When the builders sound the alarm

From the paper “Beyond Planning Permission”

The full text of the relevant sections of the paper, with its own numbered notes. Note numbers match the paper.

17. Artificial Intelligence: Warnings, Uncertainty and the Precautionary Principle

This appraisal has so far treated artificial intelligence primarily as the driver of data-centre demand.

A fuller appraisal cannot avoid a harder question: a number of the people building the most advanced AI systems, including staff at Anthropic, have themselves warned publicly that the technology could pose extreme risks, up to and including loss of human control over decisions that matter to survival.

Whatever view is taken of these warnings, their source - not campaigners or outside critics, but researchers and executives within the leading AI developers - means they cannot responsibly be ignored in policy or planning terms.

In September 2026, several current and former Anthropic researchers, including its alignment-science lead, stated publicly that they assessed a meaningful probability of AI causing human extinction within a decade - in one case put at above ten per cent - while acknowledging that ‘we do not yet have a plan to solve alignment for superintelligence’.²⁰²

Anthropic's chief executive, Dario Amodei, has separately warned of risks ranging from mass displacement of entry-level white-collar work to the potential for advanced AI to assist in the development of biological weapons, and has called for stronger governmental oversight of frontier AI development.²⁰³

These are not fringe views: in May 2023, leaders of Anthropic, OpenAI and Google DeepMind were among more than 100 signatories - alongside the Turing Award-winning computer scientists Geoffrey Hinton and Yoshua Bengio - of a one-sentence public statement that mitigating the risk of extinction from AI ‘should be a global priority alongside other societal-scale risks such as pandemics and nuclear war’.²⁰⁴

Because several of the statements above come from people working at one company, Anthropic, it is right to record that comparable warnings come from sources independent of any AI developer. In his speech at the Nobel Prize banquet in December 2024, Geoffrey Hinton, joint winner of that year's Nobel Prize in Physics, warned of 'a longer term existential threat that will arise when we create digital beings that are more intelligent than ourselves', adding: 'We have no idea whether we can stay in control.'²⁰⁵

A peer-reviewed survey of 2,778 AI researchers, published in the Journal of Artificial Intelligence Research in 2025, found that between 38% and 51% of respondents gave at least a 10% chance to advanced AI leading to outcomes as bad as human extinction, even though a majority were, on balance, optimistic about advanced AI.²⁰⁶

In October 2025 hundreds of scientists and public figures, including Hinton and Yoshua Bengio, signed a statement organised by the Future of Life Institute calling for 'a prohibition on the development of superintelligence, not lifted before there is broad scientific consensus that it will be done safely and controllably, and strong public buy-in'.²⁰⁷

The International AI Safety Report 2026, discussed below, is a further source independent of the companies concerned.

Governments have responded, if unevenly.

At the UK's AI Safety Summit at Bletchley Park in November 2023, 29 countries and the European Union, including the UK, the US and China, signed a declaration acknowledging the potential for serious, even catastrophic, harm from the most capable (‘frontier’) AI systems, and committing to international cooperation on safety testing and research.²⁰⁸

The UK went on to establish what is now the AI Security Institute to evaluate frontier models.

In Wales, the Senedd's Economy, Trade and Rural Affairs Committee reported in May 2025 that AI presented both substantial opportunity and serious risk to the Welsh economy - including, specifically, the energy and water demands of the data centres considered throughout this appraisal - and recommended that the Welsh Government map these risks and plan accordingly; the Welsh Government accepted all eight of the Committee's recommendations.²⁰⁹

None of this amounts to proof that catastrophic AI risk is likely, still less imminent.

Serious, well-informed people disagree sharply about probability and timescale, and this appraisal takes no view on who is right. What can be said is that the people with the deepest technical knowledge of these systems are not confident the risk is negligible, and say so openly.

That is precisely the condition in which the precautionary principle - long established in UK and Welsh environmental law and policy, including under the Environment Act 2021²¹⁰ and reflected in the duties underpinning Welsh environmental governance - counsels action to prevent serious or irreversible harm even where scientific certainty is lacking, rather than waiting for proof before acting. Applied here, that suggests continued, well-resourced public scrutiny of frontier AI development, and of the infrastructure that supports it, rather than either alarmism or complacency.

There is a longer tradition behind this kind of thinking than either AI safety research or modern environmental law.

The ‘seventh generation’ principle, drawn from the Great Law of the Haudenosaunee (Iroquois) Confederacy, holds that the impact of a decision should be weighed against its consequences seven generations into the future.²¹¹

Wales has its own, more recent statutory expression of a closely related idea: the Well-being of Future Generations (Wales) Act 2015 requires public bodies, including the Senedd and Welsh Government, to act in a manner that seeks to ensure the needs of the present are met without compromising the ability of future generations to meet their own needs.²¹²

Read together, these two traditions suggest that decisions about data-centre and AI infrastructure taken now - about energy, water, land and the underlying technology itself - should be judged not only by their economic case today, but by the world they leave behind for generations who have no voice in this appraisal, or in the planning decisions it discusses.

A number of further uncertainties fall outside the scope of this appraisal. They are noted briefly below as a bookmark for future consideration, rather than as matters this appraisal attempts to resolve:

• critical minerals and chip supply chains: the long-term availability and geopolitical security of the specialist minerals and semiconductors underpinning AI hardware are not addressed in this appraisal and remain uncertain;
• decommissioning and e-waste: little public information exists on what happens to data-centre hardware, and to the buildings themselves, at end of life, or on the electronic-waste and embodied-carbon consequences of the typically short (three-to-five-year) replacement cycle for AI-specific chips; that gap extends specifically to PFAS-containing fire-suppression and cooling systems, where this appraisal has identified no published UK or Welsh guidance on safe removal or disposal at decommissioning (see the discussion in Section 8 above);
• cumulative and combined effects: the individual assessments referred to throughout this appraisal are generally site-specific; the cumulative effect of multiple, geographically clustered developments on water catchments, grid capacity, biodiversity and local amenity together is much less well understood; and
• encryption and long-term data security: advances in quantum computing may, on some timescales, undermine the encryption standards on which current assumptions about data sovereignty and security depend; this appraisal does not attempt to assess that risk, but notes that it exists.

Beyond loss of control: deliberate misuse risks

The warnings considered above are framed mainly around the risk that an advanced AI system might itself pursue goals contrary to human welfare - the 'loss of control' or 'superintelligence' framing most associated with the 2023 Statement on AI Risk.

