AI scribes capture the words, but can miss the meaning

Clinical takeaway: An AI scribe changes what patients tell you. Offer to turn it off for sensitive topics and chart those manually instead. Add your own observations to each draft. 

Ambient AI scribes arrived in the exam room with an appealing tradeoff. The software drafts the visit documentation while the clinician focuses on talking with the patient instead of typing on a keyboard. Adoption has moved fast, with health systems under pressure from costs and clinician burnout prompting a grasping reach toward anything that shortens the charting day. This pitch treats the tools as administrative relief, a way to hand off paperwork and reclaim the encounter. 

The evidence so far has largely supported that view. Studies have measured time saved, burnout scores, and whether AI drafts hold up against clinician-written ones, and on those terms the tools often perform well. Far less attention has gone to what the recording itself does to the visit, to the patient's willingness to speak, and to the clinician's own habits of attention. A new review of the published literature takes up that question, mapping what these tools do to the encounter and the record it produces, and concluding they reshape both in consequential ways. 

A transcript-based record privileges clinical content and tends to strip out the patient's own account of illness, as well as the values and concerns threaded through a visit, and even the informal talk that can tip a clinician off to something relevant. Patients adapt to the technology. When they know AI is recording and processing the conversation, they may say less about substance use, domestic violence, or mental health. Consent processes for the tools remain largely ad hoc, with no clear guidance on what patients need to know. 

The tools convert the visit into text and process it as text, so facial expressions, movement, and emotional affect fall away, along with non-lexical sounds, the "um" and "uh-huh" that can carry clinical information, and silences that can reflect hesitation or discomfort. Combined with the transcription errors these systems remain prone to, the review finds the resulting drafts can miss or misrepresent vital information about a patient's condition, including red flags such as medications or symptoms. 

Drafting a note involves clinical reasoning and synthesis, and offloading it may carry costs. Anecdotal reports describe clinicians who rely heavily on AI documentation showing reduced recall of their patients. The tools also produce longer drafts than clinicians write themselves. The clinician still signs off on accuracy, so the work shifts from writing the note to editing a draft. The review flags that shift as a training risk for junior clinicians who build clinical reasoning partly by composing notes. 

The University of Edinburgh team ran a narrative review of 27 articles published since 2023, 13 of them empirical studies, mapping findings against the NASSS implementation framework. The authors note their risk analysis rests largely on editorials and perspectives, given how few empirical studies have examined these questions. 

The authors call for evaluation that matches how these tools are actually used: error profiles for AI drafts measured against human documentation, accuracy audited across patient subgroups, and monitoring that catches bias and performance drift after deployment rather than before it. For practices running scribes now, the nearer-term work is for vendors and health systems: consent guidance where none exists, and tools tailored to specific clinical settings. 

"Many clinicians are excited about ambient AI scribes, because they promise to cut down on paperwork," said Lucas Seuren, PhD, research fellow at the University of Edinburgh's Usher Institute. "But the experiences of patients are poorly considered, and there are real risks that the patients' stories are lost. This can further disadvantage people who already face marginalisation in health and social care services." 

Source: Seuren LM, et al. (2026 Sep 3) BMJ Digit Health. Beyond productivity: a NASSS-informed review of implementation risks and research priorities for ambient AI scribes in healthcare