AI Scribes for Clinical Practice

AI scribes have moved from novelty to near-routine in many clinical settings. For most clinicians, that move has happened faster than their training in how to use them safely. Here we pick out a handful of highlights from a free full-length webinar on the topic.

What does an AI scribe actually do?

The AI scribe listens to the conversation between you and your patient, converts the audio into a text transcript using automated speech recognition, and then passes that transcript through a large language model to generate a structured clinical note for you to review.

How accurate are AI-generated notes, really?

If the patient visit is routine and well structured, accuracy can be high. In acute or complex scenarios where problems unfold by the minute or decisions shift in real time, studies report accuracy dropping to around 65%, meaning up to a third of the note may need correction. Common problems range from AI inventing information (hallucination) or omitting information. This presents as potentially made up drug names, weird looking vital signs or key information missing such as a patient being allergic to a medication. 

What is the “hidden transcript” — and why does it matter?

Most clinicians focus on the finished note and never think about what sits between the recording and the output. However, every AI scribe encounter generates an intermediate transcript. This data is stored for up to 30 days or longer, and not always subject to the same privacy regulations as your institution. Many clinicians are not aware the transcript exists at all, let alone who holds it or what it can be used for. This has real implications for patient privacy, medico-legal accountability, and informed consent.

What does proper patient consent actually require?

Ideally the practitioner should inform their patients that the conversation is being recorded, who stores it, where and how long it’s kept, and that they can say no without it affecting their care. Recording a clinical conversation also falls under wiretapping laws in many states and countries, although laws vary. The real difficulty is that many clinicians don’t actually know where their patient’s data goes, which makes that conversation hard to have honestly. 

Can AI scribes carry bias into clinical notes?

Yes. AI scribes are trained on large collections of real clinical documentation, including notes written years ago when patients rarely had access to their medical records. Because the AI learns patterns from those notes, it may sometimes reproduce language that reflects outdated attitudes or assumptions, such as terms like “non-compliant” or “drug-seeking.” While the technology can save time, it cannot judge whether the language it generates is fair, objective, or appropriate. That’s why it’s important to review every AI-generated note carefully and remove any wording that could introduce bias. 

What can you do during the consultation to get a better note?

Tell the AI exactly what you want documented. As you examine the patient, say your findings out loud (for example, “lungs clear bilaterally”) and clearly state important negatives (such as “no chest pain” or “no haemoptysis”). When you move between parts of the visit, use simple verbal cues like “my assessment is…” and “my plan is…”. The AI can organise and format the note, but it can only work with the information you provide. Be specific and avoid vague statements like “the exam was normal,” as this may cause the AI to fill in details that were never actually assessed. 

Who is responsible if the note contains an error?

You are. Vendor contracts make this explicit. This becomes especially important as you grow more familiar with the technology. When previous notes have been accurate, it can be tempting to trust the AI without reviewing its output carefully, often called “automation complacency”. However, even small mistakes or missing details, such as an important sign or symptom, can have serious consequences for patient care. For now, the primary safeguard is to review the note immediately after the consultation, while the details of the encounter are still fresh in your mind. 

References

• Verma R. AI Scribes for Clinical Practice. Medmastery Free Webinar
• ChatGPT Essentials for Clinicians. Medmastery
• Guilleminot S. AI in Healthcare. LITFL