Six answers to one question

Every government now regulating artificial intelligence is answering the same question — what a state should do about a technology whose deployments outrun any single review cycle — and arriving at genuinely different answers. This is not five variations on one design with different knobs turned. The EU has built an ex-ante, horizontal, risk-tiered statute. The US governs chiefly through sector-specific agency guidance and executive action a change of administration can reverse in a single order. China treats algorithmic systems as an extension of content and information governance, requiring registration and security review rather than a risk-based licence. The UK has declined to legislate in the EU’s manner at all, asking existing sector regulators to apply five non-statutory principles instead. Around all four sit two further layers that are neither law nor government policy: voluntary commitments developers have made to each other and to summit hosts, and international declarations and reporting mechanisms that fall well short of a treaty.

The temptation, faced with this list, is to rank the approaches by effectiveness. That temptation should be resisted. The approaches differ in scope, in what counts as success, and — critically — in how long they have actually been in force; a regime with years of enforcement data is not comparable to one whose central obligations do not apply until 2027. What follows is a structural comparison: what each approach attaches obligations to, what is currently in force versus proposed versus voluntary, and where a head-to-head verdict on “which works” is not available from the record as it stands.

The European Union: risk-tiered, horizontal, ex-ante — and partly deferred

The EU AI Act is the only major framework built as a single horizontal statute covering AI systems across sectors, keyed to a hierarchy of risk rather than to an industry or an agency’s existing jurisdiction. Four tiers do the work. At the top, Article 5 prohibits eight practices outright: manipulative or subliminal techniques that materially distort behaviour and cause harm, exploiting the vulnerabilities of specific groups, social scoring, biometric categorisation inferring race, political opinion, religion or sexual orientation, untargeted scraping of facial images to build recognition databases, emotion inference in workplaces and schools, predictive policing based solely on profiling, and real-time remote biometric identification in public spaces outside narrow law-enforcement exceptions [1]. Unlike almost everything else discussed here, this tier is not aspirational: it has applied since 2 February 2025.

ADVERTISEMENT

Below the prohibitions sits the high-risk tier. Article 6 sets two pathways: a system is high-risk if it is a safety component of a product already subject to EU harmonised product legislation requiring third-party conformity assessment, or if it falls within one of the use-case categories listed in Annex III — employment, education, access to essential services, law enforcement, migration and the administration of justice — unless a provider documents that its system performs a narrow procedural task, improves a completed human activity, or otherwise does not create significant risk, in which case it escapes the tier [2]. That escape clause matters: classification runs through a provider’s own documented judgement, reviewable but not independently verified at market entry for most systems.

Above this sits a general-purpose-model layer, keyed to the model rather than the use case, carrying its own documentation and systemic-risk duties — a design this publication’s companion piece on rule attachment covers in detail.

What has changed most since the Act’s obligations began phasing in is not the architecture but the calendar. On 19 November 2025 the European Commission proposed a Digital Omnibus amending the Act, and on 27 July 2026 that Omnibus entered into force. Its central effect is deferral: high-risk obligations for stand-alone Annex III systems now apply from 2 December 2027, and for AI embedded in regulated physical products from 2 August 2028, because the Commission judged the harmonised technical standards needed for conformity assessment were not ready [3]. It also narrowed the definition of “safety component” so systems that merely assist or optimise, without their failure creating a health or safety risk, do not automatically fall into the high-risk tier, and extended proportionate compliance treatment to small mid-cap companies previously limited to SMEs [3].

The honest summary is a split docket. Prohibitions and general-purpose-model transparency duties are live law with real compliance dates behind them. The high-risk tier — the part of the Act most often described in the press as simply “the EU AI Act” — will not bind most of the systems it names until the end of 2027 at the earliest. A comparison of “how the EU regulates AI” that treats the whole Act as equally in force today describes a statute that does not yet operate in practice.

