Brussels was already probing 30-plus AI companies when the G-20 agreed to hold new rules back

The document governs future rulemaking, which is the reason it costs its signatories so little. It asks governments to reserve new AI regulation for novel considerations ^[2], and, on the account of implicator.ai, that phrase comes from a description of the draft, with no statement of what counts as novel ^[3]. The document covers pro-innovation policy, opportunity and prosperity, technical workforce development, intellectual property for AI, standards and industrial investment in supply chains ^[8], and stops there; enforcement and liability sit outside all six pillars ^[9]. It has no enforcement mechanism to suspend.

Which leaves the enforcement that exists. The Commission's requests went to more than 30 companies worldwide, concentrate on safety and copyright compliance, and are preliminary rather than formal, though they can open into proceedings under the AI Act ^[6]. Confirmation came on the Tuesday the Chapel Hill meeting opened ^[5], one day before ministers adopted the framework ^[7]. The sequencing carries the point: the inquiry track did not wait for the declaratory one.

A communique that 20 governments were willing to sign, China among them ^[4], asks nothing of anyone in the near term. That is precisely what makes it durable for capitals that have not yet written AI statutes, because restraint is now a position they have endorsed in writing. The effect lands on the next decade of national lawmaking rather than on the next compliance review.

G-20 aligned on restraint, and regulatory risk can look lower for it. But that reading skips the second half of Washington's own approach, which pairs Kratsios's argument against treating each emerging technology as a first-of-a-kind policy problem ^[13] with voluntary prerelease testing, government review of the most powerful models, and a brief June 2026 shutdown of two Anthropic models over security concerns ^[12]. That is state contact routed through discretion rather than statute. Discretion is harder to plan against than a published obligation.

The industry testimony in the room did not run one way either. Jensen Huang told ministers to regulate actual and pragmatic harm rather than hypothetical harm ^[14], and said advances in model training were already making systems safer than rules proposed by the Biden administration and others ^[15]. Sam Altman, who called adoption non-negotiable, also told them worry is warranted and that global standards and oversight retained a place ^[16]. Anyone invoking industry consensus is choosing between two available quotes.

For planning purposes the useful line came from Virkkunen, who said European law addresses conduct in advance while US courts often examine the same companies and practices after disputes arise ^[10]. Under a regime that acts in advance, the safety and copyright documentation that answers a request has to exist before the request lands. That is this quarter's work, and it settles next quarter's question, which is whether those preliminary inquiries stay preliminary. One caution about the record: this rests on a single outlet's report, published as an AI-generated summary reviewed by an editor ^[18].