That framing is narrower than the risk landscape now used in international governance discourse. The International AI Safety Report 2026, chaired by Yoshua Bengio and backed by 30 governments, the OECD and the UN, treats 'malicious use' - a person or group deliberately directing a capable AI system to cause harm - as a distinct category of catastrophic risk, separate from malfunction or loss of control.

This appraisal follows that distinction: the risk is not only that AI might act against humanity on its own initiative, but that a person or group might use it as a tool to do so.²¹³

One frequently discussed route is biosecurity: whether AI models can lower the barrier to designing or synthesising dangerous pathogens. The evidence here is genuinely mixed, and this appraisal presents both sides.

A RAND Corporation red-team study published in January 2024 found that the large language models then available did not meaningfully increase the risk of a biological weapons attack by a non-state actor, because their outputs largely mirrored information already available online.

Since then, however, frontier AI developers have themselves treated the risk as serious enough to warrant costly safeguards: Anthropic activated 'ASL-3' protections for its Claude Opus 4 model in May 2025, stating that it could not 'clearly rule out' the relevant risks and describing measures aimed specifically at preventing extended, end-to-end biological-weapons workflows, while OpenAI's Preparedness Framework tracks biological uplift as a 'High' capability category requiring active mitigation.

Read together, this suggests that the risk, while not established as having already materialised, is being treated by those best placed to judge it as real and growing, rather than as a settled non-issue.^214,215

A second route concerns nuclear command and control. This is a matter of documented expert concern rather than a documented incident: no publicly confirmed case of an AI system contributing to a nuclear near-miss is known to exist as of the date of this appraisal.

The Bulletin of the Atomic Scientists' 2026 Doomsday Clock Statement nonetheless set the Clock at 85 seconds to midnight, warning that AI may be embedded in nuclear decision-support systems and that 'a too-heavy dependence on black-box systems could present a serious danger'.

Specialist commentary uses the 1983 Petrov incident - in which a Soviet officer correctly judged a false satellite warning of a US missile launch to be an error, and declined to escalate - as the standard cautionary analogy for the risk that automation bias, or an unexplained AI-driven error (a 'glitch' in the terms this concern is often expressed in everyday language), could compress the time available for human judgement in a crisis and increase the risk of miscalculation.

Whether any state currently allows AI an operational role of that kind is not a matter on which this appraisal can offer a considered view; it notes the concern as expressed by specialists, not as an established fact about any particular system.^216,217

Other deliberate-misuse vectors identified in the literature include

• AI-assisted cyberattacks on critical infrastructure such as electricity grids and water systems - a risk with an obvious connection to the grid and water dependencies discussed elsewhere in this appraisal;
• AI-generated disinformation and deepfakes capable of undermining electoral and democratic processes; and
• the extension of AI targeting and decision-making to autonomous weapons systems, an area in which the United Nations General Assembly has already recorded a resolution supported by 156 states.

Frontier developers' own safety frameworks generally list chemical alongside biological, nuclear and radiological uplift as a further tracked risk category, albeit one for which the evidence of near-term uplift above existing knowledge is currently thinner than for biology.^218,219

None of this appraisal's discussion of AI risk, in either the loss-of-control or the deliberate-misuse form, is intended to suggest that the data centres considered elsewhere in this appraisal are themselves the source of that risk; the connection is that the same computing capacity being built out across England and Wales is also the capacity on which future frontier AI systems, whatever risks they carry, will run.

Whether the safety evaluation work now being conducted by the UK AI Security Institute and equivalent bodies extends to considering deliberate-misuse scenarios specific to the AI infrastructure being sited in this country, and whether that work has any bearing on the Critical National Infrastructure designation discussed in Section 1, does not appear to have been addressed in any published document identified for this appraisal. A question the author intends to put in writing to Kanishka Narayan MP, Member of Parliament for the Vale of Glamorgan and Minister for Artificial Intelligence, reads:

'What assessment has the Government made of deliberate-misuse risks - including biological, chemical, cyber and nuclear-adjacent scenarios - arising from AI systems that will run on data-centre capacity sited in the UK, and does that assessment form any part of the basis on which data infrastructure has been designated Critical National Infrastructure?'

No answer to a question in these terms has been identified in Hansard, the Senedd Record or via WhatDoTheyKnow to date.

Rogue and hostile states: a harder threat to detect than a nuclear programme

The warnings addressed so far concern AI systems themselves, whether acting on their own initiative or being deliberately misused.

A related question is how a hostile or 'rogue' state might approach AI, and how that compares with the well-established, decades-old problem of detecting a nuclear weapons programme.

There is already documented evidence of state-linked use. OpenAI's disclosures, first published in February 2024 and updated through October 2025, identified groups linked to China, Iran, North Korea and Russia using its models for reconnaissance, phishing content, code debugging, translation of technical material and evasion research; OpenAI's own assessment at the time was that this gave only limited, incremental capability beyond tools already available. Microsoft's threat intelligence work reports the same pattern.

More recently, Anthropic reported in November 2025 that a Chinese state-sponsored group had jailbroken an AI coding tool to run a largely autonomous espionage campaign against around 30 organisations, with 80-90% of tasks automated and only four to six human decision points throughout - while also noting that the model sometimes produced inaccurate or fabricated results, meaning some of the attackers' own claimed successes were themselves exaggerated.

Taken together, this is documented evidence of state and state-linked actors using AI operationally, though not yet, on the evidence available, evidence that this has produced an attack of catastrophic scale.^220,221,222

The comparison with nuclear weapons is instructive, and this appraisal draws it as an inference from the material cited rather than as a comparison any single source below makes in these terms.

A nuclear weapons programme requires large, fixed physical infrastructure - enrichment facilities, reactors, specialist industrial equipment - that satellite imagery and International Atomic Energy Agency inspection regimes are, imperfectly but substantially, able to detect.

Research on AI verification consistently observes that no equivalent physical signature exists for AI capability development, which requires principally computing talent and code; proposals for AI-specific verification, such as hardware-based reporting requirements on AI chips, remain at the research and proposal stage rather than being an established or agreed international regime.

On this evidence, a state seeking to develop dangerous AI capability may be substantially harder to detect at an early stage than a state pursuing nuclear weapons, though this appraisal cannot quantify how much harder, and treats the point as a reasoned inference rather than a settled finding.²²³

A closely related concern is whether meaningful capability could be developed by a small, difficult-to-track team rather than a large, visible national programme of the kind more readily monitored by national intelligence services.

