Executive Summary
Valentine's week in the AI in the NHS group proved that the community's deepest affection remains reserved for clinical safety, data sovereignty, and the perfect LLM subscription tier. The week's standout theme was a passionate debate on NHS patient data sharing, triggered by new GPES directions for consented research, which exposed deep fault lines around trust, anonymisation, and the Federated Data Platform. A shocking revelation from an emergency department, where clinicians were overheard pasting patient notes into ChatGPT, ignited a searing discussion on shadow AI in clinical settings and the urgent need for sanctioned tools. Rumours that EMIS/Optum may be changing hands sent ripples through the community, raising fresh concerns about data sovereignty and foreign ownership of UK health infrastructure. And as the week drew to a close on Valentine's Day itself, the group turned its critical eye to an Oxford study on LLM diagnostic accuracy, finding that while the research methodology was sound, the models tested were already relics of a bygone era, raising fundamental questions about academic publishing's ability to keep pace with AI's breathless rate of change.
The Week's Big Conversations
1. "It's Always About Trust" - The NHS Data Sharing Debate (Tue 10 Feb)
The week's most impassioned exchange was sparked by news that NHS England has issued new GPES directions enabling the extraction of GP data for participants in Biobank, Our Future Health, and Genomics England. What began as an informational share rapidly evolved into a community-wide reckoning with some of primary care's most fundamental tensions.
A data governance specialist provided crucial context: the directions apply only to patients who have already given consent through those research programmes, will not override Type 1 opt-outs, and repurpose the existing GDPPR pandemic dataset. Importantly, consent itself is not the legal basis for the data flow; rather, the Secretary of State direction provides the legal route, limited to consented individuals.
The group's clinical safety contingent pushed back firmly. One informatician with deep technical knowledge of NHS systems noted that "GPs in unlimited liability partnerships have to be very careful about turning on data flows voluntarily to organisations who say they have consent." The discussion illuminated the uncomfortable reality that even when every regulatory box is ticked, data controllers face genuine personal liability risks.
A health informatics lead crystallised the trust deficit: "I fundamentally do not want my NHS held data shared with anyone without a written commitment from people in power that they'll go to jail if it's ever sold for commercial gain or misused." This prompted a nuanced exchange on anonymisation robustness. One contributor asked the crucial question: "How confident are we that 'properly anonymised' data can't be re-identified, especially when different datasets get linked over time?" The consensus was that large-scale aggregation combined with proper anonymisation offers strong protection, but that systems architectured like OpenSAFELY, which keep data in situ rather than extracting it, represent a fundamentally more trustworthy approach than the Federated Data Platform.
As one innovation-focused GP summarised: "It is fundamentally about trust." The group agreed unanimously, even as they diverged on the solutions.
2. Shadow AI: The Emergency Department Wake-Up Call (Wed 11 Feb)
Perhaps the week's most alarming moment came from a group member sitting in an emergency department who reported hearing residents say "paste my medical notes into ChatGPT." The revelation, delivered in real-time from the clinical frontline, electrified the group.
An AI-focused GP and medical device developer shared striking data from their own teaching sessions: approximately 70% of GP trainees admitted to using off-the-shelf AI tools during consultations. This figure, they noted, has remained stubbornly consistent across multiple teaching events. They also disclosed having to remove a colleague from their practice for the same behaviour, noting "We have clear policies."
The implications cascaded. A Gulf-region statistic was cited claiming 80% of clinicians have used unsanctioned generative AI in patient care. One contributor working in Northern Ireland noted that LLM access is blocked entirely on both primary and secondary care networks there. Others pointed out that in Norfolk, ChatGPT and Perplexity are freely accessible on NHS networks, whilst approved clinical AI tools remain blocked.
The discussion exposed a generational shift with worrying undertones. The AI-focused GP shared an anecdote from consultant teaching: junior doctors have stopped asking supervisors questions between patients, instead consulting ChatGPT. "As a result we don't know the skills and development that they are having and find it harder to manage complexity at arm's length."
The group's clinical safety advocates were firm. As one put it: "If you have a rule you don't enforce, you'd be as well not having it." Another countered that purely punitive approaches don't build trusted cultures and that the real solution lies in providing clinicians with sanctioned, safe AI tools. The thread ended with broad agreement that this represents perhaps the most urgent patient safety challenge the NHS currently faces in the AI space.
3. EMIS on the Block? Rumours of a Sale Send Shockwaves (Thu 12 Feb)
A GP and digital health advocate attending the HETT Leaders Summit dropped what became the week's most consequential rumour: that Optum/UnitedHealth has sold EMIS, with an investment and development freeze in place. "You heard it on the AI in the NHS group first!" they declared, with characteristic flair.
The rumour was subsequently corroborated by multiple members and supported by a MarketsGroup report indicating TPG is eyeing acquisition of UnitedHealth's Optum UK business. A Pulse article from the previous month had also reported the potential sale at a valuation exceeding £1 billion.
