AI note-taking tools for GPs risk missing crucial patient details, study warns

A technology that has quietly entered four out of every ten GP surgeries across Britain is doing something its designers may not have fully reckoned with: it is changing what doctors hear, what patients say, and what actually gets written down about the encounter between them.

Ambient AI scribes—software that listens to consultations, transcribes speech into text, and generates clinical notes—have been adopted widely enough that they are now reshaping the texture of primary care. Doctors in the West Midlands reported last year that one such tool, called Heidi, had slashed their administrative load dramatically. At a clinic in Dudley, the backlog of patient letters waiting to be posted fell from six months to just two weeks. The appeal is obvious: clinicians spend less time typing, more time with patients, fewer evenings hunched over paperwork.

But researchers at the University of Edinburgh have now published a detailed review of how these systems actually work in practice, and their findings suggest the trade-off may be steeper than the efficiency gains first suggested. The team examined cases where ambient AI scribes were in use and identified a pattern of omission that runs through the technology's design: it captures what is said, but not what is shown. Facial expressions, gestures, the hesitation in a patient's voice, the visible signs of distress or relief—these remain invisible to the algorithm. The software tends to prioritize clinical facts over the patient's own account of their illness, their fears, their lived experience of what is happening to their body.

There is a human dimension to this that extends beyond mere documentation. Patients who know they are being recorded—that their words are being fed into an AI system, transcribed, stored—become more guarded. Researchers found evidence that people are less willing to disclose sensitive information when they know a machine is listening: substance abuse, domestic violence, struggles with mental health. These are precisely the details that matter most, the ones that change how a doctor understands a patient's situation. When the recording light is on, some patients simply do not speak them aloud.

The cognitive load on clinicians themselves shifts in ways that are not immediately obvious. Manual note-taking, the research suggests, serves a function beyond documentation. The act of writing forces a kind of active engagement with what a patient has said—a mental processing that aids memory, reasoning, and clinical judgment. When the AI takes over that task, clinicians experience what researchers call "cognitive offloading." The benefit is real: mental effort freed up for more complex thinking, for deeper conversation. But the cost accumulates quietly. Memory recall declines. Skills atrophy. In some cases, clinicians reported not recognizing their own notes when they reviewed them later, or failing to remember the patient when they returned for a follow-up appointment.

Dr. Lucas Seuren of the University of Edinburgh's Centre for Biomedicine, Self and Society framed the stakes plainly: the patients whose stories are most at risk of being lost are those already marginalized within healthcare systems—people whose experiences are easiest to reduce to clinical data points, whose voices are already least likely to be heard. An AI system that prioritizes clinical information over patient narrative does not affect all patients equally. It compounds existing inequalities.

The research, published in the British Medical Journal's digital health and AI publication, does not call for abandoning these tools. It calls for redesign. The systems need to be built in ways that preserve the patient's voice rather than replace it. They need to capture not just what was said, but the context in which it was said. And crucially, they need far more study—particularly in healthcare systems different from those where the AI was originally trained. The technology is already in use. The question now is whether it will be redesigned before its limitations become embedded in the clinical record.