BIP America

collapse
Home / Daily News Analysis / UK can’t rely on AI that could be switched off, says Turing head

UK can’t rely on AI that could be switched off, says Turing head

Oct 11, 2026  Twila Rosenbaum  72 views
UK can’t rely on AI that could be switched off, says Turing head

Key facts at a glance

  • George Williamson, new head of the UK's national AI institute, says the UK must not depend solely on foreign AI systems that can be restricted or switched off.
  • He warns that political and commercial pressures could interrupt access to top models, citing Washington's temporary block on foreign access to Anthropic's leading systems in June.
  • Defence, national security, critical infrastructure and healthcare are areas where AI is now embedded.
  • Williamson wants the UK to have some AI it does not need anyone's permission to use.
  • The institute cannot match big AI labs' budgets, but it could build smaller, nationally owned models for start-ups.
  • National resilience will be its main focus, from protecting critical systems to spotting state-backed disinformation.
  • He is sceptical of claims that AI will wipe out humanity, saying he has not seen convincing evidence behind the quoted percentages.
  • European Central Bank President Christine Lagarde has similarly warned that Europe risks being cut off from US and Chinese AI.

The UK must not become wholly dependent on foreign artificial intelligence systems that could be restricted, inspected by others, or switched off at short notice, according to George Williamson, the newly appointed head of the country's national institute for AI and data science. In his first major intervention on AI policy, Williamson warned that relying solely on models built and controlled elsewhere leaves the UK exposed to political and commercial shocks. He said the country should have some AI capability that it does not need anyone else's permission to use.

His comments come after Washington temporarily blocked foreign access to Anthropic's most advanced models in June, a move that underscored how quickly access to frontier AI can become a geopolitical bargaining chip. Williamson pointed to defence, national security, critical infrastructure and healthcare as areas where AI is now deeply embedded, meaning any interruption could have serious consequences.

Why sovereign AI has moved up the agenda

For years, the dominant assumption in many capitals was that AI would be developed by a handful of large laboratories, mostly in the United States and China, and that other countries could simply consume the resulting services. That assumption is now being tested. Export controls, national security reviews, cloud restrictions and corporate terms of service can all determine whether a government, hospital, bank or defence contractor can use a particular model. The result is a growing debate about AI sovereignty: the ability of a country to access, inspect, control and deploy AI systems without depending on decisions made elsewhere.

Williamson's argument is not that the UK should build everything from scratch. He acknowledged that the national institute cannot match the budgets of the largest AI labs. Instead, he suggested a more targeted approach: smaller, nationally owned models that can support start-ups, public services and specialised applications. That approach would not try to copy frontier labs but would aim to provide a resilient base layer for areas where control and accountability matter most.

The Anthropic access episode

The temporary US block on foreign access to Anthropic's top models was a turning point in the debate. It showed that even commercial AI services can be withdrawn or limited for reasons that have little to do with the user's needs. Williamson said both political and commercial pressures mean the UK may not have access in all circumstances when it wants it. That is a blunt assessment of the risks of depending on foreign infrastructure. It also explains why he believes the UK should have some AI it does not need anyone's permission to use.

Anthropic is not unique. Other leading AI developers are subject to export controls, sanctions rules and national security laws. Cloud providers can suspend accounts. Model providers can change terms. Governments can impose restrictions on weights, APIs or chips. In a crisis, these levers could be pulled quickly. For critical services, that is a systemic risk.

National resilience as the new mission

Williamson took over the institute this year after running the government unit that builds technology for UK spy agencies. His appointment signalled a change of direction. The government, which is the institute's main funder, had demanded a sharper focus on national resilience. Williamson has now made that the organisation's central mission, covering everything from guarding critical systems to spotting state-backed disinformation.

That shift matters because AI resilience is not only about having a model. It is also about understanding where models come from, what data they were trained on, how they can be audited, and what happens if a supplier changes its terms. For defence and intelligence, the ability to inspect and control systems is essential. For healthcare, patient safety and data protection depend on knowing how decisions are made. For critical infrastructure, an outage or a malicious update could cascade across energy, transport or communications.

The institute's new focus also reflects a broader recognition that AI is no longer a niche research topic. It is embedded in public services, security operations and economic planning. That makes resilience a whole-of-government problem, not just a technical one. It requires procurement rules, testing regimes, skills pipelines and international partnerships that can survive political disruption.

From spy gadgets to AI policy

Williamson's background gives him an unusual perspective. He previously led the government unit responsible for building gadgets for UK spy agencies, a role that required balancing secrecy, operational reliability and rapid technological change. That experience appears to inform his approach to AI: he is less interested in hype cycles than in systems that work when they are needed, and that can be trusted by the people who depend on them.

He has also been sceptical of some of the more apocalyptic claims about AI. He said he has not seen convincing evidence behind the percentages some in the industry quote for the risk of AI wiping out humanity. That does not mean he dismisses all risk. It means his priority is nearer-term resilience: disinformation, critical system failure, dependency and strategic vulnerability. Those are concrete problems that governments can act on now.

Smaller national models versus frontier labs

The idea of building smaller, nationally owned models is gaining traction in several countries. These models may not match the raw capabilities of the largest frontier systems, but they can be fine-tuned for specific languages, sectors and legal frameworks. They can be hosted on national infrastructure, audited by domestic authorities and used by start-ups that need a reliable base. For a country like the UK, that could mean models specialised for healthcare triage, defence logistics, public administration or cybersecurity.

