Anu Bradford’s 2012 Northwestern University Law Review article identified the precise conditions under which a single jurisdiction can globalise its regulatory standards without treaties, coercion, or the consent of other states. The framework was constructed from the EU’s record in antitrust, privacy, chemicals, and food safety. It was not built with artificial intelligence in mind. But the EU AI Act — Regulation (EU) 2024/1689, published in the Official Journal on 12 July 2024 — presents the framework with its most consequential test case since GDPR. The question this reading asks is whether the AI Act satisfies Bradford’s five conditions, which conditions it satisfies cleanly, which are structurally contested, and what the analysis implies for a specific policy problem that has received less attention than it deserves: whether U.S. federal agencies operating under the NIST AI Risk Management Framework can demonstrate substantial compliance with AI Act high-risk requirements through publicly available documentation, or whether the two frameworks are structurally incommensurable in ways that no amount of good-faith voluntary alignment can bridge. The Brussels Effect framework does not answer that question directly. But it supplies the vocabulary for understanding why the answer is harder than either the proponents of regulatory convergence or the sceptics of EU overreach have so far acknowledged.1
↑ N° 22 · Continues from the foundational annotated reading of Bradford’s 2012 article — the conditions, mechanisms, and limits of the Brussels Effect as theory.The framework before the test
Bradford’s five-condition framework is not a checklist. It is a logical structure in which each condition is individually necessary and all five are jointly sufficient. Any application to a new domain must begin by holding that structure firm before loading it with empirical content.
The five conditions Bradford specifies are: market power sufficient to make exclusion from the EU costly; regulatory capacity to produce and enforce binding rules; a domestic preference for strict standards that prevails over competing preferences by virtue of being more demanding; a predisposition to regulate targets that are structurally inelastic — that cannot escape the strict regime by relocating; and nondivisibility of the regulated conduct or production, whether through legal indivisibility, technical impossibility, or the economics of uniform-standard cost advantage.2
The conditions do not operate independently. Nondivisibility is the mechanism through which the preceding four conditions generate global effects: it is what converts a regulation that formally applies only within EU territory into one that actually governs behaviour worldwide. A jurisdiction can satisfy the first four conditions and still fail to generate a Brussels Effect if the regulated domain is divisible — if firms can cheaply maintain a European-compliant configuration and a non-European configuration simultaneously. Conversely, a domain characterised by strong nondivisibility but weak market power or inadequate regulatory capacity will generate only partial or provisional effects, as the GMO case demonstrated. The framework is multiplicative in structure: the five conditions must all be present, and the strength of the Effect is a function of how strongly each is satisfied.
The AI Act’s extraterritoriality provision — Article 2(1)(c), which applies the regulation to providers and deployers established in third countries whose systems are placed on the EU market or whose output is used within the Union — is the structural analogue of the extraterritorial application clauses in GDPR, REACH, and the EU’s competition rules. It is the legal mechanism that enables the Brussels Effect, though it does not by itself generate it. What generates the Effect is the combination of that extraterritorial reach with market power sufficient to make compliance cheaper than exclusion, regulatory capacity sufficient to make enforcement credible, and nondivisibility sufficient to make dual-standard architectures uneconomical. Whether those conditions are satisfied for AI is what the analysis must establish.
Market power and regulatory capacity — the uncontested conditions
Two of Bradford’s five conditions are satisfied by the EU AI Act without serious analytical contest. The argument for their presence does not require heroic assumptions about European market size or institutional development; it is grounded in the same structural facts that made GDPR a global standard without a treaty.
Market power is the EU’s most durable asset in regulatory globalisation, and its force in the AI context is at least as strong as in data protection. The EU’s internal market of roughly 450 million consumers represents a substantial share of global AI deployment for most commercial applications: enterprise software, financial services, healthcare diagnostics, insurance underwriting, hiring and personnel management, and the public-sector applications that Annex III places in the high-risk category. Firms building AI systems for global markets cannot afford to treat the EU as an optional market; the opportunity cost of exclusion from a market of that affluence and scale is prohibitive for most commercial AI providers.3 The Annex III list of high-risk use cases is, by design, concentrated in exactly the sectors where EU market access is most valuable — healthcare, finance, employment, and public administration are each sectors in which European demand is large and EU-specific regulatory requirements have historically been among the most demanding globally.