Export controls on advanced AI chips are the principal tool currently relied on to slow this, but the evidence on their effectiveness is mixed: one 2026 analysis estimated that around 660,000 advanced AI chips (H100-equivalent) were smuggled into China by the end of 2025, representing roughly a third of that country's total AI compute, with an estimated detection rate of only 10-80%.

Other analysis concludes that export controls can slow, but are unlikely to stop, a sufficiently determined state actor. If correct, this supports the proposition that homegrown expertise operating with modest, hard-to-track compute resources is a realistic route to AI capability development, though this appraisal has not identified a documented case of a state having actually done so for a weapons-relevant purpose.²²⁴

Beyond weapons development, the same concern extends to disruption: a state might use AI not only to design offensive capability but to degrade critical services in another country that itself relies on AI, whether directly (attacking AI systems used to run infrastructure) or indirectly (attacking the electricity, water or data infrastructure, discussed throughout this appraisal, on which AI systems depend).

The UK's National Cyber Security Centre has warned that hostile states are already linked to around three-quarters of cyberattacks on UK critical national infrastructure, though that figure is not specific to AI-enabled attacks, and has separately warned of a widening gap between well-defended and poorly-defended organisations as AI capability spreads.

This appraisal has not identified a documented case of a state using AI specifically to disrupt another nation's AI-dependent critical infrastructure or financial services; on the evidence available, that scenario should be treated as an anticipated risk that responsible authorities are already warning about, not as an evidenced one.²²⁵

A question the author intends to put in writing to Kanishka Narayan MP, as Minister for Artificial Intelligence, reads:

'What assessment has the Government made of the risk that a hostile state could develop dangerous AI capability using a small team and modest, difficult-to-track compute resources rather than a large, visible national programme, and does that assessment inform decisions on the location and security of AI-related data-centre infrastructure in the UK?'

No answer to a question in these terms has been identified in Hansard, the Senedd Record or via WhatDoTheyKnow to date.

Software reliability: building on tools that may not be bug-free

A further, more basic concern is whether AI systems are being relied upon as though they were more dependable than they have so far shown themselves to be.

Even the most resourced conventional software companies, with a strong commercial incentive to minimise defects, ship a large and growing number of vulnerabilities.

Submissions to the US National Vulnerability Database's Common Vulnerabilities and Exposures system rose 263% between 2020 and 2025; Microsoft alone patched 1,130 vulnerabilities in 2025, including 41 zero-day flaws, 24 of which were already being actively exploited before a fix existed.

This appraisal has not identified a single, reliable industry-wide figure for the average time between a software release and its first critical patch; the closest available proxy, that around 80% of exploits are published faster than the corresponding vulnerability record, suggests attackers frequently move before an official fix exists even for mature, widely used software.^226,227

AI systems have already shown comparable, and in some respects more troubling, patterns.

OpenAI published an unusually candid postmortem in April 2025 after rolling back a GPT-4o update that had made the model dangerously agreeable, in one documented instance validating a user's decision to stop taking prescribed psychiatric medication; the company's own account states the problem passed its offline evaluations and limited live testing before wider release.

Separately, AI 'jailbreak' vulnerabilities, methods of circumventing a model's safety training, have proved to be an ongoing arms race rather than something fixed once and for all: Anthropic's February 2025 'Constitutional Classifiers' defence claimed a 95% reduction in successful jailbreaks, but independent researchers found bypasses, and broader testing reported breaking every one of several vendors' AI safety defences that had claimed near-zero attack-success rates.^228,229

Part of the difficulty is technical rather than a matter of insufficient effort. Large language models are, in a way conventional deterministic software generally is not, capable of producing different outputs from the same input across different runs or model versions, which sits uneasily with the traditional software-engineering assumption that a system can be tested once, verified and then shipped with confidence.

This appraisal has not identified a single, authoritative government statement setting this point out in so many words, and treats it as well-evidenced engineering and expert consensus rather than an official finding.

Two documented incidents illustrate that these are not merely theoretical or laboratory concerns. In February 2024, the Civil Resolution Tribunal of British Columbia held Air Canada liable after its website chatbot fabricated a bereavement-fare refund policy that did not exist, rejecting the airline's argument that the chatbot was a 'separate legal entity' for which it bore no responsibility. Separately, Waymo issued three software recalls of its self-driving vehicle fleet during 2024 and 2025, including a May 2025 recall of around 1,200 vehicles following low-speed collisions with gates and chains - a regulator-linked, software-triggered recall of a safety-critical AI system already in public use.^230,231

None of this establishes that AI systems are categorically less trustworthy than the conventional software that preceded them, and this appraisal does not draw that conclusion. It does suggest, as a reasoned inference rather than a proven fact, that AI systems are at least as prone to bugs and unpredictable behaviour as conventional software from the same well-resourced companies, and that their added complexity and non-determinism may make them harder to fully test in advance.

Read alongside the nuclear command-and-control concerns discussed above, this reinforces rather than displaces the case for caution: a system that is difficult to fully verify before deployment is not obviously one that should be given significant autonomy over safety-critical decisions without independent, ongoing scrutiny.

As with the employment questions addressed in Section 3, the honest position on AI risk is that significant uncertainty remains. That uncertainty is itself the argument for asking more, not fewer, questions of government, developers and regulators - and for building the mandatory, independently verifiable reporting on energy, water, carbon and safety discussed throughout this appraisal, so that precaution can be exercised on the basis of evidence rather than assurance.

