The European Union wrote the first comprehensive AI law, the United States has swung between executive action and state law, and China moved early on generative AI. Nobody has yet proved that any of these models governs a frontier technology that refuses to stand still.
Published 14 October 2026 · 06:00 GMT
Every government on earth now agrees on one thing about artificial intelligence: somebody should regulate it. They agree on almost nothing else. Not on what counts as dangerous, not on who should decide, and not on whether the point of the rules is to slow the technology down, speed it up, or simply make sure someone can be blamed when it goes wrong.
That disagreement is now the story. The European Union has built the most complete legal machine. The United States has produced the most powerful models and the least settled federal law. China has moved fastest on specific applications, above all generative AI, while keeping the state at the centre of enforcement. Around them, dozens of other countries are choosing which template to copy, adapt, or resist. The result is not one global rulebook but three rival ones, and the gaps between them are where the real power now sits.
The EU AI Act, adopted in 2024, is the most ambitious attempt yet to govern AI by statute. Its central idea is risk. Systems judged to pose an unacceptable risk, such as social scoring by public authorities or certain forms of biometric identification in public spaces, are banned. High-risk systems, including AI used in hiring, credit scoring, education admissions, and critical infrastructure, face obligations on data quality, documentation, human oversight, and testing before deployment. Limited-risk systems, such as chatbots, must disclose that users are talking to a machine. Everything else is left largely to voluntary codes.
The law also created a special regime for general-purpose AI models, the large foundation models that underpin chatbots and agents. Providers of the most capable models face extra duties on evaluation, risk assessment, and incident reporting. The design is deliberately tiered: the more general and powerful the system, the heavier the burden. Brussels has used this architecture before, in data protection and product safety, and the AI Act applies the same logic to a technology that changes far faster than either.
The strongest argument for the European approach is predictability. A company that complies in Europe knows what the law requires, and the Act gives regulators powers to investigate and fine. The strongest argument against it is speed. Critics in industry and in several European capitals say the compliance burden falls hardest on smaller firms and open-weight developers, while the largest labs can afford the lawyers. Supporters answer that predictable rules are what let markets form at all, and that the alternative, a decade of court cases, would be slower still. Both claims will be tested as the phased obligations take effect.
The United States has taken the opposite route. It has produced the dominant frontier labs, the largest cloud platforms, and the deepest pools of AI capital, and it has no single federal AI statute. Federal action has come through executive orders, agency guidance, procurement standards, and sectoral law, and the direction of that action has shifted with changes of administration. What one White House encouraged, the next has revisited. The result is a federal layer that is real but reversible.
Below the federal level, states have moved. Several have passed laws on AI in hiring, on deepfakes in elections, on automated decision-making, and on transparency for consumers. California, Colorado, Texas, and others have each taken different paths, and companies operating nationally must track them all. Industry groups call this a patchwork and warn that it fragments the American market. State legislators answer that Washington left a vacuum and that voters in their states expect protection now, not after Congress agrees.
The American debate also turns on a question Europe largely settled: whether to regulate the model or the use. American policy has leaned toward regulating applications and harms, such as discrimination or deception, rather than the underlying model itself. Defenders say this keeps the law technology-neutral and avoids penalising research. Critics say it misses the point, because a general-purpose model can be repurposed faster than any application-specific rule can follow. The fight over that distinction is the central argument in American AI policy.
China moved early on the applications that reached the public first. Its interim measures for generative AI services, issued in 2023, require providers of public-facing services to carry out security assessments, label AI-generated content, and ensure that training data and outputs comply with Chinese law. Earlier rules had already covered recommendation algorithms and deep synthesis. The pattern is consistent: regulate the deployed service, hold the provider responsible, and keep enforcement inside the state apparatus.
The Chinese approach has a second strand that outsiders often miss. Alongside content and security rules, Beijing has promoted AI development as a national priority, with industrial plans, standards work, and support for domestic champions. The model is not regulation instead of development; it is regulation as part of development, with the state setting the rails and the companies running on them. Foreign observers debate how much room that leaves for open research. Chinese officials answer that every country restricts some AI outputs, and that the difference is who sets the boundaries.
The practical consequence is a Chinese AI market that is large, fast-moving, and visibly bounded. Generative services reach hundreds of millions of users, but within rules that require real-name registration, content filtering, and licensing for public services. For global companies, the result is a split: products built for the Chinese market and products built for everywhere else, with different models, different data, and different guardrails.
Western coverage, from Brussels to Washington, reads the race as a contest over who sets the default. Europe bets that its market is large enough that global companies will build to its standard, the way they did with data protection. The United States bets that innovation outruns law, and that the country which builds the best models writes the de facto rules. Both bets can be partly right, and both carry a cost: Europe risks regulating itself out of the frontier race, while America risks discovering that the absence of rules is itself a choice with consequences.
The Western lens also stresses the safety institutes and the testing agenda that emerged from Bletchley Park. In this telling, the most important convergence is not legal but technical: governments learning to evaluate the same models with shared methods. If that cooperation holds, the argument goes, the three rulebooks can diverge on values while agreeing on evidence. If it fails, each bloc will test the same systems and reach different verdicts, and the divergence will harden into rival standards.
Eastern coverage reads the same race as a sovereignty story. From Beijing, the lesson of the last decade is that dependence on foreign platforms is a vulnerability, and AI is the most strategic platform yet. Regulation is therefore inseparable from industrial policy: the rules protect the public, and they also protect the space in which domestic firms grow. The generative AI measures are presented in this telling not as censorship alone, but as the price of a controlled, orderly rollout.
The Eastern lens also notes that the other blocs are converging toward its methods. Europe wants audits and registrations. America wants evaluations and incident reporting. Everyone wants some version of a licence to operate at scale. The differences are real, but the direction is shared: no major government now believes frontier AI can be left entirely to its makers. The argument is over the terms, not the principle.
Global South coverage asks the question the three rulebooks mostly skip: who is at the table. Most countries in Africa, Latin America, and South and Southeast Asia are rule-takers in AI, not rule-makers. Their data trains foreign models, their workers label foreign datasets, and their citizens use foreign services under terms set elsewhere. The African Union, ASEAN members, Brazil, and India have each begun drafting frameworks, but their leverage is limited by the concentration of compute, capital, and talent in a handful of countries.
The Southern lens therefore measures the race by a different yardstick: whether regulation opens access or entrenches dependence. Rules that require expensive audits can lock out smaller firms in poorer countries. Rules that mandate local data, local testing, or open weights can do the opposite. The South is watching for the version of AI governance that treats access to the technology itself as part of the bargain, not an afterthought.
