How the world governs AI

Models cross borders. Training data crosses borders. Compute supply chains, synthetic content and the harms all cross borders. The laws do not. That mismatch is the central fact of AI governance in 2026, and it is worth understanding precisely why it persists.

Why no universal regime exists

1. Sovereignty. States genuinely disagree about privacy, speech, surveillance and the state's relationship with platforms. A single rulebook would have to pick winners among constitutions.
2. Strategic competition. Compute, chips and models are treated as strategic assets. States hesitate to accept verification or constraints that might advantage rivals.
3. Security carve-outs. Civilian rules can advance while military and intelligence uses stay contested or excluded.
4. Moving definitions. A treaty negotiated around one architecture may age before it is ratified.
5. Sector diversity. A clinical device, a hiring ranker and an industrial controller need different evidence and different regulators.
6. Enforcement across borders. Developers, deployers, compute providers and users span jurisdictions; causal chains are complex.
7. Capacity gaps. Many countries lack compute, testing labs or bargaining power, and reasonably resist rules written only by model-producing economies.
8. Slow ratification. Even agreed treaties take years of signatures, ratifications and implementing law.

Why a treaty is still the answer to inter-governmental conflict

Consider the alternatives actually on offer. Unilateral extraterritorial law reaches across borders but breeds friction: duplicated audits, forum shopping and rules written by whichever market is largest. Soft law converges vocabulary but cannot bind a rival, compensate a victim or compel an investigation. Export controls manage a chokepoint but govern access, not behaviour, and the chokepoints move.

Only a treaty creates reciprocal obligation between states that do not trust each other. That is what treaties are for. Verification, dispute settlement, mutual assistance and capacity funding cannot be improvised bilaterally at planetary scale; they need an instrument states have consented to. The research library's recommended architecture is deliberately modest: a thin binding floor for cross-border and severe risks (prohibited uses, incident reporting, contestability and remedy, whistle-blower protection, regulator access), a standards and mutual-recognition layer so evidence travels, national enforcement so constitutions keep their variation, shared science and incident learning built on the UN's new panel, and capacity and benefit-sharing so the regime is representative enough to be enforceable.

The pieces exist. The Council of Europe Convention is the strongest current treaty template, open beyond Europe, with one of the five required consents filed so far, by the European Union on 15 May 2026. The UN Global Dialogue gives every state a seat, though it is explicitly non-negotiating. UNESCO provides the broadest ethical baseline; the OECD and G7 provide operational convergence; ISO/IEC provides the evidence format. None of this is a world regulator, and none of it should be oversold. It is a credible pathway, and pathways are how international law has always been built: slowly, then suddenly.

For seventy years, "strategic technology" meant things that explode, encrypt or fly. In 2026 the list has a new entry, and you can tell because governments are behaving exactly the way they always behave around strategic technology: counting it, restricting it, subsidising it and writing security exceptions into treaties about it.

Four tells

First, the counting. Compute is now measured like a strategic stockpile. South Korea's AI law triggers safety duties at a numeric compute threshold. Export rules define chips by performance density. Analysts publish national capacity in gigawatts and dies. When states start measuring something in official units, it has become strategic.

Second, the restricting. Advanced chips are subject to United States export licensing, which shifted in January 2026 from a presumption of denial to conditional, case-by-case review, with ownership-based rules affirmed in June and preferential treatment extended to trusted partners in July. Access to compute is now a graduated, negotiated status between governments, which is precisely how strategic materials have always been handled.

Third, the subsidising. The EU's InvestAI aims to mobilise 200 billion euros; sovereign wealth funds closed a 49-billion-dollar AI fund and a 20-billion-dollar infrastructure partnership within seven months; hyperscaler capital spending guided toward roughly 700 billion dollars for 2026, which cited analyses put close to the investment intensity of the dot-com peak as a share of GDP. States describe these programmes in the language of independence, not profit: the stated goal of most sovereign-AI plans is not to out-build anyone, but not to depend on anyone.

Fourth, the carve-outs. The first international AI treaty contains national-security and defence exceptions. The UK renamed its evaluation body the AI Security Institute. The one domain where every governance framework goes quiet is the domain states consider most strategic. Silence, in treaty drafting, is a statement.