The community's reaction blended concern with dark humour. One veteran quipped that "the freeze came into force in 1997." A data specialist observed drily that "the PID in it is worth more than the IP. That's what's likely got buyers sniffing around." A digital health leader voiced the group's primary concern: "If I were the NHS procurement folk, I would not allow a sale to anyone who isn't operating via a legally separate subsidiary/group company based solely in the UK."
The discussion broadened into reflections on data sovereignty. Members noted that Singapore, UAE, and Ethiopia are all pursuing sovereign AI approaches. As one contributor observed, "Quite a few US companies think the NHS is an inefficient cash cow just waiting for milking. Then they get here and find out that it's not quite that simple." A healthtech industry veteran confirmed this from direct experience, noting they'd "worked for a Korean market entrant with same misconception."
4. Training Tomorrow's Centaurs: The Curriculum Gap (Fri 13 Feb)
Friday's discussion turned to a question that connects every preceding theme: how do we prepare the next generation of clinicians for an AI-augmented future?
The conversation was catalysed by a GP reflecting on the pace of change: "Given the pace of change, how do we best ensure the next generation of GPs are prepared for the future role of centaur/human in an almost autonomous loop? There is still no clearly defined 'medical informatician' or Clinical AI guardian training pathway."
The group identified a systemic failure. A health policy analyst noted that "If professional bodies don't step up then it's likely groups of practices and hospitals will define their own modern AI-orientated professions." The curriculum update process itself was identified as a bottleneck: "When you try to get something added to the curriculum it can take a couple of years, by which time it's already outdated."
One experienced clinical safety trainer pointed out that their CSO/CRM training was just two days and argued there is "absolutely room in a 3-year UK NHS GP curriculum for 2 days of CRM training." Another revealed that one country is already offering a free six-month online AI in health course for all doctors on its national register.
A trainee shared a sobering anecdote: during regional teaching with over 50 GP trainees, only 6 expressed genuine interest in learning about AI. Meanwhile, a digital health specialist shared their determination not to wait for institutional change: "I was asking for this in the early 90s. I've now got lectures booked with medical students at some medical schools, and I'll keep pushing at this. It's frankly lunacy to have no core components for something as integral to the practice of medicine."
5. When Research Can't Keep Up: The Outdated Models Problem (Sat 14 Feb)
Valentine's Day brought a characteristically thoughtful exchange on a fundamental challenge for AI in healthcare: academic research timelines versus the pace of model development. A digital health specialist highlighted a LinkedIn post examining a recent Oxford study on LLM diagnostic accuracy, noting that the models tested (GPT-4o from May 2024, Llama 3 from April 2024, and Command R+) are already significantly outperformed by their successors.
The group was careful to praise the study's rigour whilst identifying the systemic problem. A health policy analyst observed it was "a good paper and rigorous methodology let down by the incentives and processes of academic publishing." Others agreed: the Ethan Mollick GPQA Diamond benchmark chart was shared, showing GPT-class models scoring approximately 42% in April 2024 versus o1-high achieving 91%+ by late 2025.
The concern was that as these findings filter through the media, they risk creating a misleading impression that AI diagnostic tools are not ready for clinical application, when in reality the current generation of models may perform dramatically better. As one contributor noted: "We need to see more research on agentic and 'low-challenge' one-shot tasks, as this is where the bulk of benefits will be seen right now."
Lighter Moments
The week offered its customary blend of wit and warmth. When a practice manager described asking Claude whether upgrading was worthwhile, the AI apparently replied, to paraphrase: "No, you are a dullard. Nothing I am doing for you is very complicated." The group moderator's verdict? "Proof positive that they've worked on the sycophancy issue."
The same practice manager later returned, having been seduced into a Pro subscription after "an idea that turned into a project that turned into a lot of code that Claude did that meant I needed to upgrade." When asked about their investment strategy, they revealed they had "diversified: 3D printer filament, Haribo, and AI subs."
A GP's introduction to Openclaw prompted the question "Is that a fighting style?" to which the reply came: "'Your crane style is no match for my openclaw' swish / whoosh." A clinical safety expert's response set the tone: "Organic and ethically farmed from quality suppliers please. Be kind."
When sundials were mentioned in the context of an AI benchmark joke, one member noted wryly: "We all know the main problem with sundials in the UK at the moment." And an occupational health GP shared the group's collective relationship status with AI subscriptions: "I'm also subscribed to Google LM, Perplexity, ChatGPT, and Gamma AI. I think I might have an addiction issue." The response: "Welcome to the club."
A GP musician topped off the week by releasing their latest EP, "IG: The Signoff Sessions," following up the previous week's album "SARs Are for the Living." Genre: primary care clinical jazz.
