Regulating AI — why Europe, the United States, and China don't speak the same language

In brief

Article revised on 6 September 2026: the European timeline changed in July 2026, and the United States opened an internal conflict over who is entitled to regulate.

Regulating artificial intelligence is a global challenge — but the responses diverge radically across continents. Europe bets on risk-level classification. The United States oscillates between technological leadership and national protection. China prioritizes content control. These three approaches rest on different assumptions about what AI is — and what we should fear from it.

In short: three regulator philosophies, three metaphors. Europe as building architect: classify the rooms (living room, kitchen, fire escape) by risk level to set adapted standards. The United States as a CEO: build fast and fix afterwards if the house collapses. China as a librarian-censor: don’t regulate the house but what is said inside it. None of the three is complete, and their convergence is unlikely.

The regulator’s paradox

There is a tension at the heart of AI regulation.

Researchers measure the danger of a model through its emergent capabilities: what it can do that its creators did not anticipate, such as solving advanced chemistry problems or writing malicious code. These capabilities are difficult to predict, difficult to measure, and sometimes only appear after deployment.

Regulators, on the other hand, measure in FLOPs — floating point operations, the unit that quantifies the computing power mobilized to train a model. The EU AI Act sets the systemic risk threshold at 10²⁵ FLOPs. The American executive order of 2023 used 10²⁶. This is concrete, verifiable, and objective.

It is also an imperfect proxy. A model can be dangerous with few FLOPs if its architecture is efficient. A model can consume 10²⁶ FLOPs and have no problematic capabilities whatsoever. Kapoor and Narayanan (2024) documented these limitations in detail: the FLOPs threshold measures investment, not danger.

The regulator is constrained to measure what it can measure. This is the first structural difficulty of this field.

The European approach: classifying risk

The EU AI Act (Regulation 2024/1689, entered into force July 2024) is the world’s first binding legislation on AI. Its architecture rests on four risk levels:

• Unacceptable: prohibited. Social scoring, subliminal manipulation, real-time facial recognition in public spaces (with exceptions).
• High risk: permitted under strict constraints. AI in critical infrastructure, employment, education, justice. Requires technical documentation, human oversight, compliance before market placement.
• Limited: transparency obligations. A chatbot must declare itself as such.
• Minimal: no particular obligations. The majority of current AI applications.

For foundation models (GPAI — General Purpose AI), articles 51 to 56 introduce specific obligations: transparency on training data, copyright policy, assessment of systemic risks beyond the 10²⁵ FLOPs threshold.

The timeline moved — and that is the most useful thing on this page

As of 6 September 2026, the law is no longer what commentary from 2024 and 2025 described. Two movements in opposite directions occurred six days apart, and they must be told apart.

What genuinely applies since 2 August 2026. The transparency obligations of Article 50 came into application, and they depend on no risk tier: a system that converses with a person must tell them they are talking to a machine, unless it is obvious; synthetic content must carry a machine-readable mark; deployers must disclose emotion recognition, biometric categorisation and deepfakes. A short window runs to 2 December 2026, but only for marking systems already on the market before 2 August. The same day, the Commission gained its power to fine providers of general-purpose models: up to 3 % of worldwide turnover or 15 million euros, whichever is higher.

What was postponed. Regulation (EU) 2026/1744, the digital omnibus on AI, was published in the Official Journal on 24 July 2026 and entered into force on the 27th — six days before the deadline it moves. Obligations on stand-alone high-risk systems under Annex III shift from 2 August 2026 to 2 December 2027; those on systems embedded in products already covered by EU product-safety law (Annex I) to 2 August 2028. The same text adds two prohibitions to Article 5 — systems producing non-consensual intimate images and child sexual abuse material — with a transition to 2 December 2026, and widens the AI Office’s investigation and on-site inspection powers.

The reason given by the Commission is not a change of doctrine: the harmonised standards were not ready, and the conformity-assessment infrastructure the regulation presupposes did not yet exist. In other words, the text was written before the instruments meant to make it applicable.

What the postponement says about the rest. A sixteen-month delay on the core of the scheme — the high-risk part is the one touching employment, education, justice and critical infrastructure — while the transparency part starts on time, is an admission of sequencing: Europe knows how to require a disclosure; it does not yet know how to verify a conformity.

The underlying logic: AI is a tool, and like any tool, its danger depends on use. The law regulates use, not the technology itself.

The main limitation: who actually audits the FLOPs declared by labs? The technical capacity to verify the training of a large model cannot be decreed — the 2026 postponement is the administrative proof. Eiras et al. (2024) already raised this enforcement problem; two years on, it has been settled by the calendar rather than by method.

The American approach: competitiveness first

The United States has no federal law on AI. What it had was an executive order.

EO 14110 (October 2023), signed by Biden, mobilized the Defense Production Act to require safety reports beyond 10²⁶ FLOPs. It positioned NIST as a reference through the AI Risk Management Framework (RMF 1.0, 2023), and created the US AISI within NIST to coordinate frontier model evaluation.

In January 2025, Trump revoked this order (EO 14179) and replaced it with a text centered on competitiveness: removing barriers to American innovation, maintaining leadership against China.

This reversal illustrates the fragility of governance by presidential decree. An executive order does not create stable law — it can disappear within weeks. American regulation remains voluntary and sectoral: the commitments made by labs before the White House in 2023 (watermarking, red-teaming, information sharing) have no binding character.

The NIST RMF and ISO 42001 (2023) exist as best practice standards — but without obligation or sanction mechanisms.

