US pushes G20 toward light-touch AI rules, claims China has signed on

The US government has asked G20 members to refrain from drafting new rules for the sector, urging instead that existing law be applied to whatever novel situations artificial intelligence creates. The pitch came from Michael Kratsios, the White House's senior technology adviser, at a two-day G20 Innovation Ministers meeting in Chapel Hill, North Carolina, on Tuesday, and is now being marketed abroad as the Carolina Principles.

Kratsios co-chaired the gathering alongside counterparts from other G20 members and told reporters that countries signing the principles had agreed to that approach. He added that China had also signed. He did not produce a copy of the document. China's embassy in Washington did not respond to a request for comment.

What the Carolina Principles actually propose

The substance of the US proposal is restraint. Governments should reach for tools they already have: consumer protection law, competition enforcement, product liability, data protection, sectoral regulators. New rules should be reserved, the argument runs, for situations that cannot be handled with what is already on the books.

That posture sits in deliberate contrast to the European Union, which spent years negotiating its AI Act and now sorts AI uses into four risk tiers, with the heaviest obligations reserved for high-risk applications such as biometric identification, recruitment software and components in safety-critical machinery. The American position, if adopted by other capitals, would mean most AI products on the market would not be considered novel enough to warrant fresh legislation. That includes the generative systems now being integrated into search, customer service and software development.

Kratsios told the meeting that policymakers should not treat every emerging technology as a first-of-its-kind policy problem. The framing reflects a wider Trump-era belief that regulation slows American AI companies and cedes ground to Beijing, a position that has shaped Washington's posture in trade talks as well.

Industry takes the stage in Chapel Hill

Ministers were not the only ones with a platform. Three of the most prominent chief executives in American technology used the meeting to make the case for lighter rules.

Demis Hassabis, chief executive of Google DeepMind, called on governments to agree on standardised safety tests for advanced AI systems. The proposal has gathered momentum since 2023, when the United Kingdom launched the AI Safety Institute at Bletchley Park and counterparts opened in Washington and Tokyo. The idea is to give regulators and the public a way to judge whether a frontier model can, for example, help a non-expert build a bioweapon.

Mark Zuckerberg, chief executive of Meta, used his slot to argue against restrictions on open-weight AI models, software whose trained parameters are released publicly so others can build on them. His intervention follows Meta's recent drive to position itself as the standard-bearer for the open approach, through its Llama family of models. That stance distinguishes Meta from OpenAI and Google, which have kept their most capable systems under tighter control.

Musk turns his fire on Brussels

Elon Musk, chief executive of SpaceX, was less diplomatic. He attacked European Union technology regulation and told the meeting that the bloc's policies held back progress for companies that operate there. Musk has clashed with Brussels repeatedly, particularly over content moderation under the Digital Services Act and the bloc's data protection rules. His intervention at Chapel Hill aligns with the Trump administration's own complaints that European digital rules function as an export barrier for American technology.

The Tesla chief made a separate pitch on energy, urging countries other than China to develop new power sources for data centres. AI infrastructure is among the most energy-intensive industrial activities currently being built; Musk's call reflects a concern, voiced across the industry, that grid constraints could throttle how fast AI capacity can be added in Europe and North America alike.

The European counter-model

The EU is pursuing the opposite course. The AI Act, agreed by the European Parliament and Council in 2024, classifies AI uses by risk and imposes obligations on providers of general-purpose models above defined compute thresholds. The first set of prohibitions took effect in February 2025, with the bulk of the obligations phasing in across 2026 and 2027. The Commission has framed the law as a way to anchor the EU's regulatory model around the world, a phenomenon observers have called the Brussels effect after the GDPR's earlier success in setting global data protection norms.

Whether that effect still operates as it once did is a separate question. Several large American companies have argued the AI Act is too prescriptive and creates compliance costs that European rivals cannot match. Washington has complained, in particular, that the EU's treatment of providers, bringing them within scope once their models are placed on the EU market regardless of where the developer sits, amounts to extraterritorial regulation.

Beijing's competing vision

China has spent the past three years promoting its own AI regulatory templates, including the Interim Measures for the Management of Generative AI Services that took effect in 2023, while exporting open-weight models through developers such as DeepSeek. Beijing's official line at international fora has been that AI governance requires multilateral cooperation, with rules written through bodies such as the United Nations, and that safety obligations should be tightened.

Kratsios's claim that China had signed the Carolina Principles is therefore striking. The American framework calls for restraint, the Chinese public position calls for binding safety rules. No text was shown to reporters in Chapel Hill, and the Chinese embassy in Washington did not respond. That leaves the claim in limbo, an unusual moment of public divergence between the US technology adviser's account and the messaging from Beijing's diplomats.

Three blocs competing for the rulebook

The episode exposes how three distinct approaches to AI governance are now being marketed abroad. Washington wants existing rules applied and minimal new constraints. Beijing, at least in public, calls for multilateral governance and stronger safety standards while promoting its own open-weight ecosystem. Brussels is exporting its risk-based model and pushing the Council of Europe to agree a separate convention on AI.

The contest is not only about safety or innovation. The country with the largest installed base of AI systems, the argument goes in Washington, will shape the standards others adopt. Reports last month suggested the US administration was considering asking partner governments to choose between aligning with American or Chinese AI ecosystems, a step that would formalise a technological bloc split. Chapel Hill suggests the US would rather avoid that choice by recruiting other capitals, including Beijing's, behind a light-touch declaration first.