Left unchecked, a handful of billionaires in California and Beijing will decide whether humanity survives its next technological leap. That's the blunt warning Liberal Democrat leader Sir Ed Davey delivered at his party's annual conference in Brighton in September 2026. While most politicians in Westminster treat artificial intelligence like just another economic growth metric, Davey broke ranks to propose something radical. He wants a global non-proliferation treaty to halt the development of superintelligent AI altogether, backed by an international watchdog modelled directly on the International Atomic Energy Agency (IAEA).
It's a bold swing. Frankly, it's also a rare moment of political candor in a debate usually dominated by corporate lobbying.
If you've been following UK tech policy lately, you know the prevailing orthodoxy has been to cozy up to Silicon Valley, deregulate fast, and pray the tax receipts roll in. Davey's speech flipped that script on its head. By framing artificial superintelligence not as a consumer gadget or a productivity tool, but as an existential threat on par with nuclear weapons, he forced a question onto the national stage that most leaders would rather dodge. Can you actually put the genie back in the bottle once an arms race has already started?
What Ed Davey Actually Proposed in Brighton
Strip away the conference hall applause, and Davey's pitch rests on three concrete pillars.
First, he called for an immediate international pause on the development of artificial superintelligence—systems designed to vastly outperform human cognitive abilities across every domain. He pointed out during his September 22, 2026 address that some of the very engineers and researchers building frontier models have openly warned that superintelligent systems could "kill us all". "That is not science fiction," Davey told delegates in Brighton. "That is a live warning, from the people who know."
Second, he proposed establishing a global oversight body modeled on the IAEA. When nuclear fission became a reality in the mid-twentieth century, world powers didn't trust private defense contractors to police themselves. Nations built an inspection regime with teeth, monitoring enriched uranium and centrifuge facilities across borders. Davey argues frontier AI compute clusters require the exact same level of intrusive international verification.
Third, he attacked the concentration of raw political and economic power behind current AI labs. In Davey's view, artificial intelligence isn't merely a technical safety puzzle. "The problem is fundamentally one about power," he argued. "It's too concentrated, and we must break it up."
Why He Targeted Andy Burnham and US Tech Deals
Davey didn't just aim his fire at Silicon Valley executives like Elon Musk. He used the Brighton stage to attack Andy Burnham over a bilateral US-UK technology partnership announced during Burnham's visit to New York for the United Nations General Assembly.
According to Davey, signing bilateral tech agreements with the Trump administration—specifically arrangements tied to defense and war departments—completely misses the scale of the danger. He accused Burnham of making Britain "more dependent on a Trump administration in the pocket of the tech bros".
Look at the political calculation here, and it makes total sense. While Labour figures try to prove they are pro-business by welcoming American data centers and defense-tech partnerships, the Liberal Democrats are carving out space as the party willing to stand up to big tech oligarchs. Davey even drew parallels to climate diplomacy, reminding voters how European leaders helped forge the groundwork for the Paris Agreement rather than waiting for Washington to lead.
Why the Nuclear Non-Proliferation Analogy Matters
Whenever someone compares AI regulation to nuclear treaties, tech optimists roll their eyes. They'll tell you software isn't plutonium. You can't detect lines of code with a Geiger counter, and you can't easily track weights moving across fiber-optic cables.
They're half right. And completely missing the point.
Having spent years tracking international technology governance and compute supply chains, I can tell you that modern frontier AI isn't built in a teenager's bedroom. Training a model that approaches or exceeds human-level general intelligence requires tens of thousands of specialized accelerators, gigawatts of dedicated electrical power, and massive physical cooling infrastructure. You cannot hide a hyperscale data center under a tarp.
How an IAEA for AI Would Actually Work
If governments ever took Davey's proposal seriously, enforcement wouldn't rely on inspecting every laptop on earth. It would target physical chokepoints:
- Semiconductor fabrication and chip tracking. Advanced AI chips come from a tiny handful of foundries and rely on extreme ultraviolet lithography machines built by a single Dutch company, ASML. Hardware-level telemetry and export controls already exist; a treaty would formalize global audits of where high-end GPUs go.
- Compute threshold caps. Just as the Nuclear Non-Proliferation Treaty distinguishes between civilian nuclear power and weapons-grade enrichment, an AI treaty would allow narrow, beneficial AI—like medical diagnostics or climate modeling—while banning training runs above a specific floating-point operations (FLOP) threshold.
- Whistleblower protections and facility inspections. International inspectors would verify that major compute clusters in the US, UK, Europe, and China aren't secretly training unaligned superintelligent systems.
The IAEA itself was founded in 1957 during the height of the Cold War, proving that bitter geopolitical rivals can still cooperate when mutual survival is on the line. The United States and the Soviet Union didn't trust each other an inch, yet they signed the Nuclear Non-Proliferation Treaty in 1968 because neither wanted an uncontrolled cascade of apocalyptic weapons.
