Review finds AI medical scribes miss non-verbal clinical cues

A review of published research found that “ambient” AI tools used by doctors to generate clinical notes can miss potentially vital information conveyed through non-verbal cues, according to The Independent. Researchers at the University of Edinburgh examined 27 papers and concluded that systems built around speech recognition and generative AI largely treat the consultation as an audio-to-text conversion problem. The result is a medical record that captures words but can lose meaning carried in gestures, facial expressions, movement, and emotional affect.

The pitch for AI scribes is straightforward: less paperwork, more time for patients. That enthusiasm has translated into rapid uptake. The Independent cites an estimate that up to four in ten UK general practitioners are using AI scribes, a striking adoption rate for a tool that changes how the clinical encounter is documented. Yet the Edinburgh researchers argue that most studies have focused on productivity and technical performance rather than what happens to the patient’s story once it is filtered into text.

The limitation is not subtle. If the system “listens only to audio,” it cannot directly record a patient’s physical discomfort, a hesitant nod, or a clinician’s visual assessment—signals that often shape diagnosis and risk assessment. The researchers describe a “written language bias,” where the consultation is converted into a textual artifact and then treated as if the text is the encounter. That can reshape behavior on both sides: patients may adapt to the presence of AI in the room, while distrust may reduce disclosure. When documentation becomes the product of a tool optimized for transcription, clinicians may end up spending their regained time correcting omissions or negotiating what counts as recordable reality.

Implementation also changes work organization. Ambient scribes are not just software upgrades; they can alter documentation practices, who is responsible for what, and how care is coordinated across a health system. The review, published in BMJ Digital Health and AI, calls for a stronger evidence base before pushing large-scale deployment. It also warns that losing nuance in patient narratives can further disadvantage people who already face marginalization in health and social care services, where being understood often depends on context rather than perfect phrasing.

In practice, the promise is a quieter clinic and cleaner notes. The risk is that the record becomes tidier precisely because it excludes what the microphone cannot hear.