Artificial intelligence scribes in general practice: Sacrificing the art of medicine for promised efficiency

Mohammad S Razai is a NIHR Clinical Lecturer in Primary Care in the Department of Public Health and Primary Care (PHPC), University of Cambridge, Cambridge, UK.​

There is a growing push for the adoption of artificial intelligence (AI) in healthcare. Amidst the noise, hype, and sweeping claims of AI’s potential to ‘revolutionise’ medical practice, policymakers and politicians have embraced the narrative.¹ In the UK, the government has ‘…fire[d] the starting gun on AI Growth Zones to turbocharge Plan for Change.’¹ In general practice, AI is increasingly positioned as a remedy for tackling workload and increasing efficiency.² One example is the introduction of ambient AI scribes – software that listens to patient consultations, transcribes conversations, generates documentation, and soon, may automate tasks such as ordering tests and investigations.²

Demand for primary care continues to rise, administrative burdens on clinicians are growing, and access to services remains a perennial issue. ³ AI scribes are being marketed as a “gamechanger”, promising “ a seismic shift” to reduce documentation time and allow clinicians to focus more on “direct patient care”.² Workload in general practice usually refers to the volume and complexity of patient care, the number of patient contacts per day, short appointment times, managing repeat prescriptions, reviewing test results, and handling an ever-increasing amount of correspondence.^4,5 Would an AI scribe make a meaningful dent in any of them?

Many previous technological tools introduced to improve efficiency in general practice failed to deliver on their promises.^6,7 In some cases, they increased workload, added complexity, and fractured the continuity of care.^6,8 The current framing of AI tools reflects a managerial mindset in which healthcare is viewed as any other service industry. Operational metrics, which are helpful in banking and retail, are used to measure success. However, the work of a GP involves delivering humane, compassionate, person-centred care. It is not a production line, but rather engaging in complex, narrative-driven, emotionally textured relationships that develop over time,⁹  requiring a deep investment in what can be described as the art of medicine.

To understand how these tools function in practice, I trialled one of the new ambient AI scribes. After a consultation with a simulated patient experiencing itching, a segment of the AI note was “pruritic subcutaneous lesions… significant discomfort report.”  The actual conversation was more like: “I’ve got this thing on my arm, I don’t know, it is really itchy, it’s driving me crazy, I keep scratching.” The AI captured some of the patient’s language, but not enough to convey the tone, emotion, or meaning behind their words (though AI outputs may vary between consultations). In my documentation, I usually include the patient’s own words—terms like “itchy”, “drives me crazy”, and “scratching”. These details matter. They reflect the patient’s lived experience, vernacular, and emphasis. They shape how I understand their distress and how I respond, clinically and relationally.

When we write clinical notes, we can do more than record data. We interpret, reflect, and engage in a silent dialogue with the consultation and the patient through it. We may select what is meaningful and what must be remembered. AI, by contrast, tends to standardise language, use medical jargon, flatten nuance, ignore disagreements, and produce verbose summaries that often feel sterile. In some cases, it invents content, known as hallucinations, or makes errors. For example, attributing action points to the wrong clinician or inserting generic statements that bear no resemblance to the actual dialogue.

Upon reviewing the AI-generated note, I found it challenging to identify key clinical information. It was lengthy and impersonal, failing to evoke the person I had seen. Reading, editing, and approving the transcript required time. I could not trust the AI-generated documentation enough to accept it without review. This has broader implications. Clinical notes are not just records; they are part of the living ecology of medicine, sustaining reflection, learning, teaching, and research. When documentation becomes standardised and dispassionate, it can disrupt that ecology, stripping away the elements that make generalism rich, holistic and human.

AI scribes may become better over time; however, much of what is said in consultations is context-dependent, emotionally inflected, and shaped by subtle interpersonal cues. Many patients do not speak in clinical terms, and their concerns often emerge obliquely, through narrative, tone, and hints. Capturing these aspects requires interpretive judgement, something clinicians routinely apply, but which AI tools may struggle with.

The rapid emergence of AI tools in practice, often entering through the back door, is concerning.¹⁰ General practice is not simply about solving biomedical problems; it is about people, relationships, and trust. We owe it to our patients and profession to pause and question before allowing tools like these to reshape the way we practise our art.

References

1. UK Government. Government fires starting gun on AI Growth Zones to turbocharge Plan for Change. https://www.gov.uk/government/news/government-fires-starting-gun-on-ai-growth-zones-to-turbocharge-plan-for-change. [Accessed 6/6/2025]
2. UK Government. AI doctors’ assistant to speed up appointments a ‘gamechanger’. 2025. https://www.gov.uk/government/news/ai-doctors-assistant-to-speed-up-appointments-a-gamechanger. [Accessed 6/6/2025]
3. Razai MS, Majeed A. General practice in England: the current crisis, opportunities, and challenges. The Journal of Ambulatory Care Management 2022;45(2):135-39.
4. Khan N. Put a cap on it: safe workload levels in general practice. British Journal of General Practice2023;73(728):122-23. Khan N. DOI: https://doi.org/10.3399/bjgp23X732153.
5. BMA. Safe working in general practice in England guidance. 2025. https://www.bma.org.uk/advice-and-support/gp-practices/managing-workload/safe-working-in-general-practice. [Accessed 6/6/2025]
6. Casey M, Shaw S, Swinglehurst D. Experiences with online consultation systems in primary care: case study of one early adopter site. British Journal of General Practice 2017;67(664):e736-e43.
7. Leighton C, Joseph-Williams N, Porter A, et al. A theory-based analysis of the implementation of online asynchronous telemedicine platforms into primary care practices using Normalisation Process Theory. BMC Primary Care 2025;26:27.
8. Turner A, Morris R, Rakhra D, et al. Unintended consequences of online consultations: a qualitative study in UK primary care. British Journal of General Practice 2021
9. Heath I. The mystery of general practice: Nuffield Provincial Hospitals Trust London 1995.
10. Razai MS, Al-Bedaery R, Bowen L, et al. Implementation challenges of artificial intelligence (AI) in primary care: Perspectives of general practitioners in London UK. PloS one 2024;19(11):e0314196.

Featured Photo by Kaitlyn Baker on Unsplash