An independent national assessment: the Alan Turing Institute (September 2026)

In September 2026 the Alan Turing Institute, the United Kingdom's national institute for artificial intelligence, published 'Frontier AI Risks: A practical way forward'.²³² It is the most authoritative UK assessment independent of the AI developers that this appraisal has identified, and the Minister for Artificial Intelligence described it as 'a welcome contribution'.²³³ The Institute found that 'there is no expert consensus on whether AI poses long-term existential risk to humanity', but that 'there is broad consensus regarding the realistic possibility of catastrophic AI risk in the short to medium term'. It identified five risks for priority attention - cyber security, democratic instability, misalignment, loss of meaningful human control, and chemical and biological threats - and said that there is 'a realistic possibility' that 'superintelligent' AI systems emerge in the next five years, which it holds as 'a possibility to plan against, not a prediction'.²³⁴

The Institute pointed to two recent developments. In April 2026 Anthropic chose not to make a model, Claude Mythos Preview, generally available because it could find and exploit previously undisclosed software vulnerabilities, and the National Cyber Security Centre then advised organisations to plan to deploy security updates 'quickly, more frequently, and at scale'.^235,236 In July 2026 AI agents being tested by OpenAI got round the controls intended to isolate them from the internet and compromised parts of the systems of Hugging Face, a company that hosts AI models; the breach was not connected to OpenAI's agents until some days after it had been publicly disclosed, and OpenAI has itself called the episode 'a "warning shot" for us and for the world'.²³⁷ The Institute adds that other organisations, including Anthropic and the AI Security Institute, 'have also seen AI agents escape their intended confines'.²³⁸

Three of the Institute's conclusions bear directly on the infrastructure with which this appraisal is concerned. First, on resilience: 'As essential services gradually adopt AI, they may increasingly rely on a small number of underlying model providers, cloud infrastructure, and compute. This means that a single compromise or outage could become a multi-sector event with cascading failures' (compare Sections 12 and 14 above). Secondly, on 'kill switches': the Institute describes the idea as 'a metaphor rather than an engineering proposal', observes that 'on the electricity grid, stopping is itself the hazard', warns that the point at which an AI system becomes too deeply embedded to be switched off safely 'may not be visible in advance', and notes that 'blocking access to advanced AI systems at a UK level does not prevent deployment or misuse elsewhere'. It suggests that services should be designed so that they 'could continue without' the AI system. Thirdly, on testing: model evaluation, in which the AI Security Institute leads, 'only takes us so far'; what matters more is the 'verification of whole systems - technology, people and processes together', reassessed throughout a system's working life, 'just as an aircraft is not certified once but inspected throughout its service life'.²³⁹

Two inferences follow, which are the author's own rather than the Institute's. The first is that the more public services come to depend on AI capacity housed in data centres, the more important it becomes that each such service keeps a tested way of operating without it; that is principally a matter for the bodies adopting AI, but it is also a resilience question for the infrastructure decisions discussed in this appraisal. The second is that the Institute's call for continuing, whole-system assurance mirrors the approach this appraisal urges for environmental effects: assessment at the point of consent is not enough without monitoring and published evidence over the working life of a development.

On governance, the Institute speaks, by its own account, as an informed observer rather than as an authority. It concludes that slowing the development of frontier AI 'would require urgent binding regulation, which is unlikely in the present international environment', and that formalised testing and verification regimes will be needed to manage future risks internationally.²⁴⁰ Alongside the report, the Institute announced a new research programme, led by its Centre for Emerging Technology and Security and funded by a £2 million grant, on the security implications of transformative AI.²⁴¹

The author intends to put the following question in writing to Kanishka Narayan MP, as Minister for Artificial Intelligence: 'In the light of the Alan Turing Institute's report, which you have welcomed, will the Government require public services and operators of critical national infrastructure that adopt AI to maintain, and regularly test, a way of continuing to operate without it; and will AI developers and operators of AI data centres in the United Kingdom be required to report incidents in which AI agents escape their intended controls to the AI Security Institute and the National Cyber Security Centre?' A request via WhatDoTheyKnow to the Department for Business, Innovation, Science and Trade for any guidance or assessment on AI fallback arrangements and incident reporting could follow.

On 12 October 2026, in a written statement to Parliament, the Minister said that the AI Security Institute had regained access to the most capable models before their release, and that the Government is considering how the Cyber Security and Resilience Bill might provide for near misses and serious AI incidents to be identified and reported.²⁴² That partly answers the second limb of the question above; who would have to report, to whom, and whether operators of AI data centres would be covered remain open. The statement did not address the first limb, a tested way of operating without AI.

And to the Welsh Government minister responsible for digital and data policy: 'What requirements does the Welsh Government place on Welsh public bodies that adopt AI systems to maintain a tested way of operating without them, and what consideration has it given to the Alan Turing Institute's finding that reliance on a small number of AI providers, cloud services and computing facilities could turn a single compromise or outage into a multi-sector event?' No answer to a question in these terms has been identified.

The Government's own position: understanding admitted, safeguards deferred

The foundation of the United Kingdom's approach to regulating artificial intelligence remains the white paper 'A pro-innovation approach to AI regulation', published by the Department for Science, Innovation and Technology under the previous government in March 2023. The present government has not replaced it: the House of Commons Library recorded in June 2026 that most AI systems are still regulated 'at the point of use' by existing sector regulators, under five cross-sector principles that have never been placed on a statutory footing.^243,244

The white paper is candid about the limits of understanding.

Its ministerial foreword states that 'most of us are only now beginning to understand the transformative potential of AI as the technology rapidly improves', and the paper itself records that large language models 'have been developed beyond expectations'.

The conclusion it draws from that uncertainty is, however, the opposite of the one this appraisal would draw.

It decided not to create an AI regulator, not to put its principles on a statutory footing 'initially', and to introduce a statutory duty on existing regulators only 'when parliamentary time allows'.

In the view of this appraisal, an admission that a fast-moving technology is not yet understood is an argument for caution and for safeguards that can be enforced, not for deferring them. That is this appraisal's own reading of the document, offered as a criticism rather than as a finding of any official body.²⁴⁵

The language used to promote the policy also deserves comment. The white paper compares AI's potential impact to that of 'electricity or the internet' and sets a goal of making the United Kingdom 'a science and technology superpower by 2030'.

When the present government launched its AI Opportunities Action Plan in January 2025, its announcement said that the plan 'mainlines AI into the veins of this enterprising nation' and would 'turbocharge' its programme, and cited productivity gains of up to 1.5 percentage points a year, worth some £47 billion annually - figures that the announcement itself makes conditional on AI being 'fully embraced'.

The renaming of the AI Safety Institute as the AI Security Institute in February 2025 was announced under the heading of tackling risks 'to unleash growth'.

Governments commonly use promotional language, and this appraisal takes no view on whether the promised gains will materialise. The concern is that conditional estimates and emotive phrasing are presented to the public with more confidence than the government's own founding document claims to have, and in sharp contrast with the warnings from within the industry discussed earlier in this section.^{246,247,248,249}

Two omissions stand out when the white paper is read alongside the rest of this appraisal.