Bradford’s analysis of market power distinguishes between absolute market size and the ratio of exports to the strict jurisdiction relative to sales in home and third markets. This ratio matters enormously in the AI context and produces a distribution. For large U.S. technology firms — the hyperscalers and major enterprise software vendors — the EU represents a substantial but non-dominant share of global revenue. For mid-tier providers, particularly those building vertical AI applications in regulated sectors, the EU may represent the largest single compliance environment they face. For open-source foundation model developers who do not directly sell to EU users but whose models underlie EU-deployed systems, the market power condition is mediated through their downstream deployers. The point is that market power is unambiguously present; its precise force varies by firm size and business model.
Regulatory capacity is perhaps the most interesting of Bradford’s conditions to evaluate in the AI context, because the EU AI Act represents a significant expansion of EU regulatory competence into a genuinely novel domain. Bradford’s 2012 analysis emphasised that regulatory capacity requires not just the institutional authority to produce rules but the technical expertise and enforcement infrastructure to make those rules credible. She noted that China’s 2008 antitrust law had created regulatory authority on paper but that the practical enforcement gap between China’s competition agencies and the European Commission measured in decades, not years.4
The EU AI Act’s institutional architecture is more developed at enactment than either GDPR or REACH were at analogous stages. The European AI Office, established within the Commission by Decision of 24 January 2024, has a mandate covering general-purpose AI model oversight, scientific advice, and international engagement. National supervisory authorities in each member state carry primary enforcement responsibility for most high-risk AI obligations. The AI Board coordinates across member states. The Commission retains jurisdiction over systemic-risk GPAI models and can impose fines of up to seven per cent of global annual turnover for the most serious violations. The formal structure is dense and, at this stage, substantially untested — enforcement experience at the scale that makes regulatory capacity genuine rather than nominal will take several years to accumulate. But the institutional scaffolding is more robust at the outset than anything the EU had in place when the 1995 Data Protection Directive was adopted, and that Directive ultimately generated a fully operational Brussels Effect in data protection.5
The distinction Bradford draws between regulatory capacity and regulatory propensity is relevant here. Capacity is institutional; propensity is political. The European Parliament’s long record of advocacy for fundamental-rights-based technology regulation, the Commission’s sustained investment in AI policy since the 2018 AI strategy, and the political salience of AI governance in all major EU member states together constitute a strong signal that the regulatory propensity is genuine — that the strict standards embedded in the AI Act reflect durable political preferences rather than a transient political moment.
Preference, inelasticity, and the conditions that constrain
The third and fourth conditions are satisfied in the AI Act context, but each carries qualifications that the 2012 framework did not need to address in the same form. The AI domain introduces structural features that make these conditions more complex to evaluate than they were for chemicals or data protection.
Preference for strict rules is the political condition, and it holds clearly for the EU AI Act. Bradford’s analysis located the EU’s regulatory stringency in its commitment to the precautionary principle, the constitutional centrality of the social market economy, and the relative ideological cohesion of European political elites on fundamental-rights questions. All three of those drivers are visible in the AI Act’s design. The Act applies the precautionary logic embedded in the EU’s approach to GDPR: the burden falls on providers to demonstrate conformity before deployment, not on regulators to demonstrate harm after deployment. The fundamental-rights framing is explicit in recitals and in the design of the prohibited-practices category, which addresses systems that manipulate, exploit, or suppress individual autonomy in ways the Act characterises as incompatible with EU values. The ideological convergence on fundamental-rights protection as a prerequisite for legitimate AI deployment has been sustained across multiple Commission presidencies and parliamentary terms.
The United States’ position relative to the EU on AI regulation mirrors — with some complications — the pattern Bradford described for consumer and environmental protection. Federal AI regulation in the United States remains fragmented: no general-purpose federal AI Act equivalent exists; the Biden administration’s October 2023 Executive Order on AI was substantially rescinded by the Trump administration in January 2025; NIST’s AI RMF is voluntary; and the primary federal AI governance instruments are sector-specific guidance documents with no binding force. State-level regulation is proliferating but heterogeneous. This regulatory posture is structurally more permissive than the EU’s — not necessarily in every individual use-case rule, but in the fundamental enforcement architecture of mandatory conformity assessment, pre-deployment documentation, and national supervisory authority enforcement. The EU is the stricter regulator by the criterion that matters for the Brussels Effect: it has a binding, comprehensive, extraterritorial framework where the United States has a patchwork of voluntary guidance, sector-specific rules, and state-level experiments.6
The structure of regulatory conflict ensures that the stricter regulator prevails, not through coercion, but because the market logic of the situation makes compliance cheaper than exclusion once the strict regulator’s market is large enough to foreclose the exit option. — Bradford (2012), applied
Inelastic targets is the condition that works differently for AI than for chemicals or data. Bradford’s central observation was that the EU avoids Brussels Effect failure by regulating consumer markets — targets that cannot move — rather than capital markets, which are mobile. The logic is straightforward: if you regulate products, the consumers who want those products are where they are; the firm must comply or exit the market. If you regulate capital, the capital can move.
