On a Saturday afternoon (10 October), I asked a room of private GPs, sponsors and exhibitors at the Private GP Summit on the Isle of Man to vote: is an AI scribe just a very good pen, or nothing like one? Over the weekend I learned that most of the private practices represented in the hall from the UK and the Isle of Man already use AI scribes (save for one hand in the audience). I also learned about the safety and compliance gap that exists in private practices. I shared what stood out to me as potential issues from the perspective of a litigator.
On the fictitious case I designed as a role play, the AI scribe errors were easy to spot. Exertional tightness became “denies chest pain”. Hydralazine became hydroxyzine. The 999 advice simply vanished. It was modelled on the kinds of errors reported in the larger of two randomised trials of ambient scribes so far, conducted with 238 US physicians, which found using these AI tools brought a real reduction in burnout and some reduction in time. Clinically significant inaccuracies still appeared “occasionally” (Lukac et al., NEJM AI, 2025).
Those errors are exactly what a tired clinician between patients might miss. When a clinician signs off on the scribe-produced notes, they endorse the notes as accurate. If a claim follows, the note becomes the main evidence of what happened in the consultation. Often times, the audio recording is not retained after the session by the AI scribe either. But these are generative AI tools after all, with all the baked-in potential for errors and biases, and courts and insurers have yet to decide what they want to do with liability apportionment, after a clinician signs off on those notes.
What struck me over three days was not resistance to AI. Quite the opposite. A number of the GPs I spoke to have already folded AI into their working day. Some shared anecdotes of how other UK clinics use AI concierge tools as the point of contact with patients, and in one story I heard, helping non-clinical staff triage them. When I raised a concern, a GP replied, reasonably, that triage is not normally done by a GP anyway.
But what if errors are not always so easy to spot? How would private GPs manage AI risks and governance issues?
In England, NHS England sets mandatory clinical standards for clinical risk management and safety (DCB0129 for the maker, DCB0160 for the deploying organisation), each requiring its own Clinical Safety Officer. But a private GP has none of that by mandate. There is often not much else the private GP has by way of compliance or safety oversight, apart from their own awareness and vigilance.
The island’s current care regulations do not mention AI at all. The delayed Regulation of Health and Social Care Bill, which would replace the Regulation of Care Act 2013, is now expected early next year. Device law does not fill the gap either: UKCA marking does not apply in the Crown Dependencies, and the MHRA’s July 2026 guidance on the status of AI scribes does not seem to cover the Isle of Man.
Since the event, I have been reflecting on how AI might impact on professional judgment.
There are several phenomena emerging from studies that are worth highlighting. It seems to me that lay people are, on average, becoming better informed with the aid of AI, because it’s easier to access a broader range of material that used to take intensive research effort to understand. AI also seems to flatten the range of reasonably held opinion, as LLMs increasingly become a built-in echo chamber, and hence an information bubble. The result seems to be that unique insight from experience becomes rarer. Finally, professionals who grow to rely on AI are likely to deskill over time. For the knowledge worker, this is really worth paying attention to.
Studies already reflect this shift
An MIT team put language models to work as traders in a simulated market. The more capable the models, the more alike they behaved (Ross et al., 2026, preprint). That is fine when their shared reasoning is right. When they share a misinformation environment, the correlation becomes the risk, and better individual models cannot remove it. A review of more than 130 studies in Trends in Cognitive Sciences makes the human version of the point: language models are flattening, or homogenising, how people write and reason (Sourati et al., 2026).
In January 2026 Anthropic ran a randomised trial on 52 junior software engineers learning a new library. Those with an AI assistant scored 50% on a later quiz, against 67% for those who coded by hand, and saved no significant time (Anthropic, 2026). How they used it mattered more than whether they did. People who asked the AI to explain scored 65% or more. People who let it write the code scored under 40%. The authors’ own line is the one I keep: “getting painfully stuck” is likely part of how mastery forms.
In four Polish endoscopy centres, doctors who had routinely used AI to spot polyps were later observed working without it. Their adenoma detection rate fell from 28.4% to 22.4%, a relative fall of about a fifth, after a few months of AI use (Budzyń et al., Lancet Gastroenterology & Hepatology, 2025). It is an observational study, so it shows an association, not proof. But it is one of the first hard numbers on clinical deskilling.
There are clients these days who seek professional advice, but really, are looking for confirmation that what their LLMs told them is correct. As AI tools become more sophisticated, relatively mundane pieces of work requiring less professional judgment might well be eaten up by tokens burned privately. But how professionals keep their knives sharp, and how they signal that differentiation in a market of ‘ingest and synthesise’ practitioners armed with agents, will probably become the next questions for professionals.