Week 21 delivered a fascinating mix of existential debates and practical innovation challenges. The group wrestled with profound questions about AI replacing GPs (sparked by a provocative poll showing 13-3 against adoption), whilst simultaneously confronting real-world barriers as Ankit.ai faced widespread ICB blocking despite being a free, well-governed administrative tool. New members brought fresh perspectives from Skin Analytics, Praktiki, and medtutor.ai, whilst the Government's announcement of AI radiology funding drew sceptical responses about "magic unicorns" versus addressing actual bottlenecks. Academic discussions ranged from clinical coding's future to OpenAI's revelation that 0.15% of ChatGPT use involves suicidal intent, bookended by an unexpectedly delightful detour into 1990s computing nostalgia. The community's personality shone through: simultaneously serious about patient safety and gleefully reminiscing about floppy disks and DOOM.
The Ankit.ai Blocking Crisis: Innovation Meets Bureaucracy
Sunday 26 October brought a rallying cry from the community when it emerged that Ankit.ai - a free, well-governed administrative tool for contractual queries and policies - was being systematically blocked by multiple NHS ICBs. The creator, a working GP, had invested personal time and resources, completed a thorough DPIA, and removed all potentially problematic elements (iframe heavy design, "generative AI" labelling, even the "buy me a coffee" link), yet still faced inexplicable barriers.
The group's response revealed deep frustration with systemic inconsistency. A clinical safety expert noted that truly non-clinical administrative tools should fall outside DCB requirements, though a lightweight risk assessment and DPIA remain helpful. When the comprehensive DPIA was shared, the consensus emerged: the problem wasn't the documentation but procurement and governance teams defaulting to "not our template" responses without understanding data protection principles.
One commentator wryly observed that Twitter and Facebook operate freely without NHS-specific DPIAs, whilst HSJ subscriptions face no such scrutiny. The suggestion that a helpdesk operator simply saw "AI" and added it to a block list without deeper assessment gained traction. A technical analyst pointed out: "Someone will have seen 'AI' and added it to a block list. Ankit's tool has enough warnings and notes that it'd take someone deliberately going off-piste to cause any harm."
Regional variation emerged as a key theme. North East London reported no blocking, with local teams being "very responsive" - a testament to progressive leadership. This geographic lottery led to calls for standardised national assessment processes rather than each of 42 ICBs conducting separate evaluations. The proposed Innovation Passport from the 10-year plan was discussed, though concerns were raised about its readiness and whether the solution matches the problem.
The creator's candid reflection resonated: "I suspect I either have to launch it as a commercial product or simply let it die. It doesn't feel that the system is keen to let things that work stick around." The discussion highlighted broader tensions between innovation, governance theatre, and the challenge of scaling grassroots solutions across a complex system.
Would You Bow to Your AI Overlords? The Great GP Replacement Debate
Late Saturday 26 October saw a GP innovator drop a philosophical hand grenade into the chat: a poll asking if colleagues would adopt a "regulated and certified tool that can work autonomously as Primary Care Clinicians." The results spoke volumes: 13 voted "no", just 3 "yes" (including one of the newsletter editors, who subsequently fielded questions).
The debate that unfolded across Saturday night into Sunday morning was remarkable in its depth. A newly CCT'd GP pleaded for "at least a couple of years before replacing me", whilst others raised fundamental questions about continuity of care, the performative aspects of medicine, and what "better" actually means in healthcare delivery.
The pragmatic perspective emerged from several quarters. One contributor noted: "I optimise for patient care. If something else genuinely does it better, I salute our replacement overlords" - though acknowledging that "better" does immense lifting in that statement. The counter-argument came swiftly: "I think we're whole generational science and tech breakthroughs away from a tech that can meet the standard of GP care I've had over the years."
The sociological angle proved particularly insightful. In a system free at point of delivery, with GPs as first contact and gatekeepers, "so much of medicine is ultimately performative/human-touch related and has little to do with true diagnostic, whether it be reassurance, signposting to other services, social issues." This observation sparked discussion about whether AI would force healthcare back to examining these fundamentals after decades of moving away from them.
Future scenarios ranged from dystopian to pragmatic. One participant suggested that "fewer GPs learning ultrasound" might ironically lead to GPs becoming radiologists whilst AI handles primary care. The liability question loomed large: manufacturers would need insurance, but who carries responsibility when a GP principal is simultaneously clinician, data controller, and practice owner?
Real-world context grounded the abstract. The observation that "differently trained people in the Medical Model" have already begun replacing traditional GP roles added bite. One experienced GP noted that straightforward presentations increasingly go to ARRS colleagues and ANPs, leaving GPs with trickier cases requiring System 2 thinking - potentially a pattern that intensifies rather than reverses.
