In the first of two parts, Robin Brown and Gregory Andrews examine why AI, climate change, pandemics and online markets repeatedly outrun national regulation, and what existing arms-control and international regimes can teach us about governing risks that cross borders.
Anthropic chief Dario Amodei warned in September that, within a year, a swarm of AI agents could be capable of taking over large parts of the internet. Commentators asked what many have since Hiroshima: can the world regulate this the way it has nuclear weapons? Alan Kohler put the uncomfortable answer plainly: probably not, the reason being structural, not technical. It is a problem of sovereignty.
Whether the harm arrives as greenhouse gas emissions, an algorithm, a parcel or a pathogen, the underlying regulatory challenge is much the same: what happens at the point where a sovereign state’s formal rule-making runs out and the problem of transnational coordination takes over.
Nuclear, chemical and biological weapons are unusual cases where international regulation has achieved substantial, though uneven, success. The Nuclear Non-Proliferation Treaty (NPT), the Chemical Weapons Convention (CWC) and the Biological Weapons Convention (BWC) did not eliminate the underlying technologies, all of which are profoundly dual use. Instead, they drew legal lines around particular purposes, weapons and activities, while preserving peaceful uses. The NPT couples non-proliferation with safeguards against diversion of nuclear material; the CWC distinguishes prohibited weapons purposes from legitimate industrial, medical and other uses; and the BWC prohibits agents and toxins where their type or quantity cannot be justified for peaceful or protective purposes. International law did not ban the science. It defined what states could not do with it and, where possible, created ways of verifying compliance.
Climate change and AI are different, but also similar to each other. The combustion that warms the planet is the same combustion that has built the global economy; the capability that might one day produce a misaligned, uncontrollable AI is the same capability that helps cure diseases and gives whichever country or company gets there first a decisive edge.
The line between the dangerous and beneficial versions of both technologies is fuzzy. They are the same technology at different scale, speed and care, which is what can make continued escalation individually rational in the short run even where the cumulative risks are shared by everyone. The emissions of a particular country, the capability gains of a particular lab, look sensible from where that country or lab stands, even as everyone acting sensibly produces an outcome nobody wants.
Here sovereignty, ordinarily a source of order, can become an obstacle. As Kohler notes, Jonathan Schell (author of The Fate of the Earth, 1982) argued that national sovereignty would become “the sturdy child of terror and the twin brother of annihilation”: a shared, catastrophic threat can make the existing order stable rather than reckless. But that depends on every sovereign entity facing the same calculus at once. Climate and AI break that condition. If Australia decarbonises unilaterally, it absorbs a real cost, while the global emissions reduce only slightly, because coal and gas keep burning cheap elsewhere. If a frontier AI lab slows down, in Amodei’s phrase, to “pace the frontier”, it hands the lead to a rival that does not. Thirty annual UN climate conferences have produced pledges and, at best, a slower rate of emissions growth. AI governance involves attempting something even harder: restraining a technology whose harms remain contested, by using institutions that are failing to restrain fossil fuels.
A related but different failure appears wherever the internet itself is the route around a nation’s regulation. Take unregulated pharmaceuticals sold through offshore websites. The nation wants to enforce its drug-safety law but can’t, because the product is manufactured in one jurisdiction, shipped from a second and advertised from a third, arriving as an ordinary parcel rather than a container an inspector can open at a port. The illegal online pharmacy trade runs to billions of dollars a year, and a single coordinated sweep across 94 countries still barely dented supply. Cross-border telehealth raises the same problem for services. A platform based nowhere in particular can sell a prescription that would not meet a national regulator’s standard. The border, as somewhere a regulator can physically inspect, is less and less where the harm occurs.
Climate and AI are collective-action problems; every actor would prefer everyone restrain themselves, but nobody wants to restrain themselves alone. Pandemics involve this kind of challenge as well as the border problem. Pathogens recognise national borders even less than pharmaceutical shipments do, while hoarding vaccines, banning protective-equipment exports and delaying outbreak reporting create the same individually rational, collectively damaging incentives seen in climate and AI. The World Health Organization’s new Pandemic Agreement, adopted in May 2025, attempts to address this. But it is not yet complete. And even then, it expressly preserves national authority over measures such as vaccination, lockdowns and border controls. Sovereignty remains both the framework through which cooperation must occur and one of the constraints on how far it can go.
For AI, the arms-control model is nevertheless immediately attractive. Amodei has sketched a three-stage version: safety standards inside companies, then common standards among democracies, then coordination with authoritarian states. In July, more than 1,100 employees at OpenAI, Anthropic, Google and Meta petitioned Washington to back that pacing.
Two more detailed versions of similar thinking have appeared in P&I within weeks of each other. Ana Palacio has argued that AI needs its own Paris Agreement: voluntary commitments, shared definitions, mandatory incident reporting, deliberately open beyond the United States and China. But naming Paris as the proposed model concedes the limitation traced above. Paris is the clearest example of a voluntary framework bending a curve without turning it. A similar AI agreement would have the same limitations.
Christopher Tang supplies the missing piece of verification. Nuclear arms control worked because satellites and inspections let each side confirm compliance. An advanced AI model depends on physical infrastructure – chips, data centres, power – which is far harder to hide than a warhead and runs through a handful of choke-point supply chains. The idea also predates Tang. Bremmer and Suleyman proposed monitoring semiconductor supply chains in 2023 through the concept of ‘compute governance’. This is a concrete mechanism for a floor, something a verifier can check. The Trump administration’s August framework, which involves scrutinising closed models while exempting open-weight ones, is inadequate since released weights cannot be un-released.
These different failures point to different regulatory responses.
- A weapons regime defines prohibited purposes and activities and, where possible, makes compliance verifiable.
- A commons problem needs a shared, credible floor – a carbon price with a real border adjustment or a hardware-anchored verification regime – plus a liability rule that shifts downstream costs of harm onto whoever captures the upsides.
- A border-erosion problem needs less a floor than a fence, via customs screening, mutual recognition of approvals and enforcement cooperation.
None overcome sovereignty outright. Each instead reduces reliance on voluntary restraint or on a border that no longer performs the regulatory work once expected of it.
As sensible as the Palacio and Tang proposals are, both work at the level of treaty design. The harder question is whether floors and fences are enough, or whether sovereignty’s deeper architecture, popularly traced to the Peace of Westphalia in 1648, needs revising outright. That question, and why the answer lies in arrangements much more recent than 1648, is the subject of this article’s second part.
Robin Brown
Robin Brown is Deputy Chair of Fairer Future. He has advised Australian and overseas governments, businesses and NGOs on consumer protection, competition policy and regulatory accountability. Formerly head of the Consumers’ Federation of Australia, he helped secure the landmark court ruling that enabled bans on second‑hand tobacco smoke and spearheaded creation of both the Consumers’ Health Forum and the Australian Communications Consumer Action Network.
Gregory is a D’harawal man from NSW. He worked at SES levels in the Departments of Foreign Affairs and Trade, Environment, Climate Change, and Indigenous Affairs. He represented Australia as an Ambassador and High Commissioner in West Africa, and negotiated for Australia in the United Nations Climate Change negotiations. Gregory was Australia’s first-appointed Threatened Species Commissioner. He led development and implementation of Australia’s first Threatened Species Strategy and first Threatened Species Prospectus. In the NGO sector, he led, grew and reformed a national Indigenous charity.
Gregory speaks French and Mandarin. He has an Honours Degree in Economics and a Masters in Foreign Affairs and Trade.