NextFin News - British Prime Minister Andy Burnham is set to use his debut address to the United Nations General Assembly on Tuesday to call for a coordinated global effort to control the threats posed by artificial intelligence, positioning the UK as a convening power on AI safety at a moment when trust between Washington and its allies has rarely been thinner.
The speech, delivered in New York on September 22 as the 81st General Assembly opens, lands between two pressures that define the dilemma for any government trying to govern frontier AI: the technology is moving faster than national regulators can follow, and the countries that control the most advanced models are increasingly treating them as instruments of geopolitical competition rather than shared infrastructure.
The Situation: A Prime Minister, a Fractured World, and a Technology That Outruns Borders
Burnham arrives in New York for the annual gathering of world leaders with AI already established as one of the central files on his desk. Since taking office on July 20, he has elevated artificial intelligence into the cabinet, appointing Kanishka Narayan as minister for AI, and has centralized AI policy within Downing Street — a recognition that the issue now cuts across national security, economic strategy, and diplomacy.
The immediate backdrop is unusually stark. On September 9, Evan Hubinger, the alignment-science lead at Anthropic, wrote on X that "we really do earnestly believe AI could kill all humans," adding: "I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to." The post was a response to Jacob Coxon, a 27-year-old researcher who resigned from Anthropic a day earlier in a seven-part thread that drew tens of millions of views. Coxon wrote that he had spent three years doing pretraining research at both OpenAI and Anthropic, that "neither company is acting responsibly," and that they were "racing straight to self-improving superintelligence and gambling with our lives."
"They are racing straight to self-improving superintelligence and gambling with our lives," Coxon wrote, announcing his departure from the company where he had worked on pretraining research.
Then came the diplomatic flashpoint. Reports emerged that Anthropic had withheld its latest model from the UK's AI Safety Institute — one of the world's leading bodies for testing frontier systems — apparently under pressure from the Trump administration. In the House of Commons, Liberal Democrat leader Sir Ed Davey asked Burnham directly whether it was acceptable for the US president to undermine Britain's efforts to keep the world safe from dangerous AI.
Burnham's answer captured the balancing act ahead: "He is absolutely right that AI poses risks to our national security, but it also could be the source of solutions to keeping us safer." He added that "the offer is open to colleagues on all sides of the House to be part of that crucial discussion."
"He is absolutely right that AI poses risks to our national security, but it also could be the source of solutions to keeping us safer," Burnham told the House of Commons.
A Cabinet Office spokesperson struck a more institutional note: "These risks do not stop at national borders and no country can tackle them alone. The UK will continue to test the most advanced models, build a rigorous scientific understanding of their capabilities and risks, and ensure policy decisions are grounded in the evidence." The office also pointed out that the AI Safety Institute had tested OpenAI's most powerful model, GPT-6 Astra, only the previous week.
That exchange — domestic acknowledgment of risk, a fracture in transatlantic safety cooperation, and an insistence that the UK remains a credible tester — is the prelude to Tuesday's speech. The question it poses is whether a middle power can convene a global governance effort when the two superpowers are pulling in opposite directions.
The Analysis: Why National Regulation Cannot Contain a Borderless Technology
The mechanism: capability without jurisdiction
The core problem Burnham is trying to solve is structural, not cyclical. Frontier AI is a dual-use technology whose risks — from biosecurity and cyber-offense to autonomous weapons and loss of control over increasingly capable systems — do not respect borders. But the levers of control are almost entirely national: export controls on chips, domestic safety institutes, national liability regimes, and bilateral agreements.
That mismatch has a name in policy circles: a collective-action problem. If one country slows deployment for safety testing while a rival races ahead, the cautious country fears it will lose economic and military advantage. The result is a race dynamic even among actors who would all prefer a slower, safer pace. Hubinger's own framing — "Anthropic is trying its best, but we do not yet have a plan to solve alignment" — is the private-sector version of the same trap: individual firms can be responsible and still lack a path to safety at the frontier.