Williamson said the institute can be ambitious without trying to copy what frontier labs are doing. That is a pragmatic position. Frontier AI requires enormous compute, data and talent. A smaller country cannot easily match the largest players, but it can build capability in areas where sovereignty matters most. It can also create standards, testing methods and assurance frameworks that make foreign models safer to use when they are available.

The challenge is funding. Large AI labs spend billions on training runs and talent. Public research institutes operate on far smaller budgets. To make national models work, governments may need to combine public funding with private investment, shared infrastructure and procurement commitments. They may also need to cooperate with allies on compute, data and standards while retaining control over critical applications.

Defence, health and critical infrastructure

Williamson identified defence, national security, critical infrastructure and healthcare as areas where AI is now built in. Each has different requirements. Defence and national security need secure supply chains, cleared personnel and systems that can operate in contested environments. Critical infrastructure needs high availability, strong cybersecurity and resilience to supply chain shocks. Healthcare needs clinical validation, privacy safeguards and accountability when AI supports diagnosis or treatment planning.

In all these areas, the risk of relying on a foreign AI service that could be switched off is not abstract. A hospital might lose access to a diagnostic tool. A utility might lose a predictive maintenance model. A defence agency might lose a data-analysis platform. Even if the probability seems low, the impact could be severe. That is why Williamson argues for some nationally controlled AI capability as a backstop.

Disinformation and state-backed threats

National resilience also includes the information domain. State-backed disinformation campaigns increasingly use AI to generate persuasive text, images, audio and video at scale. Detecting and countering these campaigns requires AI tools that are trusted, transparent and under domestic control. If the detection tools themselves depend on foreign platforms, governments may be vulnerable to the same restrictions they are trying to guard against.

The institute's focus on spotting state-backed disinformation is therefore closely linked to its wider resilience mission. It is not only about identifying false content. It is about understanding networks, narratives and influence operations, and doing so with systems that can be audited and adapted to local languages and political contexts.

European parallels

Williamson's warning echoes concerns raised by European leaders. Christine Lagarde, president of the European Central Bank, has warned that Europe risks being cut off from US and Chinese AI. That warning reflects a broader anxiety about digital dependency. Europe has strong research institutions and industrial base, but it relies heavily on foreign cloud providers, chips and AI models. The same is true for the UK.

The debate is not about autarky. No country can build every component of the AI stack alone. But governments are increasingly asking which parts of the stack are too important to leave entirely to foreign suppliers. For some, the answer is chips. For others, it is cloud infrastructure, data, models or standards. Williamson's focus on models and national resilience adds a new dimension to that conversation.

What a sovereign AI stack could look like

A sovereign AI stack would not be a single product. It would be a layered capability. At the base, it would include secure compute and data infrastructure. Above that, it would include models that can be inspected, fine-tuned and deployed under national rules. Around it, it would include testing, assurance and auditing tools. And at the top, it would include applications for public services, defence and industry.

For the UK, a realistic approach might involve several elements. First, public compute facilities for research and start-ups. Second, smaller open or nationally controlled models for priority sectors. Third, procurement rules that require suppliers to provide auditability and continuity guarantees. Fourth, international partnerships that secure access to frontier models where possible, while building fallback options. Fifth, skills and standards to ensure that domestic institutions can evaluate and govern AI systems.

The institute's role could be to coordinate research, set benchmarks and help translate national priorities into technical programmes. It could also act as a trusted broker between government, academia and industry. That would not replace the private sector, but it could reduce the risk of strategic dependency.

Funding and governance challenges

Building national AI capability is expensive. Even smaller models require compute, data, engineering talent and maintenance. Governments must decide how much to invest, where to focus and how to avoid duplicating work done by allies. They must also ensure that public investment does not crowd out private innovation or create slow, uncompetitive national champions.

Governance is another challenge. National models need clear rules on data use, security, transparency and liability. If they are used in healthcare or defence, they need rigorous testing. If they are used in public services, they need accountability. If they are made available to start-ups, they need licensing and support. These are not simple tasks, but they are manageable with the right institutions and policies.

Implications for start-ups and public services

For start-ups, access to reliable AI infrastructure can be a make-or-break issue. Many small companies depend on foreign APIs that can change price, terms or availability at any time. A nationally owned model, even if less capable, could offer a stable alternative for certain applications. It could also help start-ups meet data residency and security requirements when selling to government or regulated industries.

For public services, the benefits could be even more direct. Hospitals, police forces, local councils and transport authorities need AI tools that are compliant, auditable and resilient. If those tools depend on foreign providers that can be restricted, public services face operational risk. A national capability could provide a backstop and a testing ground for new applications.

None of this means abandoning foreign AI. Frontier models will remain important for many tasks. The argument is about balance. The UK can use foreign AI where it is safe and beneficial, while retaining the ability to operate independently in areas where control is essential. That is the essence of Williamson's warning.

His comments also highlight a wider shift in the global AI debate. After a period of rapid adoption and experimentation, governments are beginning to ask harder questions about dependency, security and resilience. The answers will shape not only national AI strategies but also international trade, defence alliances and digital rights. Lagarde's warning that Europe could be cut off from US and Chinese AI shows that the UK is not alone in asking these questions. The next phase of AI policy will be defined by how countries manage the tension between access and autonomy.


Source: TNW | Artificial-intelligence News


Share:

Your experience on this site will be improved by allowing cookies Cookie Policy