AI systems occupy a hybrid position. The compute infrastructure — the training clusters, the data centres — is mobile in a way that consumer products are not. A firm prohibited from operating certain AI configurations within the EU can, in principle, route inference through non-EU infrastructure. But for high-risk AI systems as defined by Annex III, the relevant target is not the infrastructure but the output: an AI system used for credit scoring in France, for medical diagnosis in Germany, or for recruitment in the Netherlands is regulated at the point of output, not the point of computation. The Annex III categories are, overwhelmingly, defined by the application domain and the EU population affected, not by the location of the servers. The AI Act is therefore structured to regulate inelastic targets — the EU persons and organisations that are the intended users and subjects of high-risk AI systems — and it does so regardless of where the underlying model was trained or hosted.7
There is a genuine contestability in this condition for general-purpose AI models. A foundation model developer that does not directly deploy in the EU, but whose model is accessed by EU deployers via API, faces a more complex regulatory situation. The AI Act’s treatment of GPAI model providers is addressed in Articles 51–56 and imposes obligations primarily on providers of models with systemic risk. For foundation model providers without significant EU market exposure, the inelasticity argument is weaker. For providers of high-risk application-layer systems — the Annex III use cases — it is strong.
Nondivisibility — where the prediction is made
Nondivisibility is the condition that converts the first four into a Brussels Effect. It is also the condition that is most analytically interesting in the AI context, because AI systems exhibit all three forms Bradford identified — legal, technical, and economic — in ways that make dual-standard architectures especially difficult to maintain.
Bradford distinguished legal nondivisibility (the regulated action cannot be partitioned across jurisdictions), technical nondivisibility (the architecture of the activity does not permit jurisdictional segregation), and economic nondivisibility (scale economies make dual production uneconomical even where it is legally and technically feasible).
Legal nondivisibility is the weakest of the three forms for AI, in contrast to its primacy in the merger control context. An AI system can, in principle, be deployed in a jurisdiction-specific configuration — a different model card, a different conformity assessment, a different documentation package. Unlike a merger, which either proceeds or does not, an AI system’s deployment is modular enough to permit some degree of jurisdictional customisation. The legal nondivisibility argument is strongest for AI systems whose underlying model architecture is assessed as part of the conformity assessment process, because a single model cannot simultaneously satisfy and not satisfy a structural requirement. But it is not absolute in the way merger control is.
Technical nondivisibility is considerably stronger, and arguably the dominant form for most commercial AI applications. The AI Act’s high-risk obligations — risk management system, data governance requirements, technical documentation, logging and traceability, transparency to users, human oversight mechanisms, accuracy and robustness requirements — are system-level obligations that apply to the design and architecture of the AI system, not to its geographical deployment. A firm that maintains a compliant EU version and a non-compliant non-EU version of the same AI system is maintaining two substantially different systems: different training pipelines (for the data governance requirements), different monitoring infrastructure (for the logging and traceability requirements), different user interface design (for the transparency and human oversight requirements), and different technical documentation (for the conformity assessment). The cost of that divergence is not merely the direct cost of compliance; it is the ongoing engineering and operational cost of maintaining two architectures in parallel.
For most commercial AI providers, this cost is prohibitive. The more efficient strategy — and the one that GDPR established as the industry norm for data protection — is to build to the strictest standard and deploy that standard globally. The result is economic nondivisibility reinforced by technical complexity: the EU standard propagates because the alternative is too expensive to maintain.
Economic nondivisibility reinforces the technical argument through a mechanism Bradford documented in the chemicals and environmental regulation contexts: the supply-chain dynamics of a globalised industry. The major AI supply chains — foundation model providers, infrastructure providers, enterprise software integrators, sector-specific application developers — are globally structured. A hyperscaler that builds its infrastructure to EU AI Act compliance standards for its EU business has a strong incentive to apply those standards globally: it avoids the operational complexity of jurisdiction-specific configurations, it avoids the reputational and legal risk of running two-tier AI governance, and — in an industry where enterprise customers increasingly require AI governance attestations as a procurement condition — it can use EU compliance as a global signal of governance quality.