A lateral row of four EU risk-tier case files on an office shelf, a date-stamp block caught mid-press against the high-risk folder's compliance-date label, the impression only half-transferred
Figure 1. The Act's tiers were fixed in 2024; the calendar behind them was rewritten in 2026 — the high-risk obligations most people mean by "the EU AI Act" do not bind until the stamped date says they do.

The United States: sectoral, executive-driven, and reversed by administration

Where the EU wrote one statute, the United States has produced a sequence of executive instruments and sector-specific agency guidance, and the sequence itself is now part of the comparison, because a change of administration reversed it wholesale.

ADVERTISEMENT

Executive Order 14110, issued in November 2023, required developers of dual-use foundation models to report to the federal government once training crossed defined compute thresholds — 102610^{26} operations generally, 102310^{23} for models trained primarily on biological sequence data — and required operators of sufficiently large computing clusters to report their existence and capacity [4]. It also directed agency work on safety testing, watermarking research and federal AI procurement. Three days after taking office, on 23 January 2025, the incoming administration issued Executive Order 14179, revoking 14110 outright and directing every agency to review actions taken under it and “suspend, revise, or rescind” anything inconsistent with a policy of sustaining American AI dominance free of what the order called ideological bias, giving OMB sixty days to revise its own memoranda accordingly [5]. The compute-reporting regime that had existed for fourteen months no longer exists in any federal instrument.

What replaced it, six months later, was not a new regulatory mechanism but a policy document. “Winning the Race: America’s AI Action Plan,” published by the White House in July 2025, sets out roughly ninety federal actions across three pillars — accelerating innovation, building AI infrastructure, and international AI diplomacy and security — and its posture toward regulation is explicitly subtractive: it directs agencies to identify federal rules that could be eliminated to promote AI innovation, and treats open-weight models as carrying geostrategic value worth promoting rather than a risk surface to report on [6]. On the states’ own AI statutes, the Action Plan does not attempt federal preemption by statute — a ten-year moratorium on state AI legislation was proposed in Congress in 2025 and did not pass — and instead directs the Office of Management and Budget to weigh a state’s AI regulatory posture when allocating discretionary funding, fiscal pressure rather than a legal bar [6].

What survives this reversal is the layer beneath politics. NIST’s AI Risk Management Framework, published in January 2023, remains in place across both administrations because it was written as voluntary guidance rather than a rule with a repeal switch, organising risk management into four functions — govern, map, measure and manage — for voluntary adoption by any organisation [7]. An instrument that binds nothing survives every change of administration precisely because there is nothing in it for a new administration to revoke.

States have not waited for this federal picture to settle. Colorado’s automated-decision-making statute — analysed in detail in this publication’s companion piece on rule attachment — illustrates the pattern: a developer-deployer structure enacted, then repealed and reenacted with implementation pushed to 2027, on a timeline set independent of whichever federal order happens to be current that year. The result is not federal law with state supplements; it is dozens of potential jurisdictions with no single federal floor, and an administration discouraging the more stringent among them through funding conditions rather than legal override.

A shelf of ring binders labelled with superseded federal AI policy documents, one binder half slid out of its slot while a taller replacement binder is caught half slid in beside it, neither yet seated
Figure 2. A single order revoked the reporting regime a prior order had built; the binder that named the compute thresholds is the one now being slid out of the row.

China: algorithm registration and content governance as one administrative logic

China’s framework does not organise itself around risk tiers or agency sectors at all. Its central mechanism is registration: an algorithm or a generative-AI service meeting specific criteria must be filed with the state before or shortly after it operates, and the regime predates the international “AI governance” conversation that took off after 2022.

The foundational instrument is the Cyberspace Administration of China’s Provisions on the Administration of Algorithm Recommendation of Internet Information Services, effective 1 March 2022. It applies to any operator using algorithmic technology — recommendation, ranking, filtering, scheduling — to shape what Chinese internet users see, covering platforms from social feeds to e-commerce ranking and food-delivery dispatch, foreign-operated services included [12]. Its obligations mix consumer-protection duties — letting users view and delete the tags used to profile them, banning behavioural price discrimination — with content obligations: operators must promote what the rule calls mainstream values, prevent the algorithm from manipulating trending lists or engagement counts, and detect illegal content [12]. Operators whose algorithms carry what the rule calls public-opinion properties or social-mobilisation capacity — in practice, anything with real reach — must file with a national registry within ten working days of launch, publish the resulting filing number, and update it within five days of any material change [13].