Quote Wall
"I fundamentally do not want my NHS held data shared with anyone without a written commitment from people in power that they'll go to jail if it's ever sold for commercial gain or misused." -- Health informatics lead, on the trust deficit
"If you have a rule you don't enforce, you'd be as well not having it." -- Clinical safety specialist, on shadow AI policy
"Quite a few US companies think the NHS is an inefficient cash cow just waiting for milking. Then they get here and find out that it's not quite that simple." -- Health informatics lead, on foreign investment in NHS infrastructure
"Power is shifting folks, aided by the advances with AI. Primary Care was always the absolute engine of innovation, and now you have the tools and the beginnings of repatriated power." -- Digital health specialist, on the changing landscape
"It's frankly lunacy to have no core components for something as integral to the practice of medicine." -- Digital health specialist, on AI in medical curricula
"I have diversified: 3D printer filament, Haribo, and AI subs." -- Practice manager, on investment strategy
"Why do they need to think deep? Life is pretty shallow." -- Innovation-focused GP, on Gemini 3.0 Deep Think
Journal Watch
Academic Papers and Research
"Video/World Models as the Next Step Beyond LLMs" - arXiv arxiv.org/html/2602.02603v2
LLM Diagnostic Accuracy Study (Oxford) - Nature/LinkedIn discussion LinkedIn post
AI Agent Benchmarks - IEEE Spectrum spectrum.ieee.org/ai-agent-benchmarks
Stigma in Digital Health - BMJ bmj.com/content/392/bmj.s124
AI Enters the Operating Room - Reuters Investigation reuters.com
Industry and Policy Articles
"AI Doesn't Reduce Work, It Intensifies It" - Harvard Business Review hbr.org Widely discussed HBR piece arguing that AI productivity tools create new demands rather than freeing up time. Simon Willison's companion piece: simonwillison.net
"Unhelpful: The Professional Body That..." - Substack open.substack.com
"It'll All Come Out in the Wash" - LinkedIn linkedin.com
OpenAI Ads Won't Pay the Bills - The Register theregister.com
#QuitGPT Campaign - MIT Technology Review technologyreview.com
Sue Ryder to Roll Out AI Scribing in Hospice Care - Digital Health digitalhealth.net
FDA Medical Device Rules for AI/Wearables - IEEE Spectrum spectrum.ieee.org
TPG Eyes Acquisition of UnitedHealth's Optum UK Business - MarketsGroup marketsgroup.org
EMIS Set to be Bought by US Private Equity - Pulse Today Pulse Today
Michael Burry: "Emperor Palantir Has No Clothes" - The Times archive.is/C9Let
Technical Resources and Tools
GPES Data for Consented Research Directions 2026 - NHS Digital digital.nhs.uk
Euria: Swiss Sovereign AI Chatbot - EuroToday eurotoday.org
Picoclaw - picoclaw.net
Codescan - GitHub github.com/codescan-ai/codescan
Loom - GitHub github.com/cosmix/loom
Ollama + Claude Code Integration docs.ollama.com
Gemini 3.0 Deep Think - Google Blog blog.google
AI-Driven Risk-Based Outreach for Reducing No-Shows - FutureNHS future.nhs.uk
BCS Primary Healthcare Specialist Group Conference 2025 bcs.org
SNOMED Pathology Bounded Code List Mapping diseasesdatabase.com
Events
Health AI Tech Show healthaiinsiders.com
Mistral Worldwide Hackathon worldwide-hackathon.mistral.ai
NHS IP Management Survey forms.office.com
Looking Ahead
Several threads remain unresolved and ripe for continued discussion. The EMIS/Optum sale rumours will inevitably develop further as details emerge; the community will be watching closely for confirmation and implications for GP data sovereignty. The shadow AI crisis demands practical solutions: the group has clearly articulated the problem but the NHS remains without a coherent strategy for providing clinicians with safe, sanctioned AI tools. The curriculum reform discussion is gaining momentum, with several members now actively pursuing medical school partnerships. And the recurring tension between academic research timelines and AI development velocity suggests the community may soon need to articulate what "good enough" research looks like in a field that moves faster than peer review.
The GP London Resident Doctor Annual Conference on 25 March is actively seeking speakers for a technology panel. Rewired is on the horizon. And with the group now exceeding 700 members, the community continues to grow, both in size and in influence.
Group Personality Snapshot
Valentine's week revealed a community that wears its heart on its sleeve, even when that heart is mostly concerned with RBAC permissions and DPIA compliance. This is a group where a practice manager's tale of being roasted by their own AI subscription sparks more engagement than most actual AI product launches. Where a GP musician can release an EP called "The Signoff Sessions" and receive genuine enthusiasm rather than bewildered silence. Where someone can share "Why do they need to think deep? Life is pretty shallow" and everyone understands it's both a joke about Gemini and a philosophical observation about healthcare AI.