The conversation concluded with a striking demonstration: asking Claude 100 times when autonomous GP AI replacements might arrive in the NHS yielded a distribution centred on 2038-2040, with 58% of responses falling in this tight window. The mean was 2038.8 years - approximately 13-17 years from now. Whether reassuring or terrifying depends entirely on your perspective.
Welcome Wagon: New Faces and Fresh Tools
Saturday 26 October saw the community expand with several significant additions, each bringing distinct perspectives and capabilities:
A Medical Director from Skin Analytics (former Babylon GP) joined with characteristic directness: "Thanks - looking forward to hearing and learning from this community!" His subsequent contributions on liability and indemnity proved invaluable, drawing on Skin Analytics' real-world experience navigating medical device regulation and insurance frameworks.
A pharmacist and content writer from Praktiki was welcomed alongside Praktiki's CEO and co-founder. Their arrival prompted discussions about the critical intersection of AI training tools and quality control, particularly around automated feedback systems. The consensus: having a "human in the loop" - whether trainers or "celebrity educators" - remains essential for medical education tools to maintain standards whilst leveraging AI's scalability.
A GP who built medtutor.ai ("I got sick of hearing 'AI will replace doctors' so I built something where AI replaces the patient instead") showcased patient simulations for SCA preparation with instant feedback. The tool's clever approach - allowing trainees to practice with consistent AI patients, share recordings with trainers for unbiased assessment, then adjust and try again - earned widespread praise. One educator reported using it in tutorial on Friday and finding it "very useful," whilst offers of free trainer access were gratefully accepted for wider distribution.
The warm welcomes reflected the group's culture: professional respect, genuine curiosity about new tools, and immediate engagement with the hard questions (in medtutor.ai's case, how to ensure quality control in AI-generated feedback). The mix of clinical innovation (melanoma detection), education technology (safe prescribing, SCA prep), and diverse professional backgrounds enriched ongoing conversations throughout the week.
NHS Digital Planning: The Path to "Digital by Default"
Tuesday 29 October brought extensive discussion of the newly published three-year planning framework, with its boldest target: 95% of appointments bookable through the NHS App by end of 2028-29. A detailed community analysis highlighted both ambitions and concerns about this journey toward "digital by default" delivery.
Key targets include a national product adoption dashboard tracking electronic prescriptions, e-referral interfaces, and NHS App integration by March 2028. The directive to "terminate" direct-to-patient SMS services in favour of NHS App push notifications represents significant change, potentially freeing substantial resources whilst requiring careful transition management.
The Federated Data Platform (FDP) generated most controversy. ICBs "should use the FDP for data warehousing" and implement the canonical data model - language that community members interpreted as "mandating adoption". One commented: "It seems we're being taken for idiots... If the FDP is genuinely transformational, prove it. Publish an independent evaluation, total lifecycle cost, implementation burden, data governance risks, and measurable improvement." The full project cost exceeding 1bn raised further questions.
Ambient voice technology (AVT) received prominent placement: providers should adopt "at pace," though central support was ruled out for 2025-26 implementations. The expectation that business cases will "write themselves" due to productivity improvements drew scepticism. A system supplier warned against "rushed implementations of standalone solutions that aren't integrated with clinical systems" - appearances in "short-term pilots" versus "full potential productivity benefits" of integrated solutions.
The Health Foundation's caution resonated: with healthcare providers at "widely different levels of digital maturity," supporting organisations "lower down the curve to build the infrastructure to deploy AI effectively" remains crucial. The concern: AVT represents a "wild West" market where serious problems may emerge before regulatory frameworks catch up.
Strategic concerns extended beyond individual technologies. The single patient record's notable absence from planning guidance suggested unresolved decisions about architecture. Cybersecurity's omission seemed particularly curious. The relationship between national direction and local implementation - always fraught - faces new tests when central mandates ("should use") meet constrained local resources and varying digital maturity.
One CIO's observation captured the tension: the FDP has become an "obsession" for NHS England, yet many trusts face "a lot of push-back just due to the cost of change" whilst the canonical data model remains "only partially developed." The three-year visibility represents genuine improvement over annual instability, but successful delivery requires more than targets - it demands resources, training, integration support, and realistic acknowledgement of starting positions.
"Magic Unicorns" and Radiology Bottlenecks: AI Announcement Reality Check
Tuesday 29 October's government announcement about AI-powered radiological analysis funding triggered immediate scepticism from clinical informaticians who've seen this pattern before. An image circulated showing the announcement alongside a critical observation: addressing waiting times requires fixing actual bottlenecks, not just accelerating one step in the pathway.
A consultant's reaction set the tone: "Oh for god's sake. Paging Margaret McCartney for some sense." Another noted: "Only if it's radiology reporting waiting times that are the rate limiting factor" - pointing to the fundamental flaw in assuming faster image analysis automatically equals faster patient journeys.
