The United States is using its G20 presidency to press other major economies towards a lighter model of artificial-intelligence regulation, challenging the more prescriptive European approach as substantial parts of the EU AI Act enter operation.

The White House Office of Science and Technology Policy is promoting a proposed framework known as the Carolina Principles at a G20 technology meeting in Chapel Hill, North Carolina. Michael Kratsios, the US science and technology adviser, has urged governments to rely on existing sector regulators where possible rather than creating additional AI-specific authorities and rules.

Reporting on the meeting says the principles encourage governments to reserve new regulation for genuinely novel considerations, while supporting foundational research and cooperation between government and industry on technology testing. A complete public copy of the Carolina Principles had not been located during Techopia’s sourcing process, so the framework cannot yet be assessed beyond the policy direction described publicly and in reporting from the meeting.

The proposal sharpens an argument over whether AI requires dedicated regulatory architecture or can largely be governed through laws already covering financial services, employment, healthcare, consumer protection, competition, product safety, and cybersecurity. Washington is leaning towards the latter, whereas Europe has spent several years constructing a horizontal rulebook that varies obligations according to a system’s characteristics and risk.

The difference is becoming commercially concrete as EU requirements move from legislation towards implementation. Providers and deployers increasingly have to understand how transparency, documentation, model governance, and later high-risk obligations fit into products that may also be sold under a different regulatory approach elsewhere.

Europe has regulated, then adjusted the timetable

The European position is more complicated than a binary choice between regulation and innovation. Brussels has already adjusted part of the original AI Act implementation timetable as businesses, standards bodies, and governments raised concerns about whether supporting standards and guidance would be ready soon enough.

Those changes leave Europe with a legal structure markedly different from the approach Washington is promoting. Providers of general-purpose models face defined transparency and copyright obligations, while particularly capable models can trigger additional safety and security requirements. National authorities and the European AI Office also carry enforcement responsibilities beyond voluntary cooperation between industry and government.

The US approach reflects a concern that AI-specific rules can become outdated before regulators finish implementing them. Models and applications are changing through new agent capabilities, multimodal systems, cheaper inference, and more extensive tool use, while a law designed around one generation of products may have to govern substantially different systems several years later.

Existing sector rules offer one answer because they leave responsibility with regulators that already understand the underlying activity. An automated lending decision still raises banking, credit, and discrimination questions, for example, while an AI diagnostic product remains connected to medical-device regulation irrespective of the underlying model architecture.

Relying mainly on established law creates gaps of its own, however, where general-purpose systems cross several sectors or introduce risks that do not fit neatly inside one regulator’s remit. A software agent can interact with communications, payments, data, customer records, and business applications during a single task, making sector boundaries less tidy than the policy model suggests.

Divergence becomes an operating-model problem

For technology suppliers operating on both sides of the Atlantic, competing regulatory models translate into product and compliance work. A developer may face relatively light AI-specific federal requirements in the US while maintaining separate documentation, transparency controls, risk processes, and technical assessments for European deployment.

Multinational customers face a similar challenge when rolling out the same system across several jurisdictions. A governance framework developed at a US headquarters may need additional controls in Europe, while regulated industries can encounter national obligations layered on top of wider AI rules.

That fragmentation raises cost and can influence where products launch first, which features are made available, and how quickly suppliers expand into new markets. Regulatory requirements can delay deployment when compliance is uncertain, although predictable rules can also reduce ambiguity once companies understand the conditions attached to market access.

The G20 is unlikely to produce uniform AI regulation even if governments agree on broad principles. Its members have different legal systems, industrial interests, political structures, and attitudes towards privacy, competition, national security, and the role of the state in technology markets.

Rules are becoming part of technology competition

The argument also sits inside the wider industrial contest surrounding AI. The United States hosts many of the businesses developing frontier models and advanced accelerators, China has built a substantial domestic model and infrastructure ecosystem, while Europe has fewer global-scale foundation-model providers and has placed more emphasis on regulation, industrial support, and strategic autonomy.

Regulatory design can influence competition as well as safety. A complicated compliance regime can create barriers for smaller suppliers, while weak oversight can transfer the cost of failures towards customers, workers, or citizens. Governments are trying to avoid both outcomes while simultaneously competing for data centres, model developers, capital, and technical talent.

The technical transition towards agents makes the disagreement harder because software can now use tools and take actions rather than simply produce text. That development challenges policies built either around traditional sector boundaries or around static assessments of a model, since risk increasingly emerges from what the wider system is permitted to do.

Europe’s model faces the same pressure. The AI Act was designed to be risk-based and relatively technology-neutral, but regulators still have to interpret its requirements as applications change, particularly where general-purpose models are connected to autonomous tools or high-risk workflows.

Washington’s G20 push therefore adds another regulatory model competing for international influence rather than resolving the governance question. European statutory obligations are already moving into practice, while the US is arguing that technological novelty should not automatically produce new regulators.

The two approaches are unlikely to converge quickly, leaving regulatory architecture as another design constraint for companies deploying AI internationally. Where systems launch, which controls are built into them, and how much evidence providers retain may increasingly depend as much on jurisdiction as on the underlying model.