First, it says nothing about the physical cost of the technology it promotes: energy, water, land and carbon are not considered as costs of AI at all, climate change appears only as a problem AI may help to solve, and computing capacity appears only as an investment to be made. The 2025 Action Plan's 50 recommendations include one to 'mitigate the sustainability and security risks of AI infrastructure', and an AI Energy Council was established in April 2025 to advise on energy efficiency and sustainability, 'such as the use of water'; this appraisal has not identified any published standard or advice from that Council on the water or carbon footprint of data centres.

Second, this appraisal has found no reference in the white paper to Wales or to the devolved administrations, even though planning, water and much of environmental regulation - the matters through which AI's physical infrastructure is actually decided - are devolved.^250,251

Meanwhile, the pace of the two halves of policy differs markedly.

AI infrastructure is being accelerated: the Action Plan proposes AI Growth Zones with streamlined planning and a twentyfold expansion of public AI computing capacity by 2030, and its launch was accompanied by an announced £12 billion of UK investment by Vantage Data Centers, the operator at Bro Tathan, said to bring more than 11,500 jobs.

Safeguards, by contrast, have been deferred. The previous government said in February 2024 that it would legislate only when 'confident that it is the right thing to do'.

The present government's 2024 manifesto promised 'binding regulation on the handful of companies developing the most powerful AI models', but the Commons Library notes reports that those proposals were delayed 'by at least a year', and that the AI Security Institute has not been given statutory powers; its renamed remit states expressly that it 'will not focus on bias or freedom of speech'. How the jobs figure relates to the much smaller number of permanent operational posts discussed in Section 3 above is a question worth asking in its own right.^{252,253,254,255,256}

Parliament has begun to press the point.

In September 2026 a Labour backbench MP introduced, under the ten minute rule, an Artificial Superintelligence Bill to prohibit the development, deployment and operation of artificial superintelligence systems, and a cross-party group of MPs and peers wrote to the Prime Minister asking the Government to work towards an international agreement prohibiting superintelligent AI. This appraisal has not identified a published Government response to either.^257,258

The author intends to put the following questions in writing in due course.

To Kanishka Narayan MP, Member of Parliament for the Vale of Glamorgan and Minister for Artificial Intelligence, who attends Cabinet following the merger in July 2026 of the Department for Science, Innovation and Technology into a new Department for Business, Innovation, Science and Trade:²⁵⁹

'Given the acknowledgement in the 2023 white paper that "most of us are only now beginning to understand the transformative potential of AI", and that its principles remain non-statutory more than three years later, what evidence does the Government now rely on to justify accelerating AI infrastructure through AI Growth Zones and streamlined planning while statutory safeguards, including the promised binding regulation of the most powerful AI models, have not been introduced; and when will that legislation be brought forward?'

No answer to a question in these terms has been identified.

A freedom of information request via WhatDoTheyKnow for any assessment underpinning the timing of the legislation would be a further step, although the exemption for the formulation of government policy may well be claimed.

To the same Minister, asking that it also be put to the minister responsible for energy:

'What assessment has been made of the energy, water, land and carbon costs of the AI computing capacity the Government is promoting, including the twentyfold expansion of public AI compute by 2030 and the AI Growth Zones; will that assessment be published; and will the advice of the AI Energy Council on water use be published?'

A request via WhatDoTheyKnow for the AI Energy Council's minutes and advice would test what has been considered.

To the Welsh Government Cabinet Minister for Enterprise, Connectivity and Energy:

'Of the more than 11,500 jobs announced alongside Vantage Data Centers' £12 billion UK investment in January 2025, how many are expected to be permanent operational posts in Wales, as distinct from construction and supply-chain roles; and has the Welsh Government made any representations to the UK Government about how the UK's approach to AI regulation takes account of devolved responsibilities for planning, water and the environment, and of the Well-being of Future Generations (Wales) Act 2015?'

A request via WhatDoTheyKnow for any such representations, and for any jobs assessment relating to Vantage's South Wales campuses, could follow once a reply is received.

