Ambient scribes in general practice - help or hindrance?

Advocates for AI scribes project the efficiency arguments for individual consultations to the whole system, with hopes that small-scale improvements will compound at scale. However, the introduction of such technologies occurs within a digital ecosystem already characterised by multiple platforms, portals, alerts, dashboards, templates, and icons. Each of these products was also introduced with the same belief that they would achieve significant system efficiencies. While some tools have aided administrative and communication tasks, workflow distribution processes, and overall system integration, the predicted benefits have also frequently failed to materialise. The proliferation of technologies has contributed to a hugely pressured environment. Integrating and using new digital technologies demands great cognitive, emotional, or organisational effort. When this becomes overwhelming for staff, they suffer ‘technostress’ and a sense of diminished wellbeing. 5 , 29 AI scribes risk exacerbating this pressure by diverting attention from relational interactions to ensuring the system has not misunderstood, omitted, or over-recorded information. 19 Moreover, the erosion of professional ownership and meaning introduced through digital systems can generate emotional and professional strain if tools are poorly aligned with how clinicians conceptualise good practice. Clinicians who value relational practice may find this aligns poorly with the patterns of care encouraged by AI models, resulting in reduced professional satisfaction. In an already crowded digital environment AI scribes are at risk of functioning less as a liberating innovation and more as another layer of technological demand that exacerbates fatigue and inequities, undermines autonomy, and subtly reshapes what it means to do clinical work. We provide a fictional illustrative vignette in Box 1 , based on the experiences of clinical authors. My afternoon clinic is running late as usual. The next patient, Sara, has booked about ‘ongoing pelvic pain’. I’ve met Sara several times but can’t remember seeing her for pain. Her record includes codes for ‘domestic violence’ and ‘type 2 diabetes’. I remember she’s a refugee and attends with an advocate and interpreter, usually her sister, who is also registered at our practice. She’s been seen twice in the last month for abdomino-pelvic pain. My colleagues have captured the encounters with an AI scribe. There’s a page of dense text before me. I’m behind already so just call her in. I’ll have to make my own assessment. Sara enters with her sister. I ask if she is happy with me using an AI scribe. She nods. I start it. Her sister interjects — she’s bringing her 15-year-old daughter to see me next week about her acne, but her mum is worried she’s self-harming. Could I please talk to her about it — oh and don’t mention she’s said anything. Sara describes her pain. I suggest it sounds like possible endometriosis or fibroids. I ask about pain during sex. She’s distressed and embarrassed. I ask to examine her, but she declines. She tells me my colleagues have done this — why am I not aware? She asks what happened about the tests they said they’d arrange. I can’t find this in the dense page of text, so frustratingly explain we may have to do them again. I suggest a scan, blood tests, and the contraceptive pill. Sara declines the last but agrees to trial tranexamic acid. She is tearful and I pass her a tissue, holding the silence as she thinks. I suggest we meet again to see if the tranexamic acid has helped. She nods. As they leave, her sister asks where the AI notes will go — will anyone see them? I review the notes. I see the information about the sister and niece is recorded on Sara’s record; ‘endometriosis’ and ‘fibroids’ have joined Sara’s list of coded problems; the consultation label is ‘abdominal pain’ with no mention of pelvic pain or menstrual symptoms; and there is a long page of dense text, which does not sound like either of us. There is no mention of the distress and embarrassment caused by intimate questions; no reference to the look of shock and discomfort evoked by my suggestion of the contraceptive pill, nor the multiple, eloquent glances exchanged between Sara and her sister or their Urdu interactions. The encounter is superficially (and incorrectly) descriptive. The dialogue about the previous test requests is prominent, as is the patient’s concern about why I don’t already know the relevant information. There is no humanity, no cultural sensitivity, and no representation of my relational knowledge of her family. Most importantly, little indicates the emotion and uncertainty of the complex interaction. I sigh and begin to rewrite — remembering belatedly to change the coded problems from the summary.

The recent Fit for the Future: 10 Year Health Plan for England promises a techno-utopian future. The NHS will be liberated from efficiency challenges, resource constraints, and the supply-and-demand gap by new, integrated technological advancements. 1 Similar convictions have inspired governments since the introduction of the computer in the 1980s, resulting in electronic health records (EHRs), algorithmic guidelines, systematised data collection, and centralised incentivisation systems. Such advances have transformed primary care. EHRs and population-level data have strengthened epidemiological understanding, supported clinical trials, and informed more targeted treatment and prevention strategies. Digital access and remote consulting have expanded routes into care and increased the speed and efficiency of communication. Technological platforms have enabled greater workload distribution, efficacy of administrative and transactional tasks, and contributed to greater activity. 2 However, there have been unintended consequences. Care has become increasingly transactional and fragmented, 3 compromising both quality of care and staff wellbeing. 4 , 5 Increasing activity has not necessarily translated into improvements in outcomes, and inequalities have been exacerbated by variable digital literacy and infrastructure. 5 Moreover, quantifiable, technological knowledge has been promoted above human instincts, experience, and values, leading to what some have called a ‘crisis of care’ within the NHS. 6 Artificial intelligence (AI) tools have the potential to intensify both the benefits and challenges of technology. One recent study of general practice patients and staff revealed a belief that AI could reduce clinician workload, speed up patient access, and improve safety by supporting triage, prioritisation, and decision making. 7 However, participants also highlighted the risk of depersonalising care, raised concerns about safety in complex cases and the reliability of patient data entry, and emphasised the need for strong clinical oversight to maintain trust and prevent errors. Here we consider some of the major arguments made for ‘AI scribes’ in general practice. These use ambient intelligence to transcribe and summarise consultations, generating structured notes and inputting information into EHRs. It is argued they will improve efficiency, increase the accuracy of notes, support patient participation in their health records, and yield system-wide benefits. While acknowledging this potential — and the evolutionary nature of such technologies, which will inevitably result in future improvements — we argue that the challenges, risks, and losses arising from their introduction also need careful consideration.

While we acknowledge the potential for AI scribes in general practice, we temper this with the practical, layered, and value-laden considerations described above. First, efficiency savings in information recording are countered by the failed efficiencies built into in an ever-increasing transition towards transactional, distributed, and fragmented care. Second, the data-sharing benefits of enhanced classification and codification are questionable in the context of general practice, where the emphasis is often on symptomatic presentation, situational circumstance, or biomedical diagnosis. Third, we are concerned about the risks posed by data sharing and governance, which present real dangers to patients. Last, we worry that commercial interests will drive diminished relational care and professional values, imparting pressures on the wellbeing of individual practitioners and teams. We would appeal to developers and innovators to work with practitioners and researchers to ensure such technologies are co-designed, produced, and implemented with appropriate governance around safe and effective application. We hope this would enable a positive future for such technologies within general practice encounters.