The aggregate assessment is that the nondivisibility condition is substantially satisfied for high-risk AI systems as defined by Annex III, and that the Effect is likely to propagate through the same economic logic that drove GDPR compliance — not because firms prefer EU standards, but because maintaining dual architectures is more expensive than applying the EU standard universally.
What the framework predicts for interoperability
The five-condition analysis predicts a Brussels Effect for high-risk AI systems. But the prediction is not uniform, and applying it to the specific question of U.S. federal agency compliance with AI Act requirements reveals a structural asymmetry that the convergence literature has not adequately addressed.
If the Brussels Effect analysis is correct, private AI providers operating globally will progressively converge on EU AI Act compliance standards for high-risk applications, whether or not the United States adopts equivalent federal legislation. This is the pattern GDPR established: by 2022, major U.S. technology firms had implemented GDPR-aligned data protection architectures globally, had appointed data protection officers across their organisations, and had restructured consent and rights-exercise workflows for all users regardless of jurisdiction — not because U.S. law required it, but because the cost of EU-specific configurations exceeded the cost of universal compliance. The AI Act is likely to produce an analogous dynamic, with a lag of several years while the Act’s enforcement mechanisms mature and the conformity assessment ecosystem develops.
The interoperability question, however, is not primarily about private firms. It is about whether U.S. federal agencies — which procure, develop, and deploy AI systems within the scope of Annex III — can demonstrate compliance with AI Act requirements through the documentary and procedural infrastructure they have built under the NIST AI RMF. This question has a different structure from the private-sector Brussels Effect, because federal agencies cannot simply adopt EU compliance frameworks the way a private firm can. They operate under U.S. administrative law, U.S. government-wide policies, and U.S. procurement regulations that prescribe the form of their AI governance documentation. Their AI governance documentation is produced for internal accountability purposes and for congressional oversight, not for third-party EU conformity assessment bodies. Much of their most detailed AI governance documentation — risk assessments, security evaluations, operational data — is classified or otherwise restricted from disclosure to entities outside the U.S. government.
The structural asymmetry that Bradford’s framework helps diagnose is this: the Brussels Effect predicts convergence through market mechanism, but that mechanism operates on actors who are willing to adjust their conduct in order to retain EU market access. U.S. federal agencies are not seeking access to the EU market. They are seeking to procure and deploy AI systems that serve U.S. government functions. Their Brussels Effect exposure is indirect — it arises when they procure AI systems from private vendors who are themselves subject to the Brussels Effect, and who have therefore built EU-compliant systems that also happen to be the systems the agencies procure. In this sense, U.S. federal agencies may achieve substantial de facto AI Act alignment not because they have sought it, but as a byproduct of procuring from vendors who have achieved it for commercial reasons.
The interoperability gap that remains after this de facto alignment is structural, not incidental. It concerns the documentary and procedural dimension of AI Act compliance: the requirement for technical documentation in a form accessible to EU national supervisory authorities, the EU database registration requirement, the mandatory conformity assessment by notified bodies or self-declaration in prescribed formats, and the post-market monitoring obligations with EU-mandated incident reporting. These are not requirements that de facto vendor alignment addresses. A U.S. federal agency can procure a GDPR-aligned AI system and be no closer to satisfying the EU AI Act’s procedural obligations than before, because those obligations concern the agency’s own governance processes, not merely the technical characteristics of the system it uses.8
Bradford’s framework helps sharpen the diagnosis: the Brussels Effect predicts that the AI Act will become the de facto global standard for high-risk AI system architecture. It does not predict that the AI Act’s procedural and documentary requirements — which are not embedded in the AI system itself but in the governance processes surrounding it — will be adopted by actors who have no EU market access incentive to adopt them. For U.S. federal agencies, the interoperability question is precisely in this gap: the technical architecture may converge through vendor markets, but the governance process convergence requires either a political decision to align federal AI governance procedures with EU requirements, or a bilateral regulatory recognition mechanism that does not currently exist, or a finding that the NIST RMF’s existing documentation standards constitute substantial equivalence under the AI Act — a finding that the AI Act’s current framework does not readily support because the two frameworks have not been designed with each other in mind.
The limits that matter for thesis
Bradford’s Part V on the limits of the Brussels Effect is, for this application, the most practically useful section of the 2012 article. The internal constraint analysis predicts where the AI Act’s globalisation will stop — and that prediction has direct consequences for what interoperability can realistically mean.