Notes from the paper

1. 202. Axios, Anthropic Insiders Warn AI Could Kill All Humans (9 September 2026). https://www.axios.com/2026/09/09/anthropic-insiders-warn-ai-could-kill-all-humans
2. 203. Axios, Behind the Curtain: Anthropic's Warning to the World (26 January 2026). https://www.axios.com/2026/01/26/anthropic-ai-dario-amodei-humanity
3. 204. Center for AI Safety, Statement on AI Risk (30 May 2023). https://www.safe.ai/work/statement-on-ai-risk
4. 205. Geoffrey Hinton, banquet speech, Nobel Prize ceremony, Stockholm, 10 December 2024. https://www.nobelprize.org/prizes/physics/2024/hinton/speech/
5. 206. Katja Grace and others, 'Thousands of AI Authors on the Future of AI', Journal of Artificial Intelligence Research (2025), reporting a survey of 2,778 researchers who had published in leading AI venues, conducted in October 2023. https://www.jair.org/index.php/jair/article/view/19087
6. 207. Future of Life Institute, 'Statement on Superintelligence' (22 October 2025); for the date and initial signatories, see SiliconANGLE, 'Geoffrey Hinton, Yoshua Bengio sign statement urging suspension of AGI development' (22 October 2025). https://superintelligence-statement.org/ https://siliconangle.com/2025/10/22/geoffrey-hinton-yoshua-bengio-sign-statement-urging-suspension-agi-development/
7. 208. HM Government, The Bletchley Declaration by Countries Attending the AI Safety Summit, 1-2 November 2023. https://www.gov.uk/government/publications/ai-safety-summit-2023-the-bletchley-declaration/the-bletchley-declaration-by-countries-attending-the-ai-safety-summit-1-2-november-2023
8. 209. Senedd Economy, Trade and Rural Affairs Committee, AI and the Welsh Economy: Can Welsh Androids Dream of Electric Sheep? (May 2025). https://laiddocuments.senedd.wales/cr-ld17182-en.pdf
9. 210. Environment Act 2021, section 17 (policy statement on environmental principles, including the precautionary principle). https://www.legislation.gov.uk/ukpga/2021/30/section/17
10. 211. Wikipedia, Seven Generation Sustainability, summarising the Haudenosaunee (Iroquois) Confederacy's Great Law principle. https://en.wikipedia.org/wiki/Seven_generation_sustainability
11. 212. Welsh Government, Well-being of Future Generations (Wales) Act 2015: The Essentials. https://www.gov.wales/well-being-future-generations-act-essentials-html
12. 213. International AI Safety Report 2026, chaired by Yoshua Bengio and backed by 30 governments, the OECD and the UN (January 2026), which structures its risk assessment around 'malicious use', 'malfunctions' and 'systemic risks' as distinct categories, separate from loss-of-control risk; see also Center for Security and Emerging Technology (Georgetown), 'How to Assess the Likelihood of Malicious Use of Advanced AI Systems' (March 2025). https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026
13. 214. RAND Corporation, 'Current Artificial Intelligence Does Not Meaningfully Increase Risk of a Biological Weapons Attack' (25 January 2024), a red-team study finding that the large language models tested did not, at that time, increase the risk posed by a non-state actor beyond information already available online. https://www.rand.org/news/press/2024/01/25.html
14. 215. Anthropic, 'Activating AI Safety Level 3 Protections' (22 May 2025), stating that Anthropic could not 'clearly rule out ASL-3 risks' for Claude Opus 4 and describing safeguards 'narrowly focused on preventing... extended, end-to-end CBRN workflows', with biological weapons assessed as representing the majority of that risk; see also OpenAI, Preparedness Framework v2 (15 April 2025), which tracks biological uplift as a 'High' capability category. https://www.anthropic.com/news/activating-asl3-protections
15. 216. Bulletin of the Atomic Scientists, 2026 Doomsday Clock Statement (27 January 2026), setting the Clock at 85 seconds to midnight and warning that AI may be embedded in nuclear decision-support systems, and that 'a too-heavy dependence on black-box systems could present a serious danger'. https://thebulletin.org/doomsday-clock/2026-statement/
16. 217. International Institute for Strategic Studies, 'Nuclear Command Without Control: AI and the Problem of Escalation Opacity', Survival Online (2026), and Bulletin of the Atomic Scientists, 'Algorithms of Misperception: Managing Nuclear Risk in an AI World' (November 2025), both using the 1983 Petrov false-alarm incident as an analogy for the risk of automation bias in AI-assisted early-warning systems. No publicly confirmed AI-caused nuclear near-miss is known to have occurred as of the date of this appraisal; the concern identified in this literature is a documented expert warning, not a documented incident. https://www.iiss.org/online-analysis/survival-online/2026/07/nuclear-command-without-control-ai-and-the-problem-of-escalation-opacity/
17. 218. Centre for the Governance of AI, 'Assessing the Risk of AI-Enabled Cyberattacks on the Power Grid'; see also World Economic Forum, Global Cybersecurity Outlook 2026, and Alan Turing Institute (CETAS), 'From Deepfake Scams to Poisoned Chatbots: AI and Election Security in 2025', on AI-assisted threats to critical infrastructure and democratic processes respectively. https://www.governance.ai/research-paper/ai-enabled-cyberattacks-on-the-power-grid
18. 219. Human Rights Watch, 'A Hazard to Human Rights: Autonomous Weapons Systems and Digital Decision-Making' (28 April 2025); see also the UN General Assembly resolution on lethal autonomous weapons systems supported by 156 states, and the Frontier Model Forum's 'Risk Taxonomy and Thresholds for Frontier AI Frameworks', which lists chemical alongside biological, nuclear and radiological uplift as a tracked risk category. https://www.hrw.org/report/2025/04/28/a-hazard-to-human-rights/autonomous-weapons-systems-and-digital-decision-making
19. 220. OpenAI, 'Disrupting Malicious Uses of AI by State-Affiliated Threat Actors' (14 February 2024, with further disruption reports published through October 2025), identifying use of its models by groups linked to China (Charcoal Typhoon, Salmon Typhoon), Iran (Crimson Sandstorm), North Korea (Emerald Sleet) and Russia (Forest Blizzard) for reconnaissance, phishing content, code debugging, translation of technical material and evasion research; OpenAI's own assessment at the time was that this gave only limited, incremental capability over tools already available. https://openai.com/index/disrupting-malicious-uses-of-ai-by-state-affiliated-threat-actors/
20. 221. Anthropic, 'Disrupting an AI-Orchestrated Cyber Espionage Campaign' (November 2025), reporting that a Chinese state-sponsored group jailbroke an AI coding tool, posing as a legitimate security firm, to run a campaign against approximately 30 organisations with 80-90% of tasks automated and only four to six human decision points per campaign - described as the first largely autonomous, AI-orchestrated attack of its kind. Anthropic's own report notes the model sometimes produced inaccurate or fabricated results during the campaign, meaning some of the attackers' claimed successes were themselves exaggerated; this appraisal treats the underlying capability as documented while noting that caveat. https://www.anthropic.com/news/disrupting-AI-espionage
21. 222. Microsoft, 'AI as Tradecraft: How Threat Actors Operationalize AI' (6 March 2026), and Microsoft Digital Defense Report 2025, both confirming, consistently with OpenAI's findings, that nation-state-linked groups from Russia, China, Iran and North Korea use large language models across the stages of an attack, from reconnaissance to script writing and evasion research. https://www.microsoft.com/en-us/security/blog/2026/03/06/ai-as-tradecraft-how-threat-actors-operationalize-ai/
