Written evidence from James Spalding, Founder at 8GI Foundation (PMA0107)

Front page — summary and policy recommendations

Summary

This submission is made four days after the launch of the Government's Sovereign AI Unit on 16 April 2026 (Department for Science, Innovation and Technology; Chair: James Wise of Balderton Capital; £500m fund; first cohort of seven companies announced at Wayve HQ, King's Cross). We welcome the Unit as a necessary first step. Our submission to this Committee is that the Unit, as currently designed, is a venture-capital-style equity vehicle targeting upstream-of-clinic research (drug discovery, molecular binding, biomanufacturing, defence world models). It is not yet a public-utility layer for NHS-facing, underserved-population-first personalised medicine. That complementary layer is the subject of this evidence.

This submission is made by the 8GI Foundation, an international civil-society organisation working on AI governance from a sovereignty-and-access-first perspective. Our argument to the Committee is that the most important question about personalised medicine and AI in the NHS is not whether AI will transform care, but who will own the resulting medical intelligence, on whose terms the NHS will access it, and which populations are served first as AI capability grows exponentially.

Our submissions to date include the Oireachtas Regulation of AI Bill consultation (Ireland), and a joint submission in progress to the South African Department of Communications and Digital Technologies on the National AI Policy Framework. This is our first submission to a UK Parliamentary inquiry.

We write from three combined perspectives: as a civil-society foundation focused on AI governance, as builders of open-source assistive AI tools for neurodivergent and disabled users, and as a parent of a non-speaking disabled child navigating disability services (in Ireland, not the NHS, but with transferable observations about how state health and care systems procure and deploy assistive technology).

The Committee's call for evidence asks how AI and personalised medicine can benefit NHS patients, what unlocks that benefit, and what government must do. Our answer is that the technology is not the bottleneck. The bottleneck is governance capture, data asymmetry, and a procurement culture that tends to buy access to American frontier models rather than grow a sovereign NHS-governed alternative. As AI capability compounds exponentially, the distance between populations served by commercial frontier models and populations structurally deprioritised by them will widen, not narrow. This is the K-shaped cut at the heart of our argument: without deliberate sovereign public-utility intervention, the upper arm of the K accelerates and the lower arm is cut away from the system entirely.

Policy recommendations

  1. Extend the Sovereign AI Unit mandate to include a public-utility NHS-governed foundation model. The Unit launched 16 April 2026 as an equity-and-compute vehicle backing a first cohort of seven upstream-of-clinic companies. That work is welcome and necessary. It is not, however, a delivery-phase personalised-medicine instrument. We recommend the Committee urge Government to add a complementary mandate: an open-weight, NHS-owned, UK-hosted medical foundation model trained on NHS-consented data, available to clinicians and accredited researchers under open licensing, governed as a public good in the same category as NHSmail or the NHS App. The Unit funds the upstream. The NHS needs the downstream.
  2. Mandate a data-sovereignty preference in NHS AI procurement. Where a UK-hosted, UK-governed model can do the task, it should be preferred to a foreign closed-weight model, and the procurement trail should make that preference explicit and auditable.
  3. Require governance-capture disclosure in all NHS-AI partnership contracts. Every vendor shaping clinical AI policy, integration standards, or commissioning frameworks should disclose commercial interest in the outcome. This is a Spalding-Paradox guardrail: the entities with the most commercial interest in minimal regulation are currently dominating the governance conversation.
  4. Prioritise underserved populations first, not the largest markets. Disabled children, non-speaking patients, rare diseases, mental health, and neurodivergent adults should be first-tier beneficiaries of NHS personalised-medicine AI, not the last. This is both a moral and a strategic position: the highest personalisation-value cohorts are precisely the ones commercial markets routinely deprioritise.
  5. Return AI-derived insights to patients under their own data rights. Any insight generated from NHS-held data should be returnable to the patient it originated from, in a portable, machine-readable form, under the patient's control. Data sovereignty is not only an institutional question. It is also an individual one.
  6. Require model provenance labels for clinical AI. Any AI used in NHS clinical decision-making should carry a provenance label: what data trained it, who owns the weights, what jurisdiction the weights and inference sit in, and whether those weights can be withdrawn by the vendor. Clinicians and patients should be able to read this label the way they read a drug leaflet.
  7. Treat public AI as a public service, in the same category as the NHS itself. The NHS-governed foundation model recommended at (1) is the clinical instance of a broader principle: any AI system that shapes how citizens reason about their own health, education, welfare, or rights is not a consumer product. It is civic infrastructure. If the UK allows frontier AI to become the default interface between citizens and state services, it has outsourced the value judgements that shape public life, what is safe, what is fair, what is true, what is normal, to a handful of foreign commercial vendors whose incentives are not civic. As decisions at the margin (triage, eligibility, appeals, follow-up, risk explanation) are progressively handed off to AI systems, the vendor of the model becomes, in effect, an unelected regulator of public life. The UK has a century of public-service precedent (BBC, NHS, Royal Mail, National Grid) to draw on. Public AI should sit in that lineage, not be treated as a Software-as-a-Service line item.