ADVERTISEMENT

The 2023 Interim Measures for Generative AI Services layered a parallel regime on top for text, image, audio and video generation. Their scope is explicitly public-facing: providers offering generative AI services to the public inside mainland China are covered, while internal research and development is expressly excluded [13]. Providers must use lawfully sourced data, respect intellectual property, obtain consent for personal data, and meet data-quality obligations around authenticity and objectivity. Services carrying public-opinion attributes or social-mobilisation capacity must additionally pass a security assessment before the same ten-day filing applies [13].

The comparison worth drawing is not “China regulates AI more or less strictly than the EU.” China’s regime answers a different question. The EU’s tiers ask what risk a system poses to health, safety or fundamental rights. China’s filing and security-assessment triggers ask whether a system can shape public opinion or mobilise people at scale — an objective centred on information control and social stability, administered through pre-launch registration rather than post-market classification [23]. A registration number and an EU conformity file are not two implementations of one instrument; they answer to different theories of what needs watching.

An algorithm-filing ledger open on a desk with a row of completed numbered entries, a fresh entry's filing stamp caught in mid-descent just above the page, its impression not yet made
Figure 3. The filing number is what makes an algorithm legible to the registry; until the stamp lands, the entry beside it is provisionally unregistered.

The United Kingdom: principles without a statute, regulators without a rulebook

The UK’s starting position, set out in its March 2023 white paper, was a deliberate rejection of the EU’s model. The paper declined to create a dedicated AI regulator, reasoning that a new cross-sector body would duplicate expertise sector regulators already hold, and proposed five cross-sectoral principles instead — safety and robustness; transparency and explainability; fairness; accountability and governance; contestability and redress — for existing bodies such as the financial, medicines and communications regulators to apply within their own remits [10]. The principles were issued on a non-statutory basis from the outset, with a statutory duty requiring regulators to “have due regard” to them floated only as a possible later step [10].

That later step has not arrived. A private member’s Artificial Intelligence (Regulation) Bill was relaunched in the House of Lords and passed its first reading there in March 2025, but it is not a government bill and has not become law. The government’s own promised legislation for the most powerful models — repeatedly signalled, including around possible statutory powers for the UK’s AI Security Institute — has not been introduced, and as of 2026 there is still no single UK AI Act [11]. In February 2025 the government restated its position that most AI systems should be regulated at the point of use by existing expert regulators, a formulation that reads less as a new decision than as confirmation that the 2023 default has held [11]. In practice, AI use in the UK falls under a patchwork of instruments mostly not written with AI in mind — UK GDPR, an Information Commissioner’s Office code of practice, the Online Safety Act, sector guidance — plus the extraterritorial reach of the EU AI Act for any UK firm serving EU users [11].

This is a genuinely different design choice from the EU’s, not an earlier version of it. Where the EU asks which risk tier a system’s use case belongs to, the UK’s live position asks which existing regulator already has jurisdiction, and whether the AI system changes that regulator’s analysis. The tradeoff is explicit in the white paper’s own framing: flexibility and continuity of existing expertise, purchased against the absence of one predictable cross-sector standard and, for now, any statutory teeth behind the five principles at all.

Five printed cross-sectoral principle cards laid in a row on a desk, one card caught mid-lower onto a sector regulator's guidance booklet with a corner still lifted clear of the page
Figure 4. The UK's principles have no rulebook of their own; they are checked, case by case, against whatever a sector regulator already had on its own shelf.

Voluntary commitments: safety policy as private ordering

Beneath government regulation sits a layer that is neither law nor government guidance: commitments AI developers have made to each other, to summit hosts, or to a standards body, none of which any government compels.

