Analysis
The G20 innovation ministerial ran Sept. 1 and 2, co-hosted by the Commerce Department and the White House Office of Science and Technology Policy, with OSTP director Michael Kratsios leading alongside Commerce Secretary Howard Lutnick. The US came to press a hands-off approach to AI regulation, and did not get a clean result, Axios reported.
Kratsios defended the Carolina Principles, under which governments commit to avoid creating too many rules governing AI deployment. That is the administration's core export: rather than negotiate a binding international instrument, get peer governments to pre-commit against one. It is the mirror image of the approach the EU took with the AI Act.
The disagreement came from inside the tent. Demis Hassabis, whose DeepMind sits under Alphabet and is among the labs the principles are meant to protect, called for the United States to create an organization to test powerful AI systems before release. David Sacks, Trump's former AI czar, said he opposed the creation of any such bodies. Elon Musk, appearing virtually on the first day alongside Sacks and Meta's Dina Powell McCormick, defended AI data centers and attacked EU rules.
“Kratsios defended the Carolina Principles, under which governments commit to avoid creating too many rules governing AI deployment.”
Hassabis's position is not abstract. The UK AI Safety Institute and its US counterpart have both run pre-deployment evaluations, and the past week produced two demonstrations of why that matters: OpenAI delayed a model after determining it crossed the company's own critical cyber threshold, and Anthropic paused a training run after Claude took unauthorized actions during evaluations. Both were self-reported by the labs. Hassabis is arguing that self-reporting is not a durable regime.
Sacks's counter is the standard accelerationist case: a pre-release testing body becomes a licensing body, licensing favors incumbents, and incumbents are the labs already large enough to staff a compliance function. That argument has more force in a market where Anthropic, OpenAI and Google collectively hold most of the frontier capability than its proponents usually admit -- a gatekeeper regime would be written largely by the three companies it governs.
What did not happen at the ministerial is any binding commitment. The Carolina Principles are a statement of intent among governments that mostly already regulate AI differently from one another, and the EU is not going to unwind the AI Act because a summit in North Carolina asked it to. For US founders the practical consequence is unchanged: build for the EU regime if you sell in Europe, and expect the domestic environment to remain permissive at the federal level while getting steadily harder at the county level, where the data centers actually go.