The technical details mattered. Questions emerged about device classification, clinical effectiveness validation, and whether public assessment data exists for review. The companies mentioned (Lucida, Quibim) have solid credentials, and integrating with digital pathology AI could indeed "support better pathways following on from diagnostics" - but the announcement's framing as a major breakthrough ignored systemic constraints.
Implementation reality provided sobering context from a practice using Skin Analytics for melanoma detection. Their deliberate approach includes built-in delays: scanning clinicians don't get instant results, patients receive callbacks the next day, and pathways explicitly avoid "instant gratification" precisely because this is "a serious medical device and the pathways must reflect that on how it impacts patients." Average time from GP visit to outcome: 2.8 days, considerably better than hospital routes. The contrast with government announcements promising rapid transformation couldn't be starker.
One digital health lead characterised this as the "shallowness of health policy... prior to the election GPCE were concerned it would be all magic unicorns, and here we are." The phrase "magic unicorns" captured the frustration: announcements that sound transformative but ignore the unglamorous work of pathway redesign, staff training, system integration, and addressing actual rate-limiting steps.
The underlying pattern reflects broader tensions. New technology alone rarely solves complex system problems. MRI scanner capacity, radiologist numbers, pathway design, and onwards referral routes all constrain patient flow. Accelerating image analysis without addressing these factors simply moves the bottleneck elsewhere - potentially creating new problems (overflow, loss of contemplation time, information overload) whilst claiming victory.
The community's response demonstrated sophisticated understanding: celebrating genuine innovation (AI radiology has proven value) whilst maintaining scepticism about implementation rhetoric that oversells and underdelivers. As one commented: "This is absolutely the headline that highlights the shallowness of health policy."
The End of Clinical Coding? LLMs and Healthcare Language
Sunday 27 October saw renewed debate about whether AI will eliminate clinical coding, sparked by a LinkedIn post highlighting fundamental tensions between structured ontologies and natural clinical language. A digital health specialist's response captured decades of accumulated wisdom: "Clinical Coding is important, of course, but it has always been an effort to take humans towards speaking in a language that can be easily understood and transferred digitally."
The historical context matters. SNOMED codes now exceed common English vocabulary by 5:1 - "almost no human being can remember all the Snomed codes!" One contributor noted coding served two purposes: knowing exactly what clinicians meant, and enabling statistics (machine readable data). But with SNOMED's explosion, the system has become unwieldy whilst remaining useful for EHR searches and potentially as training data tokens for AI models.
The LLM revolution offers a different approach: "Why reduce the complexity of clinical reality to a hierarchical graph, impressive as they have become, when you can have it embedded in several thousand dimensions without the need to be limited by human comprehensibility?" This philosophical shift - from forcing human thought into structured codes to allowing machines to process natural language in its full complexity - represents genuine transformation.
Practical considerations tempered pure enthusiasm. One expert suggested codes still serve roles: triggering pathways and actions in rule-based systems, maintaining consistency across healthcare informatics infrastructure, and providing searchable structured data. "Ideally hospitals would have an orchestration bus connecting all systems that used (probably several) LLMs to code stuff but the edges still need coded info, unless you want an LLM in every system, with all the possibility of variance that comes with that."
The transition challenge emerged as key. Healthcare operates on existing infrastructure built around coded data. Moving to LLM-processed natural language requires careful orchestration: which systems need codes, which can work with semantic understanding, how to maintain interoperability, and how to avoid introducing new failure modes. The technical debt of decades of coding-centric systems won't disappear overnight.
The conversation reflected broader themes: technology enabling return to more natural human communication patterns, the challenge of maintaining consistency whilst embracing flexibility, and the tension between revolutionary possibility and evolutionary reality. Clinical coding may not disappear completely, but its role will fundamentally shift as LLMs mature - from primary data structure to legacy compatibility layer.
AI Slop and the Value of Authentic Writing
Monday 27 October saw a Northern Ireland GP identify an emerging pattern: "Noticing a HUGE trend in emails in NHS and beyond in folks using AI to reply. Em dashes, three point phrasing in sentences, the lot." His poll showed overwhelming irritation: 16 voted "yes" (AI replies irritate), just 1 voted "no". The discussion that followed explored why authentic writing increasingly matters in an age of AI generation.
The nuanced reality emerged quickly. For "purely transactional things, absolutely not" irritating. But "for things that I expect a human eye on, yes I do and I will judge the company." The expectation that companies should "read my mind to get that line right" captured the impossibility: AI-generated content works for some contexts, fails dramatically in others, and the difference hinges on emotional stakes and relationship expectations.
The strategic insight came from an innovation-focused GP: "Your writing will be your biggest moat in the world filled with AI Slop. Protect it." The phrase "AI slop" - low-effort, generic AI-generated content - gained immediate traction. In a world where anyone can generate plausible-sounding text instantly, authentic human voice becomes differentiating. The comment "The finger is mightier than the Clanker" captured this sentiment.