 

Numbered written evidence

Section 1 — About the submitter

  1. 8GI Foundation is an early-stage international civil-society organisation focused on AI governance, sovereignty, and equitable access. We are not an NHS body, not a clinical organisation, and not a UK-resident entity. Our relevance to this inquiry is structural, not clinical.
  2. The author has no formal medical training and makes no clinical claims. Any clinical illustrations in this submission are drawn from publicly reported NHS evidence or from the author's lived experience as the parent of a non-speaking disabled child (in Ireland), which is disclosed wherever it informs a point.
  3. The author builds open-source AI tools, including a personal AI operating-system kernel (8gent-code, Apache 2.0) and an AI environment for neurodivergent children (8gentjr). Both are free and local-first. This shapes our bias towards sovereign, open-weight, patient-owned AI over rented access to foreign frontier systems. The bias is stated, not hidden.

Section 2 — Near-term opportunities for patients (Committee question: most significant near-term opportunities)

  1. The strongest near-term opportunities are not in rare-disease diagnostics or high-cost oncology alone, though those are important. The strongest near-term opportunities are in the unglamorous, under-invested long tail: AAC (augmentative and alternative communication) for non-speaking patients; personalised mental-health triage for neurodivergent adults and children; personalised medication-review AI for polypharmacy in older patients; and AI-assisted care-pathway navigation for families with disabled children.
  2. These cohorts share a common feature. They are expensive to serve at population scale through traditional clinical labour. AI is cost-asymmetric in their favour. A sovereign NHS-governed model, trained on NHS data with informed consent, could serve them at a cost per interaction that no commercial market will ever match, because the commercial incentive does not point at these populations.
  3. The author's son is a non-speaking disabled child. In Ireland, the state's assistive-communication device pipeline is measured in years. A parent with five hours and a modest laptop can today build a more personalised AAC interface than the state will ship in a decade, using open-weight AI. This is not a theoretical near-term opportunity. It is a deliverable-this-year opportunity. The question is whether the NHS captures it or outsources it.

Section 3 — Major gaps in understanding (Committee question: where gaps exist)

  1. Gap one — data governance vocabulary. NHS and DHSC public materials consistently treat "data sharing" and "data sovereignty" as interchangeable. They are not. Data sharing is a permission problem. Data sovereignty is an ownership problem. The UK is losing the ownership question while winning some of the permission ones.
  2. Gap two — weight governance. Public discussion focuses on whether AI vendors can access NHS data. It does not focus on who owns the resulting model weights. A model trained on NHS data, owned by a foreign commercial vendor, and accessible only via their API is not a UK asset. It is a UK-subsidised foreign asset. This gap is not technical. It is contractual.
  3. Gap three — population priority. "Personalised medicine" is discussed in terms of genomics and oncology. The populations who would most benefit from AI-driven personalisation are arguably non-speaking patients, neurodivergent children and adults, those with complex comorbidities, and patients historically failed by the one-size-fits-all clinical pathway. They are largely absent from the framing.

Section 4 — What is needed to unlock opportunities (Committee question: what is needed)

  1.                     The NHS needs, in priority order:

(a) A UK-hosted, NHS-governed foundation model as a public good, in the same category as NHSmail or the NHS App.

