The framework before the test
Bradford identifies conditions for regulatory influence through markets. Applying them to AI requires evidence about particular firms and systems. This reading uses the framework to organise that inquiry; it does not calculate a probability of global convergence.
Bradford identifies five conditions: market power, regulatory capacity, a preference for strict rules, regulation of inelastic targets and nondivisibility of production or conduct. The last two concern whether the regulated activity can move and whether firms can maintain different standards across markets. Each requires evidence in the domain being examined.1
On page 17, Bradford separates the ability to impose conditions on market access from the decision to extend a standard worldwide. A provider might adapt its EU offering while maintaining another configuration elsewhere. The analysis must therefore examine whether separation is feasible and worthwhile. It cannot infer global adoption from the existence of EU obligations alone.
The original Article 2 distinguishes providers placing systems or GPAI models on the Union market from third-country system providers and deployers whose output is used there. The Act includes exclusions. These legal connections help define the inquiry; they do not establish that a provider applies the same practices worldwide.
Market access and the capacity to enforce
A large market creates an incentive, but its value differs across providers. A legal power to investigate creates another incentive. Neither tells us, by itself, how a particular company will respond.
On page 12, Bradford links the value of market access to adjustment costs and alternative markets. That comparison gives this reading a practical starting point. For a hypothetical hiring-software provider, the evidence would include European revenue, compliance costs and opportunities elsewhere. A population count cannot replace those figures.
The same inquiry should distinguish selling a model from operating an application. A provider’s position in the supply chain affects which commercial relationships must be examined. This reading has not measured those relationships. It therefore does not establish that leaving the EU market would be prohibitive for most AI providers.
Bradford also distinguishes market size from regulatory capacity. Expertise and enforcement resources matter. An assessment of the AI Act would need dated evidence of staffing, investigations and enforcement activity. The existence of an institution or a statutory power is relevant, but it does not demonstrate effective enforcement.
Article 101 illustrates the distinction. In the original 2024 Act, the Commission’s ceiling for specified GPAI-provider infringements is 3% of the preceding financial year’s worldwide turnover or EUR 15 million, whichever is higher. That provision is not limited to systemic-risk models. A penalty ceiling does not show how frequently it is used. See Article 101.
For this thesis, those observations set a limit on the explanation. If a vendor adopts a shared documentation practice, market pressure is one possible reason to investigate. The adoption alone would not establish its cause. A procurement requirement, an internal policy or another applicable rule could also matter.
Strictness and the target of regulation
Which requirement is stricter, for whom, and at what date? That comparison must identify the activity and the responsible actor. A general contrast between Europe and the United States cannot settle it.
A voluntary framework and a binding obligation have different legal roles. But the NIST AI RMF does not represent every rule governing a federal agency. Section 05 therefore considers OMB M-25-21 separately. The relevant comparison must specify the agency, system, use and applicable instruments before drawing a conclusion about relative demands.
A hypothetical hiring system makes the problem concrete. One document might describe how applicants receive information; another might assign human review responsibilities. The comparison should examine each practice and its legal context. It should not classify the whole U.S. system as permissive because one framework is voluntary.
Bradford’s inelastic-target condition asks whether the regulated activity can escape the jurisdiction. Her discussion contrasts consumer markets with more mobile capital. Applying that distinction to AI requires separating server location from the legally relevant activity. Moving computation does not, by itself, establish that the provider or deployer falls outside the Act.
The original Article 2 provides distinct connections through market placement and use of a system’s output in the Union. Its exclusions still matter. Territorial scope is also separate from high-risk classification. The original Article 6 distinguishes the product route involving Annex I from Annex III uses, with specified qualifications. A broad label such as healthcare is insufficient to complete either analysis.2
For the hypothetical hiring service, the inquiry should identify its intended purpose, users, outputs and deployment arrangements. It should then explain the legal basis for any classification. Only after that can the analysis ask how market access affects the provider’s choices. Jurisdiction, incentives and worldwide adoption remain distinct steps in the argument.