This is why the call for a global effort is not rhetorical padding. Without some form of cross-border coordination — shared testing protocols, model-registration requirements, compute monitoring, and incident-reporting obligations — national regulation is porous by design. A model trained in one jurisdiction can be deployed globally in minutes; a safety standard that applies only in London or Brussels can be routed around.
The geopolitical reality: two poles are already forming
The world Burnham is walking into is not a blank slate. In July 2026, the UN held its first Global Dialogue on AI Governance in Geneva, convened under a General Assembly mandate and informed by an Independent International Scientific Panel on AI co-chaired by Yoshua Bengio and Maria Ressa. That panel warned in June that AI could "cause catastrophic harm, either on its own or due to malicious users," and that the technology is "outpacing both scientific understanding and governments' ability to adapt."
But even as the UN process gathered, a competing pole emerged. On July 16, 2026, representatives of 29 countries signed the agreement establishing the World AI Cooperation Organization in Shanghai. The WAICO framework emphasizes state sovereignty, capacity-building, and security over the universal human-rights language that Western democracies have pushed. The result is two multilateral tracks, each courting the Global South — and each offering a different answer to the question of who gets to set the rules.
There is evidence that this fragmentation is not merely theoretical. A study released by the Meta Oversight Board on July 16 tested ten commercial large-language models from Anthropic, DeepSeek, Google, Meta, OpenAI, and xAI on politically critical prompts. The models refused 14% of requests involving permissive jurisdictions such as the United Kingdom and the United States, but 34% of requests involving restrictive jurisdictions such as China and Saudi Arabia. In other words, the same model behaves differently depending on who is asking, which means governance is already being outsourced to the most permissive or the most restrictive regulator in the chain. Board co-chair Paolo Carozza called the pattern "extended censorship by proxy that goes across borders."
The UK's leverage: credible testing, limited access
Burnham's strongest card is the AI Safety Institute. It is genuinely world-leading in frontier model evaluation, and its recent testing of GPT-6 Astra before public release demonstrates that leading US labs still see value in independent assessment. The UK's pitch is that it can offer something the superpowers cannot easily provide: a trusted, technically competent neutral ground for testing and verification.
But that leverage has a ceiling, and it is defined by access. Advanced AI development is concentrated in a handful of companies in the United States and China, and it is not clear on what terms the UK — or Europe — will have access to the capabilities that will define economic potential and national power. If the US administration treats frontier models as strategic assets to be withheld, the UK's testing regime can only evaluate what it is allowed to touch. The Anthropic episode is a preview of that constraint.
This is the tightrope that Chatham House's Olivia O'Sullivan described ahead of the New York trip: "Burnham has a bit of a tightrope to walk." He must build rapport with a US president who has dismissed AI safety concerns as a media hoax, while simultaneously arguing that the technology requires binding international guardrails.
Second-order effects: when governance fragments, risk finds the weakest link
The first-order story is diplomatic: can countries agree on rules? The second-order story is about where the cost of AI risk ultimately lands. When governance is fragmented and voluntary, the cost of a catastrophic failure does not sit on any single balance sheet — it is socialized across borders. A model developed in one jurisdiction, tested under the loosest available regime, and deployed globally means that the strictest regulator in the chain does not set the standard; the weakest one does.
That dynamic distorts incentives in two directions. Developers racing for capability have an argument for moving fast: any slowdown is a gift to a rival. Regulators have an argument for moving slowly: the downside is unbounded. The gap between those two incentives is exactly where a global framework is meant to sit — converting a race into a set of shared constraints so that safety becomes a condition of market access rather than a voluntary sacrifice.
The most concrete articulation of what that framework could look like has come from Darren Jones, the former Cabinet Office minister, who has called for a multinational treaty. In a letter to Burnham, UN secretary general Antonio Guterres, and OECD secretary general Mathias Cormann, he wrote: "The debate ranges from the end of humanity to claims of 'marketing hype' pre-IPO. Either way, governments must now step in." His proposed instrument is "a new multi-national treaty for the regulated and safe development of superintelligence," explicitly "not a ban on innovation or scientific endeavour, but a safety first approach to the rapid development of this technology."
