US AI Executive Order: Voluntary Rules Split From the EU AI Act

What changed

• No US model licence. The 2 June 2026 executive order, titled "Promoting Advanced Artificial Intelligence Innovation and Security", explicitly prohibits any mandatory federal licensing, pre-clearance or permitting requirement to develop or release an AI model.
• A voluntary frontier framework instead. Trusted labs can give the government secure early access to the most capable models before public release. The stated goal is stronger cybersecurity, not a regulatory gate.
• The EU is going the other way. The EU AI Act is a binding regime with enforceable obligations and real financial penalties, phasing in across 2025 to 2027.
• The divergence is the story. If you sell into both the US and the EU, you now manage two different regimes — and the US side carries its own federal-versus-state uncertainty.

What the order actually says

On 2 June 2026, President Trump signed an executive order titled "Promoting Advanced Artificial Intelligence Innovation and Security". The headline for builders is what it refuses to do. The order states that nothing in it "shall be construed to authorize creation of any mandatory governmental licensing, pre-clearance, or permitting requirement for the development, publication, release or distribution of AI models." In plain terms: there is no federal permission slip to train, publish or ship a model in the United States.

In place of a licence, the order establishes a voluntary framework aimed at the most capable systems — the frontier models from labs such as Anthropic, OpenAI and Google. Under it, trusted partners can give the government secure early access to new models before public release. The point, as the order frames it, is to strengthen cybersecurity rather than to create a regulatory gate that a model must pass through before it reaches users. The law firm Crowell & Moring characterised the result as a "voluntary regulatory regime of frontier AI models" — cooperation by invitation, not compliance by statute. The order also expands a set of cybersecurity and federal oversight provisions, but it stops well short of the EU's structure of enforceable duties.

Several US legal teams have published analyses worth reading if you sell into the American market: Greenberg Traurig, Perkins Coie, Holland & Knight, Crowell & Moring and Lathrop GPM have all weighed in. The common thread across their notes is that the federal posture is deliberately innovation-friendly and deliberately light on hard obligations.

The federal-versus-state fight underneath it

The 2 June order does not stand alone. It builds on an earlier executive order from 11 December 2025 on federal preemption of state AI laws, which proposed to preempt state rules that are inconsistent with the federal approach. A related legislative push — referred to as the "Great American AI Act" — aims to freeze state AI laws so that a single national posture prevails over a patchwork of local ones.

That is exactly where the friction lies. States such as California and Colorado, along with several state attorneys general, have signalled opposition and likely legal pushback. The result is that the "light-touch" US picture is not actually settled. A builder reading the federal order as a green light has to also reckon with the possibility that a state regime survives, or that the preemption effort is contested in court for years. For a product team, unresolved litigation is not a footnote — it is itself a planning risk, because the rules you build to today may shift under you.

The contrast: a binding EU regime with penalties

The EU AI Act sits at the opposite end of the spectrum. Where the US framework relies on cooperation, the EU relies on law: it is a binding regime with enforceable obligations and genuine financial penalties for non-compliance, with general-purpose AI and transparency duties phasing in across 2025 to 2027. We keep the precise clause numbers and dates in our dedicated EU coverage rather than restating them here, because the timeline has been amended more than once — see our pieces on the Digital Omnibus and the high-risk deadline and the delay of high-risk rules into 2027 for the current, specific position.

The structural difference is what matters for planning. In the US, a frontier lab can choose to share early access and choose not to. In the EU, certain obligations attach by law and carry penalties if you ignore them. That is the divergence in a sentence: one regime is opt-in and cooperative, the other is mandatory and enforceable. A model that is lightly handled for a US release may still trip EU duties the moment it is offered to European users.

US voluntary order vs EU AI Act vs UK approach

The EU column is kept high-level by design; for exact deadlines and penalty figures, follow our linked EU AI Act coverage, which we update as the timeline shifts.

Where the UK sits — between the two

The UK is neither the US nor the EU. It has pursued its own frontier-AI regulatory approach, layered across existing regulators — the Bank of England, the Financial Conduct Authority and the Information Commissioner's Office each touching AI in their patch rather than one single AI statute. That makes the UK a third map to read, not a copy of either neighbour. If you sell into Britain, you must track all three regimes: the voluntary US posture, the binding EU regime, and the UK's sector-led model. Our explainer on UK frontier-AI regulation under the BoE, FCA and ICO sets out who governs what.

For a UK builder, the practical consequence is that "where is my user?" determines "which rulebook applies?" A London team shipping a hiring tool to EU employers inherits EU obligations; the same team shipping to a US customer inherits the lighter US posture plus whatever state law survives; and a domestic financial-services use case answers to the FCA and ICO. One product can therefore live under three regimes at once depending on the customer.

The India angle

Indian builders are increasingly export-first. A Bengaluru or Pune team building agents, RAG systems or fine-tuned models often sells into the US and the EU long before the domestic market matures. At home, the Digital Personal Data Protection (DPDP) framework and the IndiaAI programme set the baseline for data handling and sovereign compute. But the binding constraint on an export product is rarely the domestic floor — it is the strictest foreign market the product touches.

The rule of thumb is simple. Map your product to the strictest market you sell into. If a single EU customer can put you inside the EU AI Act's scope, the economical move is to build to that bar from the start, then treat the US and the domestic Indian baseline as subsets you already satisfy. Designing to the US voluntary posture first, then bolting on EU-grade evidence later when a European deal appears, is the expensive path — and it tends to arrive at the worst possible moment, mid-deal.

The light-touch label has an asterisk

It would be a mistake to read "innovation-friendly executive order" as "the US does not restrict AI". The same administration that is light-touch on domestic model regulation has used export controls on specific frontier models — a model can be freely releasable for domestic users yet restricted from leaving the country. We have covered one such case in depth; rather than re-report it here, see how export controls forced Anthropic to pull Fable 5 and Mythos 5. The lesson for builders is that "US regulation" is not one dial. Domestic innovation policy and export policy are separate levers, and they can point in opposite directions on the same model.

The builder playbook

Strip away the politics and the order leaves builders with a short, practical list.

• Design for the strictest market. If you sell into the EU at all, build to the EU bar and let the US and Indian baselines fall out as subsets. Building to the loosest regime and upgrading later is the costly route.
• Document provenance and risk now. Training-data sources, evaluations, limitations and oversight design. Cheap to capture continuously, painful to reconstruct under deal pressure.
• Treat US uncertainty as a risk, not a green light. The federal-versus-state preemption fight is unresolved. A surviving state law could still bind your product, so keep a watch on California, Colorado and the relevant attorneys general.
• Separate domestic from export policy. A model can be innovation-friendly to ship domestically in the US and still face export controls. Check both levers before you assume a model is freely available everywhere.
• If you sell into Britain, track three regimes. The US voluntary posture, the binding EU regime, and the UK's sector-led model under the BoE, FCA and ICO can all apply to one product depending on the customer.

The bottom line

The 2 June 2026 executive order draws a clear line: the United States will not require a federal licence to build or ship an AI model, and it prefers a voluntary, cooperation-based frontier framework over a statutory one. The European Union has chosen the opposite — binding obligations with penalties. The United Kingdom sits between them, governing AI through existing sector regulators. For an Indian or UK builder selling across borders, the takeaway is not which philosophy is right. It is that you now operate under more than one of them at once, and the cheapest way through is to build to the strictest market you touch and document your work from day one.