Practical applications varied. One experienced user employed Apple Intelligence powered by OpenAI or Perplexity: "I usually write the responses and then ask AI to enhance it. If simple responses, then give the prompt to reply with thanks." This augmentation approach - human thought, AI polish - represented a middle ground between pure automation and complete rejection.
The deletion principle offered pragmatic advice: immediately delete emails one has "no intention of responding to." AI-generated replies actually help this process: "It allows me to ignore them emails straight away." The ability to quickly identify auto-generated content creates an arms race between generation and detection, with implications for attention, trust, and communication effectiveness.
The discussion reflected deeper questions about authenticity, effort, and value in professional communication. In an era when generating plausible text costs nothing, what does it mean when someone invests time in careful writing? What signals does AI generation send about the sender's priorities? And how do recipients calibrate their responses when they can't be sure a human read their original message?
Quote Wall
"Your writing will be your biggest moat in the world filled with AI Slop. Protect it." -- Innovation-Focused GP on authentic communication in AI era
"I optimise for patient care. If something else genuinely does it better, I salute my replacement overlords. I appreciate that 'better' is doing alot of lifting." -- Digital Health Specialist on autonomous AI clinicians
"2025 luddites in full swing" -- Recently Qualified GP on Ankit.ai blocking
"So much of medicine is ultimately performative/human-touch related and has little to do with true diagnostic, whether it be reassurance, signposting to other services, social issues." -- Recently Qualified GP on the future of general practice
"This is absolutely the headline that highlights the shallowness of health policy... prior to the election GPCE were concerned it would be all magic unicorns, and here we are." -- Practice-Side GP Lead on radiology AI announcements
"Why reduce the complexity of clinical reality to a hierarchical graph, impressive as they have become, when you can have it embedded in several thousand dimensions without the need to be limited by human comprehensibility?" -- Digital Health Specialist on LLMs versus clinical coding
"I got sick of hearing 'AI will replace doctors' so I built something where AI replaces the patient instead." -- GP and medtutor.ai creator on education technology
"Someone will have seen 'AI' and added it to a block list. Ankit's tool has enough warnings and notes that it'd take someone deliberately going off-piste to cause any harm." -- Clinical Safety Expert on governance theatre
Journal Watch
Academic Papers and Key Studies
Nature Medicine: "Do AI guardians protect us from health information overload?"
Nature Medicine: "LLM Performance in Clinical Consultations" - Researchers designed an AI agent simulating human patients to test LLMs' clinical capabilities across 12 specialties. Key finding: all LLMs performed significantly worse in conversational consultations compared to exam-style questions. Sparked Sunday discussion about the gap between benchmark performance and real-world diagnostic capability.
ArXiv: "Illustrated Guide to Transformers"
ArXiv: "Diffusion Models Deep Dive"
ArXiv: "Thermodynamic Computing and P-bits"
Industry Articles and News
Guardian Long Read: "DeepSeek is humane, doctors are more like machines"
CNBC: "Amazon announces sweeping corporate job cuts"
The Hindu Business Line: "Narayana Hrudayalaya acquires UK's Practice Plus Group"
HSJ: "The Download - The path to 'digital by default' is now clearer"
LinkedIn Post: "Question of the year - generating positive ROI with AI agents"
Indeed Hiring Lab: "AI at Work Report 2025"
Technical Resources and Guidelines
Ankit.ai DPIA Documentation
OpenAI: "Strengthening ChatGPT responses in sensitive conversations" - Technical blog post revealing 0.15% of weekly ChatGPT use involves element of suicidal intent. Details improved response protocols and safety measures.
Brave Browser Blog: "Comet Prompt Injection Vulnerabilities"
VentureBeat: "When your AI browser becomes your enemy"
NHSE: "GP Clinical Systems Experience Survey"
Machine Learning University Explainers
IEEE Spectrum: "MLPerf Trends - AI Growth vs Hardware Struggles"
Policy Documents and Official Reports
Companies House Filings: TPP Director Changes
Sky News: "UnitedHealth considers Optum UK sale"
Telegraph: "NHS staff sick days cost 1bn per month"
Digital Health News: "TPP Director Changes Confirmed"
Conferences and Events
Bradford Quantum Hackathon 2025
Four Nations Conference: AI for Education
Group Personality Snapshot
Therapeutic Nostalgia: The ability to pivot from existential debate about AI replacing GPs to collective reminiscence about floppy disks and DOOM demonstrates remarkable emotional range. The retro computing thread wasn't distraction - it was processing space, allowing the group to decompress from heavy topics whilst reinforcing shared generational experiences. "I'm amazed that amongst the heroin trade there was a thriving floppy underground in 1990s Glasgow" captures the group's gift for finding absurdist humour in unexpected places.