The Hard Truths Critics Point Out
Now let's be honest about where Davey's plan hits a brick wall. Calling for a global treaty from a conference stage in Brighton is the easy part. Getting Washington and Beijing to sign on is another beast entirely.
Critics of the Lib Dem proposal—and there are plenty across Westminster and the tech sector—raise three serious objections that deserve a fair hearing:
- The geopolitical prisoner's dilemma. Right now, both the US government and the Chinese Communist Party view AI supremacy as the defining military and economic advantage of the twenty-first century. Neither side wants to hit pause if they suspect the other will cheat in secret. Without mutual trust, a pause looks like unilateral disarmament.
- Definitional chaos. Nuclear weapons are easy to define. What exactly counts as "superintelligent AI"? Researchers themselves disagree on where narrow AI ends, where artificial general intelligence (AGI) begins, and when a system crosses into superintelligence. Writing a legally binding treaty around a moving scientific target is a nightmare.
- Economic self-harm for smaller nations. Britain doesn't currently host the world's largest frontier model labs; those sit primarily in the US and China. Opponents argue that if the UK pushes for strict domestic bans on superintelligence research—something Davey also floated in interviews with Sky News—Britain simply locks itself out of the economic upside while American and Chinese firms keep racing ahead anyway.
Even Nvidia CEO Jensen Huang dismissed catastrophic AI scenarios around the same time, claiming there was zero chance AI would destroy the world by 2030 and brushing off safety warnings as "doomsday narratives". When the person selling the picks and shovels during a gold rush tells you there's no danger, you take it with a grain of salt—but his view reflects how entrenched industry resistance really is.
Why Ignoring the Warning Is Worse
Still, dismissing Davey's proposal as naive misses how fast the ground is shifting beneath our feet.
A few years ago, talking about existential risk from AI got you laughed out of serious policy rooms. Today, pioneers of deep learning—including Nobel laureate Geoffrey Hinton and Yoshua Bengio—have repeatedly warned that we are building systems we don't fully understand and may not be able to control. When the architects of a technology tell you the brakes don't work, listening to them isn't Luddism. It's basic common sense.
What Happens When Safety Is Left to Voluntary Pledges
We already tried the "trust the tech bros" approach. Remember the voluntary AI safety summits and non-binding corporate pledges from 2023 and 2024? Companies promised they would pause training if models hit dangerous capability thresholds.
Then competitive pressure kicked in. Safety teams were disbanded or sidelined, release timelines accelerated, and voluntary commitments quietly evaporated whenever a rival lab dropped a better benchmark score. Market incentives naturally reward speed over caution. Expecting trillion-dollar corporations to regulate themselves out of a lucrative market is like asking oil companies to cap their own emissions without legislation.
Davey's speech cuts right through that illusion. By insisting that democratic governments—not billionaire founders—must set the red lines, he's articulating what millions of ordinary voters already feel in their gut. People see AI disrupting jobs, flooding the internet with synthetic media, and concentrating unimaginable wealth in a few hands, all while politicians act like helpless bystanders.
How to Follow and Weigh In on UK AI Governance
Whether you think a global AI non-proliferation treaty is visionary statesmanship or wishful thinking, the policy battles happening right now in Britain and abroad will shape how this technology affects your work, privacy, and security.
If you want to track where this debate goes next—or push your representatives to take AI oversight seriously—start with a few concrete steps:
- Read the actual frontier safety frameworks. Don't rely on press releases from tech CEOs or soundbites from party conferences. Check independent assessments from the UK AI Security Institute (formerly the AI Safety Institute) to see what frontier models can actually do right now regarding biosecurity, cyber offense, and autonomous replication.
- Watch the UK Parliament's Science, Innovation and Technology Committee. That's where real legislative scrutiny happens, far away from conference-hall rhetoric. You can submit written evidence to open inquiries on AI governance and compute regulation as a member of the public.
- Challenge your local MP on bilateral tech deals. When UK politicians sign infrastructure or defense AI agreements with foreign governments or tech giants, ask what binding safety thresholds, environmental limits on data center power consumption, and democratic oversight mechanisms are attached.
- Separate narrow AI utility from unchecked frontier scaling. Use current AI tools to automate tedious work and boost your own productivity, while staying skeptical of claims that unrestricted scaling toward superintelligence is inevitable or automatically good for society.
Politicians rarely get ahead of a technological crisis before disaster forces their hand. Ed Davey's push for an international AI watchdog might face steep diplomatic odds in Washington and Beijing, yet putting a binding global treaty on the table forces every other UK party leader to answer a question they've spent years avoiding: if we won't draw a line at superintelligence, where exactly do we draw it?
For a closer look at how these proposals were framed on stage, you can watch coverage of Ed Davey's key note speech at Lib Dem conference, which captures his direct argument for pairing international AI guardrails with broader democratic reforms.