December 2025: the federal level takes on the states, not the models

On 11 December 2025, executive order 14365, “Ensuring a National Policy Framework for Artificial Intelligence,” opened a third phase. Its object is not to regulate AI: it is to stop the states from regulating it. The text directs the Attorney General to set up a task force to challenge state AI laws as unconstitutional regulation of interstate commerce or as preempted by federal law; it authorises agencies to condition discretionary grants on a state refraining from enforcing such laws; and it asks Congress to legislate to make that preemption effective. Colorado’s algorithmic discrimination law is named explicitly. Placed out of reach: child protection, state government use of AI, and infrastructure regulation other than permitting.

The weakness of the device is the same as its predecessors’, and it is structural: preemption normally flows from an act of Congress, not from a decree. Congress has repeatedly declined general preemption — in the One Big Beautiful Bill Act as in the defence authorisation act. An executive order can steer the conduct of federal agencies; it does not repeal a state law.

Position as of 6 September 2026: the United States still has no federal AI statute, some thirty states have legislated, and federal energy goes into undoing that patchwork rather than replacing it. It is the exact inverse of the reproach Washington addresses to Brussels — not too many rules, but no national rule at all and an open conflict over who is entitled to make one.

The Chinese approach: regulating content

China has adopted an inverse strategy: not regulating AI as a technology, but regulating what it produces.

A chain of successive texts, each adding a layer to the last: algorithmic recommendations (2022), deep synthesis and deepfake rules (2023), the generative AI regulation (CAC, August 2023), then mandatory labelling of generated content (in application since 1 September 2025) and, the latest link, the interim measures on anthropomorphic interactive services, issued on 10 April 2026 and applicable since 15 July 2026 — the ones governing conversational companions: emotional boundaries, anti-addiction mechanisms, conduct when a user is in distress.

The 2023 regulation remains the most revealing of the method: it requires a security assessment before deployment, mandates that generated content respect “core socialist values,” and makes operators responsible for what their models produce.

Two details deserve a European reader’s attention. First, Chinese labelling is dual — a mark visible to humans, a fingerprint in the metadata for machines — which is exactly the architecture Europe’s Article 50 made mandatory eleven months later. Second, China legislates by successive addition rather than by a single act: where Europe wrote a 180-page regulation it is already amending, Beijing stacks short measures that apply quickly, at the cost of overall coherence.

The declared objective is not technical safety in the sense of systemic risks — it is information control. The central concern is: what does the model say to Chinese users? Not: how does it technically work?

This approach is consistent with the broader Chinese regulatory architecture on the internet. It is also radically different from the European and American frameworks, which makes international convergence difficult.

Open source as a coherence test

The open source debate reveals the internal contradictions of each approach.

The EU AI Act provides an exemption for open source models: GPAI obligations do not apply to models whose weights are freely published — unless they exceed the systemic risk threshold. The logic is that open models favor research, auditability, and competition.

The problem: the weights of a published model are irreversible. You cannot “withdraw” an open source model from the world once it is circulating. If dangerous capabilities are discovered after publication, there is no recall mechanism.

The other complication: “open source” is a vague term in this context. Llama (Meta) is not open source in the classic sense — its license prohibits commercial use beyond 700 million monthly active users. Bommasani et al. (2021) had already flagged this nomenclature problem: the gray area between “open” and “proprietary” is wide, and regulators struggle to define it precisely.

The Kapoor and Narayanan (2024) report goes further: it challenges the idea that opening weights mechanically increases risks. The correlation has not been empirically established. But neither has it been refuted — which leaves the regulatory debate in uncertainty.

Brussels Effect or fragmentation?

A structural question runs through all three approaches: who sets the global norm?

The “Brussels Effect” hypothesis suggests that European regulation imposes itself de facto on global companies, as with the GDPR. An American lab that wants access to the European market must comply with the EU AI Act — and it is often simpler to apply these rules everywhere than to have different versions per market.

The counter-hypothesis is that regulatory fragmentation is already here. Labs adjust their models by geography, American states legislate independently, China imposes its own rules. The result could be not a common norm, but an archipelago of incompatible regimes.

The outcome depends partly on enforcement capacity — and there, all approaches share the same problem: regulators lack experts capable of technically evaluating what they are supposed to supervise.

Comparative synthesis

State of play on 6 September 2026.

In short: these three systems do not complement each other — they compete. An American lab that wants access to the European market follows the AI Act. If it wants to enter China, it accepts content censorship. The likely outcome is not a global standard but an archipelago of regimes — each market imposing its framework, and models adapting per segment.

Key takeaways

• The EU AI Act classifies risks by use, not by technology. Since 2 August 2026, Article 50 transparency applies to everyone and providers of general-purpose models can be fined. The core of the scheme — high risk — was however postponed to December 2027 and August 2028 by Regulation 2026/1744, for want of harmonised standards and assessment infrastructure. Europe knows how to require a disclosure; it does not yet know how to verify a conformity.
• The United States opted for regulation by executive order — a reversible mechanism, as shown first by the January 2025 reversal and then by executive order 14365 of December 2025, which does not regulate AI but seeks to stop the states from doing so. No federal statute exists, and Congress has declined preemption twice.
• China regulates produced content, not the technology, and proceeds by successive addition: mandatory labelling in September 2025, rules for conversational companions in July 2026. Its dual labelling — visible and in the metadata — prefigures the architecture Europe’s Article 50 made mandatory eleven months later.
• The FLOPs threshold is a practical but imperfect proxy: it measures compute investment, not real capabilities or their dangerousness.
• Open source creates a regulatory gray area: the European exemption rests on a definition of “openness” that the sector itself has not stabilized.