Nondivisibility — a hypothesis to test
Does a provider need one design for all markets? The answer depends on the system. Expensive regional adaptations would support an economic argument; they would not establish technical impossibility.
Bradford distinguishes legal, technical and economic nondivisibility on pages 17–18 of the 2012 article. Her examples concern mergers, data practices and chemical production. Applying those distinctions to AI requires new evidence. The article does not establish the cost of maintaining regional versions of an AI system.
Legal nondivisibility concerns whether an activity can be separated across jurisdictions. Bradford uses a global merger as an example. For an AI system, this reading would need to identify a particular legal constraint that prevents separation. The existence of a common model does not establish such a constraint.
Technical nondivisibility concerns whether separation is technically feasible. Consider a hypothetical hiring service with a shared model and regional interfaces. The inquiry would examine which records, oversight functions and processing steps can be configured separately. Shared infrastructure alone does not answer that question. A model diagram, configuration records and deployment evidence would help test it.
Economic nondivisibility concerns the relative cost of alternatives that are feasible. Maintaining regional versions might cost more than using one design. Uniform practices might also impose costs outside the EU. Neither balance is measured here. Relevant evidence would include maintenance costs, regional revenue and records explaining the provider’s design choice.
These are research questions, not ratings of the AI industry. This reading has not established that dual versions are prohibitively expensive for most providers. It also has not established that the AI Act produces worldwide uniformity. A finding of regional separation would qualify the market-diffusion hypothesis rather than invalidate the comparison of documentary evidence.
What the comparison can establish
Market influence, documentary coverage and legal compliance are different questions. A vendor may reuse an EU-facing practice elsewhere without making every customer compliant with EU law.
The unit of analysis is the AI system. The agency provides its operating context. The inquiry should identify the system, its intended use, the actors involved and the available records before mapping evidence against Articles 11–14. A shared application label does not itself establish that the AI Act applies.
The NIST AI RMF 1.0 is voluntary. That does not make every federal agency obligation voluntary. The OMB memorandum M-25-21 of 3 April 2025 directs covered agencies to implement specified practices, including minimum risk management for high-impact uses. Its scope and exceptions must be read separately. It is not a general statute regulating all private AI providers.
An assessment record may support more than one governance purpose. Its existence still leaves questions about content, timing, completeness and implementation. The analysis must distinguish what a document reports from what independently establishes that the practice occurred.
The resulting model can show supported correspondences and gaps. It cannot substitute for a legal determination of applicability, an assessment of actual compliance or a decision formally recognising equivalence. Those conclusions require their own authority and evidence.
Limits that shape the thesis
The comparison needs a defined legal baseline and an explicit account of missing evidence. Neither market influence nor a regulatory label removes those requirements.
The original Article 2 contains exclusions for specified military, defence and national-security uses. That does not exclude every system operated by an agency with such a mission. The purpose and circumstances of the particular use matter. Article 2(4) separately addresses certain international law-enforcement and judicial-cooperation arrangements, with conditions concerning safeguards.
The thesis can use Articles 11–14 as a comparative reference without claiming that every U.S. system is legally subject to them. It must state when it is testing documentary coverage against that reference and when it is making an applicability claim. Those tasks should not share an unqualified label of compliance.
Missing public records constrain this study. A record may be unavailable for reasons that require further investigation; absence from the retrieved material is not proof of noncompliance. Record the gap, its effect on the comparison and what additional evidence would resolve it.
Legal versions also matter. The original 2024 text, later amendments and application dates must be distinguished. This review does not establish current consolidated law. Before applying the comparison, identify the version governing the question and the date of the evidence.
Bradford’s framework can guide questions about the incentives for suppliers to reuse regulatory practices across markets. Whether that occurs for a particular system requires evidence. The thesis should therefore record supported correspondences, competing interpretations and unresolved gaps, without turning a market hypothesis into a finding of legal equivalence.