A Pact With No Teeth, and That's the Point
At the G20 Innovation Ministerial held September 1-2, 2026 at the Carolina Inn in Chapel Hill, North Carolina — officially co-hosted by the U.S. Department of Commerce and the White House Office of Science and Technology Policy (trade.gov media advisory) — the United States asked fellow G20 economies to sign the "Carolina Principles." The pact commits signatories to create no new AI-specific regulators, avoid model-by-model rulemaking, and "reserve new regulation for novel considerations" rather than treating each new AI capability as a first-of-its-kind policy problem. OSTP Director Michael Kratsios, who led the drafting, put it bluntly: "Policymakers do not need to approach each innovation in isolation and should not treat every emerging technology as a first-of-a-kind policy problem" (Gizmodo). Commerce Secretary Howard Lutnick co-hosted; Sam Altman and Jensen Huang appeared in person to make industry's case directly to ministers.
The document is non-binding, which is both its main limitation and its main appeal. China's technology and science minister signed on, and Kratsios told reporters the bilateral meeting with Beijing was "great" (The Globe and Mail) — a notable irony given how much of Washington's AI policy is framed around competing with China. The UK, also a G20 member, had not committed either way as of the ministerial's close.
The Case for Restraint, Stated Fairly
The strongest argument for the Carolina Principles isn't industry lobbying — it's regulatory history. Technology-specific regulators built around a single snapshot of a fast-moving field tend to calcify. A bespoke "AI regulator" created in 2026 risks writing rules for large language models just as the frontier moves to agentic systems, robotics, or something not yet named, while lacking the flexibility that general-purpose consumer-protection, competition, and liability law already has. There's also a coordination case: a patchwork of 20-plus national AI regulators, each with its own model-classification scheme and audit regime, imposes real compliance costs on smaller developers who can't staff a compliance team for every jurisdiction, while incumbents with the scale to absorb that cost pull further ahead. Kratsios's framing — invest in foundational research, use existing law where it already applies, save new rulemaking for genuinely novel harms — is a coherent, evidence-respecting position, not merely deregulatory reflex.
Where the Principles Understate the Trade-off
But "reserve regulation for novel considerations" begs the central question: who decides what counts as novel, and on what timeline? The EU's own experience this year is instructive. The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on July 27, 2026, after amending the AI Act to push back compliance deadlines for high-risk AI systems — standalone systems to December 2, 2027, and systems embedded in regulated products to August 2, 2028 (EUR-Lex; lawandtechnology.eu). That delay was driven not by philosophical disagreement but by practical failure: EU member states hadn't finished designating the national competent authorities or finalizing the harmonized technical standards the Act's own high-risk tier depends on. In other words, even a jurisdiction fully committed to prescriptive AI regulation is discovering that model-by-model, sector-by-sector rulemaking is genuinely hard to implement on schedule — which cuts both ways. It's evidence for the U.S. argument that novel, narrow rules beat sweeping upfront classification schemes. But it's also evidence that "wait until harm is novel and specific" can mean waiting until the harm has already scaled, since transparency obligations under the same Act only took effect this past August, well after many general-purpose models had been on the market for years.
EU Commission Vice President Henna Virkkunen's response to the U.S. push was pointed: "Our goal is to ensure that AI in Europe is developed, released and used safely and transparently" (Al Jazeera) — a fair statement of the opposing case that the Carolina Principles don't really answer. A framework that asks governments to promise, in advance, not to create new oversight bodies is making a bet about how AI risk will unfold that nobody, including the drafters, can actually verify yet.
The Realistic Read
None of this makes the Carolina Principles meaningless. As a statement of first-mover intent — the country that hosts the leading AI labs telling the rest of the G20 it won't tolerate a race to license-and-audit regimes before harms are demonstrated — it will shape how smaller and developing economies calibrate their own AI laws over the next two years, especially ones weighing whether to import the EU's risk-tiered model wholesale. But a non-binding pact signed alongside China while the UK sits out, deployed on the same day the European Commission sent information requests to more than 30 AI companies over safety and copyright compliance, is a marker in an ongoing negotiation, not a settled global consensus. The right regulatory question — for the U.S., EU, or anyone else — isn't whether to regulate AI, but whether a given rule targets a demonstrated harm with a proportionate, enforceable remedy. The Carolina Principles get the diagnosis right without yet answering who makes that call in practice.