(b) A procurement rule that prefers UK-sovereign AI where the task allows.

(c) A public register of every AI vendor shaping NHS-AI policy, disclosing commercial interest in the policy outcome.

(d) Ringfenced funding for AI tools targeting the populations described in paragraphs 5–6 above, on the principle that equitable AI cannot be left to commercial gradient.

(e) A patient-sovereignty layer: NHS-generated AI outputs returnable to the patient in portable form.

Section 5 — Role of AI in accelerating development and reducing cost (Committee question: role of AI in accelerating and reducing cost)

  1.                     AI's greatest cost-reduction potential in personalised medicine is not in drug discovery, where cost reduction is already being aggressively pursued by well-capitalised commercial actors. It is in the delivery phase, where AI can dramatically lower the marginal cost of personalisation at the point of care.
  2.                     A single open-weight model, hosted on UK infrastructure, could support: (a) medication-review AI across every GP in the country, (b) personalised-communication support for non-speaking patients, (c) care-plan translation and adaptation for families, (d) tailored follow-up for post-discharge patients, (e) mental-health check-ins for patients on long waiting lists. None of this requires frontier-scale intelligence. All of it requires sovereignty.
  3.                     The risk is not that the NHS fails to adopt AI. The risk is that the NHS adopts AI at a cost structure that is permanently rent-to-vendor rather than capital-owned.

13a. The K-shaped cut, applied to healthcare AI. As frontier model capability compounds exponentially, healthcare AI will not improve uniformly across populations. Capability gains will concentrate on the cohorts that commercial markets already serve (working-age adults, high-prevalence conditions, high-revenue diagnostic specialties) and will structurally bypass the cohorts commercial markets deprioritise (non-speaking patients, neurodivergent children and adults, rare diseases, severe mental illness, complex polypharmacy in older patients, care-pathway navigation for families of disabled children). The Sovereign AI Unit's first cohort confirms this gradient: Callosum, Prima Mente, Cosine, Cursive, Doubleword, Twig Bio, Odyssey, and the OpenBind Consortium are strong upstream-of-clinic choices for national capability, and none of them serve the delivery-phase cohorts above. Without a public-utility layer, the gap between served and unserved populations widens every time the underlying models get more capable. That is the cut: the upper arm of the K accelerates while the lower arm is severed from the benefit curve. Data sovereignty and serve-the-unserved-first are not two separate policies. They are the same policy: the NHS must own the infrastructure that serves the populations no commercial frontier vendor will ever serve at viable unit economics.

Section 6 — Where AI tools can be most effective (Committee question: most effective deployment)

  1.                     AI is most effective in personalised medicine where three conditions hold simultaneously: (a) the task has high variation across patients, (b) human clinical labour is the scarce resource, (c) errors can be caught and corrected by the clinician in the loop.
  2.                     By that test, the highest-value deployment targets are: AAC and assistive communication; medication review; care-pathway navigation for complex patients; personalised mental-health triage; and longitudinal follow-up. Diagnostic imaging, which receives the most public attention, is actually lower priority on criterion (b): radiology is already automatable at scale with narrow AI, and the bottleneck is commissioning, not development.

Section 7 — NHS digital infrastructure barriers (Committee question: NHS digital/IT barriers)

  1.                     The structural barrier is not the EHR, the interoperability standard, or the trust-level IT capacity. It is the commercial architecture underneath. An NHS trust choosing between a foreign closed-weight vendor and a UK-governed open-weight alternative is currently not choosing at all: the sovereign option does not yet exist at the scale required.
  2.                     Recommendation: treat the sovereign NHS foundation model as the digital infrastructure. Build it, and the trust-level deployment questions resolve downstream. Fail to build it, and every other digital-infrastructure decision locks the NHS into permanent rent-payer status.