The clearest government-convened example is the Frontier AI Safety Commitments, signed by twenty organisations — including Amazon, Anthropic, Google, IBM, Meta, Microsoft, OpenAI and Samsung, with others such as NVIDIA added subsequently — at the AI Seoul Summit in May 2024. The text is explicit about its own status: signatories “undertake to develop and deploy their frontier AI models and systems responsibly, in accordance with the following voluntary commitments” [16]. Those commitments include assessing risk across a model’s lifecycle, setting internal thresholds for intolerable risk, publishing a safety framework describing how those thresholds are operationalised, and disclosing safety information publicly “except insofar as doing so would increase risk or divulge sensitive commercial information” [16] — a carve-out that leaves each signatory the final word on what it discloses.

Individual companies have built comparable frameworks independently of any summit. Anthropic’s Responsible Scaling Policy, first published in 2023 and now in its fourth revision, defines AI Safety Levels with capability thresholds that trigger specific safeguards — the company activated ASL-3 protections in May 2025 after concluding it could no longer support a case that the relevant catastrophic-misuse risk was low [19]. The policy is, again explicitly, self-imposed: Anthropic states its purpose is to offer “an example of a framework that others might draw inspiration from,” not a standard anyone is bound to meet [19]. Several other frontier labs maintain comparable documents under different names, none enforced by a regulator.

A third register sits between these two: certifiable but still optional standards. ISO/IEC 42001, published in December 2023, is the first international management-system standard for AI, specifying requirements for governing an AI system’s lifecycle — risk management, impact assessment, supplier oversight — that an organisation can have independently audited and certified, voluntarily, by an accredited body [20]. The G7’s Hiroshima Process points the same way from the intergovernmental side: its Code of Conduct for advanced AI developers is voluntary guidance, and the OECD-built Reporting Framework launched in February 2025 lets signatories publish structured self-reports against it, explicitly to facilitate transparency and comparability rather than to enforce anything [9].

All three registers share one structural gap: each framework is written and, in practice, graded by the organisation it covers. There is no independent verification requirement, no accreditation regime for third-party safety auditors, and no penalty beyond reputational cost for narrowing a commitment quietly. That does not make the commitments meaningless — publishing a threshold is a fact a critic can hold a company to — but it is a different kind of accountability from a conformity file reviewed by a market-surveillance authority, and the two should not share the vocabulary of “compliance.”

A shelf of frontier-lab safety-framework ring binders, one pulled forward and open to a threshold page where a small version plate on the spine is caught half-lifted, a newer plate not yet seated beneath it
Figure 5. A responsible-scaling policy binds only the organisation that wrote it, and only until it writes the next version; the plate on the spine is the whole of the record that anything changed.

International coordination: declarations, not treaties

No binding international treaty governs AI, and the closest thing to a shared foundation predates the summit era entirely. The OECD’s AI Principles, adopted in May 2019 and updated in May 2024, are non-binding but have been adhered to by 47 governments including every OECD member and the EU, and their definition of an AI system has been borrowed into the EU AI Act, US federal guidance and UN texts even where nothing else about those instruments aligns [8]. What has followed since 2023 is a widening set of political declarations, expert-body reports and monitoring mechanisms built on that shared vocabulary, of increasing specificity but still consistently non-binding form.

The sequence began with the Bletchley Declaration, signed by 29 countries and the EU at the UK’s AI Safety Summit in November 2023 — a genuinely broad list including the United States, China and the EU together, which was itself the news. Its text commits signatories to collaborate on identifying frontier-AI risk, develop risk-based domestic policy while noting approaches “may differ based on national circumstances,” and support an internationally inclusive research network on frontier AI safety [14] — on its face a statement of intent rather than a binding instrument, leaving each country’s actual policy response to national discretion.

