The G20's Carolina Principles Quietly Abandon the Idea of One Global AI Rulebook
G20 governments met in Chapel Hill this month and endorsed a set of AI policy principles backed by Washington. They’re non-binding. They favour rules written for particular uses over rules written for the technology itself, and they lean on cooperation between governments and industry. The name attached to them is the Carolina Principles.
What they don’t do is create a universal regulatory regime, and that omission is the substance of the agreement.
For about three years the assumption in international AI policy circles was that something like a general framework would eventually arrive: a floor of obligations covering frontier models, agreed among the countries that host the labs, with the EU’s approach as the likely template. That assumption is now dead at the G20 level. The Carolina Principles put sector-specific regulation in its place. Medical AI gets medical rules from medical regulators. Credit AI gets credit rules. A model itself acquires no obligations by virtue of being large.
The American argument for this is competitive, and officials have made it plainly. Restrictive general-purpose rules slow Western development while Chinese models and platforms pick up users in markets that never had a rulebook to begin with. Regulation of capability, on this view, is a self-imposed handicap in a race where distribution decides the winner.
The counterargument is that use-based regulation only works when the uses are known in advance, which is precisely what general-purpose systems make impossible. A model that isn’t a medical device on Tuesday can be one by Thursday, at which point the sectoral regulator is the last party to find out. The gap between a capability existing and a sector noticing it is where the risk sits, and nothing in the Chapel Hill text addresses it.
Consider who was in the room. Huang, Altman, Zuckerberg and Musk all attended a governmental technology gathering that produced a voluntary framework endorsing closer government-industry cooperation. That’s not a scandal, and there’s no need to pretend it is. It does describe the balance of information. Governments negotiating rules for systems they can’t independently evaluate will produce rules shaped by the people who can evaluate them, whatever the seating chart says.
The practical effect over the next year is fragmentation with a friendly face. The EU keeps its framework and continues to enforce it, which is why a chatbot was classified as a search engine there last week. The US pushes sectoral rules and preemption of state-level activity. China regulates in its own direction for its own reasons. Middle-income and mid-sized states get the Carolina Principles as a template, which is cheaper than writing their own and comes with implicit access to American models.
That last point is the one to watch. Non-binding principles are how standards spread when nobody can compel anyone. A country adopting them isn’t giving up sovereignty; it’s picking a stack. The rulebook and the vendor arrive together.
Which makes the interesting question not whether the Carolina Principles are enforceable. They aren’t, and they were never meant to be. It’s how many states adopt them, how fast, and what else comes through the door behind them.