What follows from it

Three consequences matter for anyone doing compliance work. Supply chains become legal objects: where a chip may run, who may own the data centre and which subsidiaries count are now licensing questions, so provenance records and contract clauses about jurisdiction stop being boilerplate. Fragmentation becomes the default: a domestic-stack model trained wholly on one country's silicon and a sovereign cloud governed wholly by another's law are both rational responses to the same fear, and both are accelerating. And the civilian rulebooks inherit the tension: treaty carve-outs, filing regimes and evaluation institutes all draw the same line between what states regulate for citizens and what they reserve for themselves.

None of this is presented here as good or bad. It is presented as weather: the conditions under which every AI compliance programme now operates. Toto-Rex's practical read is unchanged from the rest of this site: build to the strictest regime that reaches you, document your supply chain like it will be asked for, and treat "where may this model run" as a legal question, because several governments already do.

The theory

The lazy map of AI regulation is a dial from "strict" to "lax", with Brussels at one end and everyone else strung along it. The evidence does not support the dial. What the world actually produced is four philosophies of governing anything powerful and new: the comprehensive statute (the EU, now echoed by South Korea, Taiwan, Vietnam and Peru), the principled regulator (the UK, Singapore, Hong Kong, Japan), the adversarial patchwork (the United States, where states, agencies and courts write the real rules), and the plural market of answers (Asia taken whole, from China's one-rule-per-behaviour ladder to India's considered restraint). Each is a faithful expression of the legal culture that produced it, which is exactly why none of them copied another.

Regimes drafted this differently still turn out to be commensurable, because they all answer the same six questions: what is the system's purpose, what data did it learn from, what decisions does it touch, what must be disclosed, who is accountable, and whose jurisdiction reaches it. That six-factor lens, developed in the book draft this essay leans on, is the working theory of this whole site: comparison is not about deciding who is right; it is the instrument that makes four right answers usable at once.

The current situation

As at September 2026 the scoreboard reads like this. One treaty exists and is not yet in force: the European Union ratified it on 15 May 2026 and remains the only party to have done so, while five consents, three of them from Council of Europe member states, are needed. That count is the single most watchable number in the field. One regional statute is binding and phasing in, with transparency duties in force since 2 August 2026. A statute club is quietly growing around it. The soft-law layer (OECD, UNESCO, G7) is broad, stable and increasingly operational, while the UN has built itself evidence machinery rather than a rulebook: a Scientific Panel and a Dialogue that is deliberately non-negotiating. Beneath all of it, data-protection law remains the enforcement muscle almost everywhere. And the industry has moved faster than every instrument named above: capital spending near historic records, frontier releases weeks apart, and compute treated by states as a strategic resource.

The approaches, honestly compared

Each philosophy pays its own price. The statute buys certainty and pays in weight; its 2026 simplification round proved even architecture bends. The principled regulator buys speed and pays in fragmentation; five principles across a dozen regulators means no single answer to "am I compliant?". The patchwork buys experimentation and pays in fifty answers where one was wanted. The market of answers buys fit and pays in translation costs for anyone crossing borders. Three gravities pull across all of them: Brussels through market size, Beijing through supply chains and filings, and the American states through sheer commercial reach. Distance from any of the three is not a defence.

Looking forward

The convergence that matters is not on instruments; it is on questions and evidence. Risk tiers, transparency labels, incident reporting and evaluation are appearing in every system under different names. The live questions for the next two years: whether the treaty crosses its ratification threshold; whether the UN Dialogue stays deliberative; whether incident reporting and model evaluation converge through the OECD, the G7 and the standards bodies, or fragment by bloc; and whether "sovereign AI" hardens from procurement language into law.

Possible solutions

For states, the credible path is the layered one this site keeps returning to: a thin binding floor for severe cross-border risks, interoperable evidence standards so one audit travels, national enforcement so constitutions keep their variation, shared science and incident learning, and capacity-sharing so the regime is representative enough to be enforceable. For everyone else, the solution already works today and does not wait for Geneva: read to the stricter rule. Build one system to the highest bar that reaches you, keep one proof file with jurisdiction annexes, appoint the local representatives, and treat every new instrument as a delta against the framework you already run, not a fresh headache. Four rulebooks, one floor, and the floor holds.