Falling out of love with AI

The more I used an AI scribe during consultations, the less sure I became about its benefits

I was an early adopter of an artificial intelligence (AI) scribe for GP consultations. In July 2025, I wrote in this newspaper about how the scribe had transformed my workday. I was effusive in my praise: Less typing, more freedom to concentrate on the patient, no need to keep lists, and less likely to miss something amid the multiple problems presented at each consultation.

“Once the story had been told and I had had my say,” I wrote, “AI would… formulate a comprehensive summary and plan for my records.” It would even draft a referral letter when required. I felt as if there was no going back to the old ways.

Unfortunately, within a year I had fallen out of love with my AI scribe. Niggling worries about my inexperience with data protection regulations as applied to AI made me uncomfortable. I worried about data processing, storage, and the company’s deletion/retention policy. I also worried that the company could use the data to train an AI model, fine-tune the existing model, or inform future GP textbooks or agentic GPs. Was the data truly anonymised, or was there a risk of patients being identified? Were patients giving fully informed consent without a full explanation of the terms and conditions of the company?

I tried to find satisfactory answers, but it wasn’t easy. Ambient AI tools are remarkably easy to use, but it is difficult to comprehend and navigate the legal and ethical implications for doctors and patients. Another problem with AI use is that it is advancing faster than the guidelines that safeguard users. The Data Protection Commission issued guidelines for the use of AI in healthcare in March 2026 and HIQA produced further guidelines in July 2026. Both documents are dense and difficult to interpret and both suggest that the responsibility for ensuring the company is GDPR [General Data Protection Regulation]-compliant rests with the GP.

As I wrote previously, AI did not save time in the consultation. It did, however, reduce the cognitive load and the effort of writing referral letters and reports. So, I persisted, but without my initial enthusiasm. By then, lots of GPs were using AI scribes, as were consultants and, in the UK, the NHS had introduced these tools to many hospitals. Also, AI has been used in healthcare in Australia and New Zealand for a few years with no reported difficulties I’m aware of.

After about eight months of use, I noticed that my consultation style began to change. I felt less engaged with patients and more engaged with the record of the previous consultation, with its neat, orderly summary, and plan. I voiced my examination findings aloud for the benefit of the recording, rather than for the patient. I allowed the patient more time to talk, but I didn’t have to listen as attentively, as I knew my scribe would weed out what it considered irrelevant to the problem of the day. I felt as if I was outsourcing cognitive processing to AI and I didn’t like how that felt.

My medical records were excellent (once I had corrected the hallucinations), but they did not reflect my experience of the consultation. I felt as if there was a third person in the room, a conductor who wanted the orchestra to play a particular way, and I wondered if the patient felt that presence too. Eventually, I managed to contact the company, pulled the plug, and went back to my old ways.

Other doctors I speak to have no such reservations about AI scribes. So I asked ChatGPT if there was any evidence that using scribes in a general practice setting changes the doctor-patient relationship. One 2026 study involving 12 GPs and 535 patients in the Netherlands found that while there were positive effects – slight reduction in consultation times, reduced workload for GPs, and improved communication reported by some patients – there were also negative ones.

These included inaccurate summaries, patients who were fearful of discussing sensitive subjects, and possible interference with the GP’s clinical reasoning. This study also found that AI can alter how the GP speaks, thinks, and structures the consultation. I had certainly found this in my personal practice.

A further study of GPs in New Zealand who used AI scribes revealed that while some reported shorter consultations and improved rapport with patients, many had concerns about compliance with legal and ethical frameworks, data security, and data stored outside the country. Only 66 per cent of physicians had read the terms and conditions and only 59 per cent had sought patient consent. 56 per cent agreed that the tool changed the dynamics of the consultation. These findings also reflected my own concerns about using scribes.

I am not saying that I will not revert to using AI in the future, but for now, I am happy to be alone with my patient and my flawed, typo-riddled records.