Section 8 — Research-to-clinic and regulatory framework (Committee question: translation, MHRA, NICE)

  1.                     The UK's translational weakness in personalised medicine is not primarily regulatory. It is procurement-side. MHRA and NICE are not blocking deployment at the rate commonly assumed. What blocks deployment is that NHS England buys through commercial vendor frameworks that price out small UK AI developers who cannot afford framework compliance costs.
  2.                     Recommendation: a dedicated NHS AI procurement route for UK-governed open-weight systems, with framework costs waived or subsidised for systems that meet sovereignty and open-licensing criteria.

Section 9 — Good innovation-adoption practices (Committee question: examples)

  1.                     We defer to clinical-sector witnesses on specific NHS case studies. Our observation, from adjacent systems internationally, is that the most durable adoption patterns share three traits: (a) clinician ownership of the tool's training feedback, (b) patient ownership of the resulting data, (c) open-licensing of the model weights so that the tool survives vendor exit. All three traits are currently rare in NHS AI procurement.

Section 10 — Strategic question: strengthening the feedback loop (Committee question: Government role in research/industry/NHS feedback loop)

  1.                     The feedback loop is weakened by one structural feature: the insights generated by NHS-hosted data flow out to commercial vendors under partnership contracts, and flow back to the NHS only in the form of vendor products the NHS then pays to use. The NHS becomes both the data source and the eventual customer, with the margin extracted in the middle.
  2.                     Closing this loop does not require new research funding alone. It requires asserting that the intelligence generated from NHS-held data is itself an NHS asset, not a commercial by-product. That is a contractual and legislative move, not a research-grant move.
  3.                     The Spalding Paradox, applied to the NHS: the entities with the largest commercial interest in keeping the NHS AI conversation framed around "access to foreign frontier models" are the frontier-model vendors themselves. They are well-resourced, well-represented in UK policy forums, and broadly persuasive. The countervailing voices, patient-rights groups, disability-rights groups, open-source AI developers, small UK AI firms, are structurally under-resourced. This Committee is well placed to correct that imbalance by actively seeking evidence from those voices, not only from the commercial sector.

Section 11 — Final recommendation to the Committee

  1.                     Our single headline recommendation, if the Committee chooses to adopt only one, is this: recommend that Government extend the Sovereign AI Unit's mandate, announced 16 April 2026, to include a public-utility NHS-governed foundation model. Preference for open-weight architecture, UK-hosted inference, NHS-consented training data, and explicit population-priority commissioning (serving the cohorts named in paragraphs 4-6 and 13a ahead of, not after, the high-revenue cohorts). The terms of reference should be written by a panel that includes patient-rights, disability-rights, and open-source-AI voices alongside industry, not industry alone. The Unit as currently designed funds the upstream of personalised medicine; the NHS, as a public system, requires the downstream. Both are needed. Only one is currently in motion.

24a. Public AI as a public service. The NHS foundation model recommended above is the clinical instance of a broader principle the Committee is well placed to name: public AI is a public service, not a consumer product. The systems that will increasingly mediate how citizens make decisions about health, welfare, education, benefits, and rights should not be owned by a small number of foreign commercial vendors whose values, priorities, and risk appetites are shaped by their markets, not by the British public. Every small decision handed off to a commercial frontier model (what counts as urgent, what counts as a symptom worth escalating, what counts as an acceptable side effect, what counts as appropriate advice) is, at scale, a delegation of civic judgement to a non-civic actor. Treating public AI as public service, in the lineage of the BBC, the NHS, Royal Mail, and the National Grid, is the UK's most direct route to keeping democratic control of the reasoning infrastructure of the next century. Failing to treat it that way is not neutrality. It is a default-vote for private capture of how the public reasons.

  1.                     Everything else we have written is downstream of those decisions.

 

Author note

James Spalding is the founder of the 8GI Foundation, an international civil-society organisation focused on AI governance, sovereignty, and access. He is a full-stack engineer, an AuDHD self-identified parent of a non-speaking disabled child, and an independent researcher publishing work on AI governance capture (the "Spalding Paradox"). He is not a UK resident. He is available to the Committee for oral evidence if helpful.

 

Submitted in good faith. 8GI Foundation does not receive funding from any AI vendor, NHS contractor, or pharmaceutical company. The author has no undisclosed commercial interest in the outcome of this inquiry.