The external constraint analysis is as applicable to AI as to earlier domains: international institutions are unlikely to effectively discipline EU AI Act requirements; the WTO does not have jurisdiction over most AI governance obligations; other states have the same constrained set of responses that Bradford enumerated — voluntary convergence, negotiated harmonisation, or regulatory obsolescence. The United States government’s opposition to elements of the EU AI Act, expressed through industry submissions during the Act’s drafting and through diplomatic engagement during implementation, has already demonstrated the predictive accuracy of Bradford’s account of what other states’ protests achieve.
The internal constraint analysis is more instructive. Bradford identified two principal internal limits: variation in EU regulatory competence across policy domains, and the political economy of growing internal diversity as the EU expands. Both apply to AI, with additional complications.
The EU AI Act’s scope is broad but not unlimited. Article 2 excludes AI systems used exclusively for military, national security, and defence purposes — a carve-out that reflects both the limits of EU competence (common defence is an intergovernmental domain where the Commission’s regulatory authority is minimal) and the political resistance of member states to EU-level governance of their intelligence and military AI. This carve-out matters enormously for the interoperability question: U.S. federal agencies whose AI use falls primarily within national security and defence functions are, by the Act’s own terms, outside its scope. The interoperability problem is concentrated in the civilian and dual-use government AI use cases that Annex III does cover — procurement systems, benefits administration, immigration processing, law enforcement applications outside the national security carve-out.
The internal diversity constraint also operates in an AI-specific form. The AI Act’s risk-tier architecture was itself a product of internal EU negotiation between member states and EU institutions, and the resulting framework reflects compromises that have left some provisions ambiguous, others contested, and the relationship between the general framework and sectoral AI legislation (in medical devices, financial services, and aviation) partially unresolved. As implementation proceeds through 2025–2027, the practical meaning of the high-risk obligations will be shaped by implementing acts, delegated regulations, harmonised standards developed by CEN-CENELEC and ETSI, and guidance from the European AI Office. The content of “substantial compliance” with the AI Act is not yet fully determinable because the implementing infrastructure that specifies what compliance means in operational terms is itself still being produced.
This is the most important limitation for the interoperability assessment: the Brussels Effect predicts that the AI Act will become a global standard, but the global standard is still being written. The interoperability gap cannot be precisely measured against an incompletely specified obligation. What the analysis predicts is the direction — convergence on EU AI Act standards for high-risk applications — not yet the precise destination.
The coda belongs here, outside any section wrapper.
The analysis yields three findings that are more precise than either the optimistic or the pessimistic accounts of EU AI governance in the literature.
First, the Brussels Effect is structurally likely for the technical architecture of high-risk AI systems. The five-condition framework is satisfied for this domain: EU market power is present, regulatory capacity is being built, the preference for strict rules is genuine and politically durable, the targets are substantially inelastic, and the nondivisibility of AI system architectures makes dual-standard maintenance expensive enough that universal compliance with the EU standard will be the economically rational choice for most global AI providers. The AI Act will become a de facto global standard for high-risk AI architecture in the way GDPR became a de facto global standard for data protection architecture — not through treaty or negotiation but through the rational choices of firms seeking EU market access.
Second, the Brussels Effect does not extend automatically to the procedural and documentary governance obligations that constitute a substantial part of AI Act compliance. These obligations concern the governance processes organisations build around AI systems, not merely the technical characteristics of the systems themselves. They are not embedded in the AI system in the way that REACH requirements are embedded in a chemical’s molecular formula or that GDPR requirements are embedded in a data architecture. They require affirmative institutional choices by actors who may have no EU market access incentive to make them. U.S. federal agencies are the clearest case of actors in this position.
Third, the interoperability question — whether NIST RMF compliance can constitute substantial AI Act compliance — cannot be answered from the Bradford framework alone, because the framework predicts convergence through market mechanism, and that mechanism does not reach actors who are not seeking EU market access. The answer to that question requires either a political decision to align governance frameworks — a bilateral regulatory recognition mechanism — or a technical analysis of whether the RMF’s documentation standards, in practice, produce outputs equivalent to those the AI Act requires. That technical analysis is what makes the publicly-available documentation constraint on U.S. federal agencies’ AI governance records so consequential: if the relevant documentation is classified, no equivalence finding is possible, regardless of how functionally similar the underlying governance practices are.
Bradford’s framework does not solve these problems. It diagnoses them with enough precision to show where the work needs to be done.