The AI Seoul Summit in May 2024 produced the Seoul Declaration, signed by a narrower group including the EU, US, UK, Australia, Canada, Germany, France, Italy, Japan, South Korea and Singapore, adding a specific and, in retrospect, prescient theme: the importance of interoperability between different countries’ AI governance frameworks, precisely because those frameworks were not going to converge on one design [15]. Seoul is also where the industry-facing Frontier AI Safety Commitments discussed above were signed.

The fracture became visible at the next summit. The Paris AI Action Summit in February 2025 produced a joint declaration on inclusive and sustainable AI that 61 countries and blocs — including, notably, China — signed. The United States and the United Kingdom did not. The US delegation objected to its emphasis on multilateral governance and its environmental and inclusivity references, with the US vice president telling delegates that excessive regulation would stifle innovation; the UK said the text lacked practical clarity on governance and did not adequately address national security [17]. It was the first time two of the series’ founding hosts publicly declined to sign the product of their own process: whatever convergence the Bletchley-to-Seoul sequence appeared to be building, Paris showed it was not secure.

The most recent entry, the India AI Impact Summit, held in New Delhi on 19-20 February 2026, made the divergence explicit at the level of framing rather than a single missing signature. Its official framing describes a shift toward addressing the “Global AI Divide” — the concentration of AI infrastructure among a small number of countries — organising the agenda around three principles named People, Planet and Progress and seven working groups spanning human capital, inclusion, safe and trusted AI, resilience, science, resource access and economic growth, with delegations from more than 100 countries [18]. Safety is one working group among seven oriented mainly toward development rather than the organising theme it was at Bletchley.

Multilateral bodies have moved in parallel but slower. The UN General Assembly adopted its first resolution on AI in March 2024 — on safe, secure and trustworthy AI for sustainable development — led by the United States and co-sponsored by more than 120 states, adopted by consensus [21]. Like all General Assembly resolutions it is a recommendation, not a binding obligation; its content calls on states to promote safe AI and bridge digital divides rather than creating any institution. The UN Secretary-General’s High-level Advisory Body on AI followed in September 2024 with a report, “Governing AI for Humanity,” whose seven recommendations deliberately avoid a heavyweight new institution, calling instead for “light institutional mechanisms” to complement existing efforts [22] — restraint that is itself a data point: even the body charged with imagining global AI governance concluded a single strong regulator was not what the moment supported.

A summit secretariat's preparatory desk with a draft joint declaration marked with redline edits, a row of small folded delegation place-cards along the table edge, one card caught mid-turn face up
Figure 6. Not every delegation turns its card face up on a joint declaration; the third summit in this series is the one where two of its own founding hosts did not.

Why a head-to-head ranking cannot be built from this material

This publication’s editorial standard is to flag incomparability rather than force a ranking from conditions that are not alike, and this comparison is close to a textbook case of why.

Start with the plainest problem: elapsed enforcement time. China’s algorithm-filing regime has run since March 2022 — more than four years of registrations, security assessments and enforcement activity. The EU’s high-risk tier, the part of its Act closest in spirit to a general risk-based licence, will not bind most Annex III systems until December 2027 at the earliest [3]. The UK’s five principles have applied, non-statutorily, since 2023, but a regulator applying a principle it was already free to apply under its existing powers is a different claim from enforcing a new statute. Comparing outcomes today would mean comparing a system with years of live administrative data against one with close to zero years of its central mechanism’s operation, whose independent contribution is hard to isolate from what regulators were doing anyway. Effectiveness is only meaningful once the thing measured has run long enough, under comparable conditions, to produce an outcome — informally,

Ei≈f(coveragei, enforcementi, ti),ti≪tj for several pairs (i,j) E_i \approx f(\text{coverage}_i,\ \text{enforcement}_i,\ t_i), \qquad t_i \ll t_j \text{ for several pairs } (i, j) ↗

where tit_i↗ is time-in-force for approach ii↗. When tit_i↗ differs by an order of magnitude or more across the set compared, as it does here, no amount of careful measurement of coverage or enforcement intensity fixes the fact that the outcome variable has not had time to be generated for every entrant. This is why no figure in this article, or elsewhere in this publication, ranks these regimes by effectiveness.

