automation bias: the more reliably a system usually works, the more likely it is that errors will be overlooked.
  • Data protection: recordings of consultations are health data within the meaning of the GDPR. Transparent information, consent, clear deletion periods and contracts with the provider are required. Data should be processed within the EU wherever possible.
  • Regulation: pure documentation aids are usually not considered medical devices. This can change as soon as software suggests diagnoses or treatments. In addition, AI systems that interact with people are subject to transparency obligations under the European AI Act.

What patients can do

You are entitled to ask whether a consultation is being recorded, for what purpose and how long the recording will be kept, and to refuse the recording. Refusing must not have any negative consequences for your treatment.

Conclusion

AI-assisted documentation is one of the most practical everyday applications of language models in medicine. It can lighten the workload, but it is not a sure-fire success: the benefit depends on the quality of the software, sound data protection rules and careful review.

More on this topic: AI in Medicine 2026 – an overview and Risks of AI in medicine.

Note: This article is for general information only. As of: October 2026.

Video on this article