"We need a new multi-national treaty for the regulated and safe development of superintelligence," wrote Darren Jones, the former Cabinet Office minister, in a letter to the prime minister, the UN secretary general and the OECD secretary general.
That is the hardest path to negotiate, because a treaty requires enforcement, and enforcement requires the superpowers to submit to something above themselves. But the alternative — a patchwork of national regimes with a technology that ignores borders — leaves the same collective-action problem unsolved.
The counter-thesis: coordination is a mirage, and national action is what matters
The strongest argument against Burnham's approach is not that AI risks are exaggerated, but that global coordination is a mirage — and that chasing it wastes the narrow window in which national action is still possible. Critics point out that every major AI governance initiative since 2023 has produced declarations rather than obligations. The UN dialogue produced a draft framework; the WAICO charter produced a sovereignty-first club; the Bletchley and Seoul summits produced photo opportunities. Meanwhile, capability continues to advance on a predictable cadence.
There is a harder version of this critique, and it comes from the liability camp rather than the treaty camp: enforcement works better through domestic courts than through international consensus. If AI developers face meaningful liability for harms, the discipline comes from the prospect of damages, not from an unenforceable treaty with a rival superpower. A country that cannot secure its own liability regime has little business negotiating one for the world.
This counter-thesis has force, but it misses the specific nature of the risk. Tort liability works when harms are attributable, localized, and compensable. It does not work when the harm is a borderless, irreversible event — a released pathogen designed by an AI, a cascading cyber failure, or a loss of control over an autonomous system. For those, the only mitigation is prevention, and prevention requires that the actor with the capability be bound before deployment, not sued after catastrophe. That is the case for a global effort, and it is the case Burnham will make.
The falsifying signal is specific: if, within 12 months, the US and China both establish verifiable pre-deployment testing and incident-reporting obligations for frontier models — and submit to some form of joint or third-party audit — then the structural-fragmentation thesis is wrong, and national-plus-bilateral coordination is sufficient. If instead each side entrenches its own bloc, with compute and model access weaponized, then the race dynamic dominates and Burnham's convening role shrinks to that of a respected but increasingly isolated tester.
Conclusion: What to Watch, and Who Is Exposed
The short-term read is diplomatic theater with real stakes. Tuesday's speech will be judged on whether Burnham names a concrete mechanism — a treaty process, an expanded testing consortium, a compute-tracking regime — or settles for a call to cooperate. The follow-through on the Geneva framework and the response from Washington and Beijing are the immediate catalysts.
The medium-term read is about leverage. The UK's influence depends on two things: whether US labs continue to submit frontier models to independent testing, and whether the Anthropic episode was an anomaly or a pattern. If access to models tightens, the UK's safety infrastructure becomes a lifeboat with no ships to inspect.
The long-term read is structural. The governance gap for frontier AI is not a cyclical lag that will close on its own; it is a regime problem created by a technology that is globally deployable and nationally regulated. Closing it requires either a treaty with teeth or a market mechanism that makes unsafe deployment financially impossible. Neither exists yet.
Base case: the UN process advances incrementally, with more countries signing onto shared testing standards but no binding treaty, and the US-China split hardens into parallel blocs. Upside case: the warnings from inside AI labs shock enough governments into a binding pre-deployment testing pact with verification. Downside case: capability races accelerate, safety cooperation fractures along geopolitical lines, and the UK's convening role is reduced to symbolism.
The beneficiaries of a genuine global effort are the companies that can afford verifiable safety and the countries that host credible testing infrastructure; the exposed are the developers betting that speed alone will win, and the governments that assume national regulation is enough. Burnham's central judgment is that the world is about to find out which of those bets was right.
The tightrope is real, but the alternative is worse: a world where the most dangerous technology in history is governed by whoever is willing to move fastest, and where safety is the first casualty of a race no one can afford to win.
Explore more exclusive insights at nextfin.ai.