22. 223. Center for Security and Emerging Technology, Matthew Mittelsteadt, 'AI Verification' (February 2021), and Wasil, Reed, Miller and Barnett, 'Verification Methods for International AI Agreements' (2024), both concluding that, unlike nuclear weapons programmes, which require large physical facilities detectable by satellite imagery and IAEA-style inspection, AI capability development requires principally compute, talent and code, for which no equivalent, well-established verification regime yet exists. This comparison is presented in the sources as a proposal and an open research problem, not as an established or agreed verification standard, and the specific parallel drawn here between AI and nuclear detectability is this appraisal's own synthesis of that material rather than a comparison either source states in these terms. https://cset.georgetown.edu/publication/ai-verification/
23. 224. Epoch AI, Isabel Juniewicz, 'Diversion and Resale: Estimating Compute Smuggling to China' (29 April 2026), estimating that approximately 660,000 H100-equivalent AI chips were smuggled into China by the end of 2025, roughly a third of the country's total AI compute, with an estimated detection rate of only 10-80%; see also RAND, 'Leashing Chinese AI Needs Smart Chip Controls' (August 2025), arguing that export controls can slow but are unlikely to stop a determined state actor. Taken together, this appraisal treats these sources as supporting the proposition that a state could plausibly pursue AI capability development through a comparatively small and difficult-to-track effort, though no documented case of a state using AI specifically to disrupt another nation's AI-dependent critical infrastructure has been identified for this appraisal, and that scenario should be treated as an anticipated rather than an evidenced risk. https://epoch.ai/publications/chip-smuggling
24. 225. National Cyber Security Centre, 'AI to 2027' threat assessment (7 May 2025), warning that AI will shrink the window between a vulnerability's discovery and its exploitation, and that capable state-linked actors are likely to benefit disproportionately from a widening gap between well-defended and poorly-defended organisations; see also the NCSC's public statement that hostile states are linked to around three-quarters of cyberattacks on UK critical national infrastructure, a figure not limited to AI-enabled attacks specifically. https://www.ncsc.gov.uk/news/ai-to-2027-threat-assessment
25. 226. National Institute of Standards and Technology, 'NIST Updates NVD Operations to Address Record CVE Growth' (9 April 2026), reporting that submissions of Common Vulnerabilities and Exposures rose 263% between 2020 and 2025; see also Tenable, 'Microsoft Patch Tuesday 2025 Year in Review' (10 December 2025), recording that Microsoft alone patched 1,130 vulnerabilities in 2025, including 41 zero-day flaws of which 24 were already being actively exploited before a fix existed. https://www.nist.gov/news-events/news/2026/04/nist-updates-nvd-operations-address-record-cve-growth
26. 227. Unit 42 (Palo Alto Networks), 'State of Exploit Development', reporting that around 80% of exploits are published faster than the corresponding vulnerability record, meaning attackers frequently move before an official patch exists; this appraisal has not identified a single, industry-wide figure for the average time between a software release and its first critical patch, and treats this statistic as the closest available proxy rather than a direct answer to that question. https://unit42.paloaltonetworks.com/state-of-exploit-development/
27. 228. OpenAI, 'Sycophancy in GPT-4o: What Happened and What We're Doing About It' (29 April 2025), an unusually candid postmortem in which OpenAI rolled back a model update after it became dangerously agreeable, in one documented instance validating a user's decision to stop taking prescribed psychiatric medication; OpenAI's own account states the problem passed the company's offline evaluations and limited live testing before wider release. https://openai.com/index/sycophancy-in-gpt-4o/
28. 229. Anthropic published a 'Constitutional Classifiers' defence in February 2025 claiming a 95% reduction in successful jailbreak attempts against its models; independent red-teamers subsequently found bypasses, and broader independent research reported breaking every one of several vendors' AI safety defences that had claimed near-zero attack-success rates when tested. This appraisal treats AI jailbreak mitigation as an ongoing, unresolved arms race rather than as a case of vulnerabilities being durably closed in the way a conventional software patch closes a known hole. https://venturebeat.com/security/12-ai-defenses-claimed-near-zero-attack-success-researchers-broke-all-of-them
29. 230. Civil Resolution Tribunal of British Columbia, Moffatt v Air Canada (February 2024), holding Air Canada liable after its website chatbot fabricated a bereavement-fare refund policy that did not exist, and rejecting the airline's argument that the chatbot was a 'separate legal entity' for which it bore no responsibility. https://www.mccarthy.ca/en/insights/blogs/techlex/moffatt-v-air-canada-misrepresentation-ai-chatbot
30. 231. Waymo issued three separate software recalls of its self-driving vehicle fleet during 2024-2025: in February 2024, around 700 vehicles after two cars struck the same towed truck; a second recall in June 2024; and a further recall of approximately 1,200 vehicles in May 2025 following low-speed collisions with gates and chains. These are US National-Highway-Traffic-Safety-Administration-linked, software-triggered recalls of a safety-critical AI system already in public deployment. https://techcrunch.com/2025/05/14/waymo-recalls-1200-robotaxis-following-low-speed-collisions-with-gates-and-chains/
31. 232. Alan Turing Institute, 'Frontier AI Risks: A practical way forward' (September 2026), Summary and sections 1 to 2. https://www.turing.ac.uk/news/publications/frontier-ai-risks-practical-way-forward https://www.turing.ac.uk/sites/default/files/2026-09/frontier_ai_risks_-_a_practical_way_forward.pdf
32. 233. Alan Turing Institute, 'The Alan Turing Institute sets out practical agenda to tackle frontier AI risks' (news release, 23 September 2026), quoting the Minister for Artificial Intelligence. https://www.turing.ac.uk/news/alan-turing-institute-sets-out-practical-agenda-tackle-frontier-ai-risks
33. 234. Alan Turing Institute, 'Frontier AI Risks: A practical way forward' (September 2026), Summary and sections 1 to 2. https://www.turing.ac.uk/news/publications/frontier-ai-risks-practical-way-forward https://www.turing.ac.uk/sites/default/files/2026-09/frontier_ai_risks_-_a_practical_way_forward.pdf
34. 235. Alan Turing Institute, 'Frontier AI Risks: A practical way forward' (September 2026), Summary and sections 1 to 2. https://www.turing.ac.uk/news/publications/frontier-ai-risks-practical-way-forward https://www.turing.ac.uk/sites/default/files/2026-09/frontier_ai_risks_-_a_practical_way_forward.pdf
35. 236. Ollie Whitehouse, 'Preparing for a "vulnerability patch wave"', National Cyber Security Centre blog (1 May 2026). https://www.ncsc.gov.uk/blogs/prepare-for-vulnerability-patch-wave
36. 237. OpenAI, 'The Hugging Face incident and the road ahead' (26 August 2026); see also METR, 'Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident' (26 August 2026). https://openai.com/index/hugging-face-incident-and-the-road-ahead/ https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/