Second, the regimes do not target the same harm. The EU’s tiers are built around risk to health, safety and fundamental rights. China’s triggers are built around public-opinion influence and social-mobilisation capacity. The UK’s principles name fairness and contestability, categories with no equivalent trigger in China’s filing rules. Al-Maamari’s cross-regional comparison of these frameworks reaches a compatible conclusion from a policy-analysis angle: effective AI governance has to be globally informed yet context-sensitive, and no single framework is a universal benchmark the others should be measured against [24]. A regime optimised for content control and one optimised for individual-rights risk are aimed at different targets; declaring one “more effective” requires first deciding whose target counts — a political judgement dressed as a technical one.

Third, the voluntary and international layers cannot be scored by the same method as statutes, because there is no enforcement record to examine, only disclosure a signatory itself controls [16, 19]. Judging their effectiveness against a statute means comparing a promise to a law: different remedies for breach, and for the promise, often none beyond reputational cost.

None of this means the approaches cannot be described, compared structurally, or criticised on their own terms — the rest of this article has done exactly that. It means a ranked league table of “best AI regulation” is not a defensible output of the evidence that exists in 2026, and any source presenting one is offering an opinion, not a finding.

Predictions, with the observations that would falsify them

These are forecasts, separated from the sourced analysis above. Horizon: 16 August 2029. Shared assumptions: no catastrophic AI incident triggers emergency multilateral treaty-making; no jurisdiction abandons its current architecture wholesale; summit coordination continues on roughly its current cadence.

One. The EU’s high-risk tier will be deferred again, or narrowed further, before it takes effect in December 2027, rather than applying on schedule. Indicator: a further Commission proposal or delegated act touching Annex III scope or timing, filed before December 2027. Disconfirmed if the high-risk obligations enter into force on 2 December 2027 exactly as set by the July 2026 Omnibus, with no further amendment.

Two. No binding international AI treaty will exist by the horizon date; coordination will remain declarations, reporting frameworks and voluntary commitments rather than a ratified enforcement instrument. Indicator: any multilateral body opening formal treaty negotiations with a defined ratification path. Disconfirmed if such negotiations are underway with states committed to a ratification timeline.

Three. US federal AI policy will keep using funding conditionality rather than statute to influence state law, because a direct preemption attempt failed in Congress in 2025 and the cross-party state-autonomy concerns behind that failure have not obviously changed. Indicator: a preemption bill reaching a floor vote. Disconfirmed if Congress passes, and the President signs, a statute directly preempting state AI legislation.

Four. The UK will not have passed a comprehensive government AI statute by the horizon date; its regulator-led default will still be operative, whether or not a statutory “due regard” duty has finally been introduced. Indicator: a government, rather than private member’s, AI bill with a second reading date set. Disconfirmed if such a bill has received Royal Assent by 16 August 2029.

What to take away

Four governments, a coalition of companies, and a loose set of international bodies are all, in 2026, doing something they each call AI governance, and the word is carrying more weight than it can bear. The EU is writing product-safety law for a technology that changes faster than its own review cycle, and has just admitted as much by deferring its own central tier. The US is running a policy experiment in which federal AI governance depends on which party last won an election, its durable parts being exactly the parts nobody can be compelled to follow. China is extending an existing information-control apparatus to a new class of systems. The UK has bet that its existing regulators, lightly coordinated, can do the job without a new statute, and has not tested that bet against a serious enforcement dispute. Voluntary frontier-safety commitments and a certifiable management standard fill some of the space between these positions, but they audit themselves rather than being audited. And the summit series that once looked headed toward something binding produced, at its third meeting, two founding members declining to sign.

The useful question is not which of these will turn out to have been right. It is which specific gap — enforcement timing, cross-border reach, independent verification, treaty-level coordination — a given AI system’s real behaviour is currently falling through, because the answer, in 2026, is often more than one.