Washington Pitched Deregulation at the G20, Brussels Sent 30 Letters

September 2 produced the clearest single-day illustration yet of how far apart the United States and the European Union have moved on AI governance.

In Chapel Hill, North Carolina, the US hosted a G20 innovation ministerial and used it to argue against AI-specific regulation. Michael Kratsios, tech adviser to President Trump, promoted what the administration is calling the Carolina Principles — a framework whose organizing claim is that rules should not single out particular technologies. "Policymakers do not need to approach each innovation in isolation," Kratsios said.

The same day, the European Commission confirmed it had sent information requests to more than 30 AI companies worldwide regarding AI Act compliance, focused on safety and copyright. Henna Virkkunen, the Commission's vice president for tech sovereignty, framed the objective: "Our goal is to ensure that AI in Europe is developed, released and used safely and transparently."

One capital spent the day exporting a deregulatory framework. The other spent it opening files.

What the Carolina Principles are actually arguing

The technology-neutrality argument is not new and it is not stupid. Its serious form goes like this: fraud is already illegal, discrimination in lending is already illegal, defective products already create liability. Writing a separate AI statute risks producing rules that are simultaneously over-inclusive — catching benign systems because they are called AI — and under-inclusive, missing harms that arrive through a technology the drafters did not anticipate. Regulate outcomes, not implementations.

There is real evidence behind this. Technology-specific regulation ages badly, and the EU's own AI Act has already been criticized for definitions written before the current generation of models existed.

The less serious form of the argument is that any specific rule is by definition premature, which is a position that never expires and therefore never has to be defended.

The administration brought industry backing to make the case. Mark Zuckerberg raised infrastructure demands. Elon Musk made the energy argument — "There will be a significant power shortfall next year, not [the] distant future" — and the regulatory one: innovation requires that "new things must be default legal as opposed to default illegal."

That last formulation is the actual crux, and it is more honest than most of what gets said in these rooms. Default-legal versus default-illegal is the whole disagreement, compressed. The EU chose default-illegal for a defined set of high-risk applications. The US is arguing for default-legal with after-the-fact liability.

The venue choice is the strategy

Hosting this at a G20 ministerial rather than announcing it domestically is the part worth reading carefully.

Domestic US AI policy is largely settled for the moment — a deregulatory posture with a voluntary federal review layer on top. What is not settled is what the rest of the world does, and specifically whether the EU's approach becomes the global template the way GDPR did.

The Brussels Effect is the thing the Carolina Principles are built to prevent. GDPR became the de facto world standard not because other countries adopted it but because multinationals found it cheaper to build one compliant product than several. If the AI Act works the same way, American labs end up governed by European rules regardless of what Washington does — which makes the deregulatory posture domestically hollow.

So the US is not primarily lobbying its own regulators. It is lobbying everyone else's, early, before national AI frameworks harden. Getting a G20 ministerial to entertain technology-neutral principles is worth considerably more than another domestic executive order.

The EU understands this too. Virkkunen's counterpart signal is unsubtle: enforcement began on August 2, and a month later there are 30+ live information requests. Brussels is demonstrating that the AI Act is not aspirational, at the exact moment Washington is telling the world it does not need one.

The voluntary layer, and what Altman confirmed

The domestic US position is not pure laissez-faire, and the nuance matters.

The administration operates a voluntary framework for reviewing some advanced models before release, which can include giving government officials access for up to 30 days — the structure established by the June executive order that followed Anthropic's zero-day-finding Mythos. Sam Altman confirmed this week that the administration reviewed OpenAI's Astra model, and emphasized that the process is voluntary.

Astra is a useful test of what "voluntary" is worth. It is the first model OpenAI has classified Critical for cybersecurity under its own Preparedness Framework — 100% on ExploitBench, two zero-days found and chained during testing. It went through the review. It shipped.

So the mechanism functioned as designed: the government looked, and the decision remained the company's. Whether that reads as appropriate restraint or as a review with no teeth depends entirely on what one thinks the review was for. It is, notably, the strongest available example for both sides.

What this actually means for anyone shipping

The practical consequence of divergence is not that one regime wins. It is that everyone building at scale now maintains two postures.

If you ship into Europe, the transparency obligations are live: interactive AI systems must disclose they are not human, deepfakes must be labelled, and 30+ companies just learned that the Commission asks questions in writing. Safety and copyright documentation is the immediate exposure.

If you ship only in the US, the binding constraint is not a statute. It is liability, state-level rules, procurement requirements, and the voluntary review if you are at the frontier.

If you ship both — which is every company that matters — you build to the stricter standard and argue about it in Brussels, which is precisely the dynamic the Carolina Principles exist to break.

What to watch

Whether any G20 member actually adopts the Carolina Principles in a national framework. Principles endorsed at a ministerial and principles legislated are very different artifacts.

What the Commission does with 30 sets of answers. Information requests are the step before enforcement action. The first AI Act penalty against a major lab is the event that tells you whether this regime has weight.

Whether the voluntary US review ever results in a model not shipping. Astra shipped. Until something does not, the review is a disclosure process with a 30-day window.