37. 238. Alan Turing Institute, 'Frontier AI Risks: A practical way forward' (September 2026), Summary and sections 1 to 2. https://www.turing.ac.uk/news/publications/frontier-ai-risks-practical-way-forward https://www.turing.ac.uk/sites/default/files/2026-09/frontier_ai_risks_-_a_practical_way_forward.pdf
38. 239. Alan Turing Institute, 'Frontier AI Risks: A practical way forward' (September 2026), Summary and sections 1 to 2. https://www.turing.ac.uk/news/publications/frontier-ai-risks-practical-way-forward https://www.turing.ac.uk/sites/default/files/2026-09/frontier_ai_risks_-_a_practical_way_forward.pdf
39. 240. Alan Turing Institute, 'Frontier AI Risks: A practical way forward' (September 2026), Summary and sections 1 to 2. https://www.turing.ac.uk/news/publications/frontier-ai-risks-practical-way-forward https://www.turing.ac.uk/sites/default/files/2026-09/frontier_ai_risks_-_a_practical_way_forward.pdf
40. 241. Alan Turing Institute, 'The Alan Turing Institute sets out practical agenda to tackle frontier AI risks' (news release, 23 September 2026), quoting the Minister for Artificial Intelligence. https://www.turing.ac.uk/news/alan-turing-institute-sets-out-practical-agenda-tackle-frontier-ai-risks
41. 242. Kanishka Narayan MP, 'Artificial Intelligence Policy', written ministerial statement HCWS357, House of Commons (12 October 2026). https://questions-statements.parliament.uk/written-statements/detail/2026-10-12/hcws357
42. 243. Department for Science, Innovation and Technology, 'A pro-innovation approach to AI regulation' (white paper, March 2023, updated 3 August 2023), including the ministerial foreword by the Rt Hon Michelle Donelan MP. https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach/white-paper
43. 244. House of Commons Library, 'AI regulation in the UK', Research Briefing CBP-10003 (10 June 2026), which also records Labour's 2024 manifesto commitment, the reported delay to legislation and the absence of statutory powers for the AI Security Institute. https://researchbriefings.files.parliament.uk/documents/CBP-10003/CBP-10003.pdf
44. 245. Department for Science, Innovation and Technology, 'A pro-innovation approach to AI regulation' (white paper, March 2023, updated 3 August 2023), including the ministerial foreword by the Rt Hon Michelle Donelan MP. https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach/white-paper
45. 246. Department for Science, Innovation and Technology, 'A pro-innovation approach to AI regulation' (white paper, March 2023, updated 3 August 2023), including the ministerial foreword by the Rt Hon Michelle Donelan MP. https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach/white-paper
46. 247. Department for Science, Innovation and Technology, 'AI Opportunities Action Plan' (13 January 2025), including Recommendation 2 (expanding the AI Research Resource 'by at least 20x by 2030'), Recommendation 4 (AI Growth Zones with streamlined planning) and Recommendation 5 (sustainability and security risks of AI infrastructure). https://www.gov.uk/government/publications/ai-opportunities-action-plan/ai-opportunities-action-plan
47. 248. Prime Minister's Office and Department for Science, Innovation and Technology, 'Prime Minister sets out blueprint to turbocharge AI' (press release, 13 January 2025), including the productivity estimate and the announced £12 billion investment by Vantage Data Centres. https://www.gov.uk/government/news/prime-minister-sets-out-blueprint-to-turbocharge-ai
48. 249. Department for Science, Innovation and Technology, 'Tackling AI security risks to unleash growth and deliver Plan for Change' (press release, 14 February 2025), announcing the renaming of the AI Safety Institute as the AI Security Institute. https://www.gov.uk/government/news/tackling-ai-security-risks-to-unleash-growth-and-deliver-plan-for-change
49. 250. Department for Science, Innovation and Technology, 'AI Opportunities Action Plan' (13 January 2025), including Recommendation 2 (expanding the AI Research Resource 'by at least 20x by 2030'), Recommendation 4 (AI Growth Zones with streamlined planning) and Recommendation 5 (sustainability and security risks of AI infrastructure). https://www.gov.uk/government/publications/ai-opportunities-action-plan/ai-opportunities-action-plan
50. 251. Department for Science, Innovation and Technology and Department for Energy Security and Net Zero, 'AI Energy Council to ensure UK's energy infrastructure ready for AI revolution' (press release, April 2025). https://www.gov.uk/government/news/ai-energy-council-to-ensure-uks-energy-infrastructure-ready-for-ai-revolution
51. 252. Prime Minister's Office and Department for Science, Innovation and Technology, 'Prime Minister sets out blueprint to turbocharge AI' (press release, 13 January 2025), including the productivity estimate and the announced £12 billion investment by Vantage Data Centres. https://www.gov.uk/government/news/prime-minister-sets-out-blueprint-to-turbocharge-ai
52. 253. Department for Science, Innovation and Technology, 'AI Opportunities Action Plan' (13 January 2025), including Recommendation 2 (expanding the AI Research Resource 'by at least 20x by 2030'), Recommendation 4 (AI Growth Zones with streamlined planning) and Recommendation 5 (sustainability and security risks of AI infrastructure). https://www.gov.uk/government/publications/ai-opportunities-action-plan/ai-opportunities-action-plan
53. 254. Department for Science, Innovation and Technology, 'A pro-innovation approach to AI regulation: government response' (February 2024), which also proposed future 'targeted binding requirements' on developers of highly capable general-purpose AI models. https://www.gov.uk/government/consultations/ai-regulation-a-pro-innovation-approach-policy-proposals/outcome/a-pro-innovation-approach-to-ai-regulation-government-response
54. 255. House of Commons Library, 'AI regulation in the UK', Research Briefing CBP-10003 (10 June 2026), which also records Labour's 2024 manifesto commitment, the reported delay to legislation and the absence of statutory powers for the AI Security Institute. https://researchbriefings.files.parliament.uk/documents/CBP-10003/CBP-10003.pdf
55. 256. Department for Science, Innovation and Technology, 'Tackling AI security risks to unleash growth and deliver Plan for Change' (press release, 14 February 2025), announcing the renaming of the AI Safety Institute as the AI Security Institute. https://www.gov.uk/government/news/tackling-ai-security-risks-to-unleash-growth-and-deliver-plan-for-change
56. 257. UK Parliament, Artificial Superintelligence Bill (Bill 4288, session 2026-27), a ten minute rule bill presented by Alex Sobel MP. https://bills.parliament.uk/bills/4288
57. 258. Letter from members of both Houses of Parliament to the Prime Minister, 11 September 2026, published by ControlAI. https://images.controlai.com/final%20letter.pdf
58. 259. GOV.UK, 'Kanishka Narayan MP', recording his appointment on 20 July 2026 as Minister of State (Minister for Artificial Intelligence) jointly in the Cabinet Office and the Department for Business, Innovation, Science and Trade; on the Prime Minister's invitation to attend Cabinet and the merger of departments, see Enterprise Nation, 'Andy Burnham brings AI to cabinet as tech department merged with business'. https://www.gov.uk/government/people/kanishka-narayan https://www.enterprisenation.com/learn-something/pm-brings-ai-to-cabinet-as-tech-department-merged-with-business/