After the EO leak — what binds when the federal lever softens and the EU AI Act GPAI implementing act hardens
A revised executive order draft leaked to Axios this week showing the federal AI-safety lever being walked back. Two days later, the EU's GPAI implementing act published — operationalizing the AI Act's general-purpose-AI obligations on August 2026 timeline. The center of gravity in AI regulation just shifted from Washington to Brussels and the institute-led methodologies are filling the gap.
Two policy events, one week
A revised EO draft leaked showing a softer set of obligations on frontier-model developers — voluntary cooperation language replacing some of the mandatory reporting provisions. The EU published the GPAI implementing act, the technical regulation that operationalizes how general-purpose-AI provisions of the AI Act work in practice.
What the EO walks back
The leaked draft removes or softens: the mandatory red-team-result reporting threshold, the federal pre-deployment review requirement for compute above a stated FLOPs threshold, and the export-control reporting language. What stays: voluntary cooperation, the US AISI mandate, and the federal-funded research safety conditions. The bar to deploy at the federal-policy level just got lower.
What the EU AI Act GPAI implementing act locks in
The implementing act operationalizes: technical documentation requirements (model card + capability disclosures), the systemic-risk threshold and how it's computed, the substantial-modification trigger that resets compliance, and watermarking requirements for synthetic media. August 2026 is the operative date. Vendors selling into the EU market — which is everyone with a public API — have to be compliant by then.
The center of gravity in AI regulation just shifted from Washington to Brussels. The federal lever softens; the EU hardens; the institute-led methodologies become the de facto bar.
The transatlantic divergence
This is exactly the regulatory split-personality the industry was negotiating around. Models compliant with EU GPAI obligations may exceed what US federal rules now require. Labs will choose: comply with the higher bar globally (Anthropic-style), or run two compliance regimes (Google-style), or skip the EU market for some product tiers (no one yet, but the option is real).
The forward read
- Two frontier labs adopt the EU GPAI bar globally by Q3. Compliance simplification beats geographic optimization at their scale.
- The US AISI fills the federal gap. Voluntary cooperation through the institute becomes the de facto US bar, which lands close to but not identical to the EU bar.
- The next administration's policy stance becomes the 2027 inflection. Until then, EU + institute-led methodologies are the binding constraint for product decisions.
Axios — Revised EO draft leak → · European Commission — GPAI implementing act → · Lawfare — Transatlantic AI regulation divergence analysis →