White House tells OpenAI and Anthropic to let U.S. review new models before sharing them with British testers

The White House has directed OpenAI and Anthropic to delay sharing new AI models with the United Kingdom's AI Safety Institute until U.S. government agencies complete their own review process first.

This directive reflects an emerging tension between international AI safety collaboration and domestic regulatory control. The U.K. established its AI Safety Institute as a neutral research organization to test frontier AI systems independently, partnering with leading AI companies to evaluate safety risks before deployment. That arrangement now faces friction from Washington, which wants first-look access to models heading overseas.

The timing matters. Both OpenAI and Anthropic have engaged with the U.K. institute as part of broader efforts to demonstrate responsible development practices. These partnerships carry symbolic weight in an era when AI safety governance remains fragmented across nations. Britain positioned itself as a lighter-touch regulator, fostering voluntary industry cooperation rather than mandatory compliance. The White House pressure threatens that model.

The directive signals U.S. concern over two issues. First, the administration worries about intelligence and national security dimensions of advanced AI systems. Allowing international allies early access to unreleased models could reveal technical capabilities, training methodologies, or vulnerabilities before U.S. officials understand them. Second, domestic regulators increasingly view AI development as a competitive arena where first-mover information advantage matters for shaping national policy.

The National Institute of Standards and Technology (NIST) and other U.S. agencies have been developing AI safety evaluation frameworks, but these remain in early stages. A requirement that American review comes before international sharing could slow the model release cycles that both companies currently maintain.

For OpenAI and Anthropic, the directive creates operational friction. Both companies have publicly committed to transparent safety testing and external evaluation. Complying with the White House request means explaining delays to the U.K. institute without appearing to prioritize national security over international trust. The companies already navigate complex geopolitical waters around export controls, data residency, and compute access. This adds another compliance layer.

The U.K.'s response remains uncertain. The AI Safety Institute was explicitly designed to function as an independent, trusted intermediary in a fragmented global AI governance landscape. Subordinating its review timeline to U.S. processes undermines that independence. Britain's government has positioned itself as a bridge between American and European approaches to AI regulation. This White House move suggests Washington views that bridge as secondary to its own review authority.

This also reflects broader U.S. strategy to maintain dominance over frontier AI development. America hosts the major model developers and the largest compute infrastructure. Controlling information flow about cutting-edge capabilities before they reach even close allies extends that control. It echoes Cold War-era technology gatekeeping, adapted for AI.

The precedent matters more than any single model release. If the White House successfully establishes that U.S. review comes before international collaboration on frontier models, other nations will demand similar rights. Japan, South Korea, Singapore, and the EU will expect early access too. Companies could face impossible coordination challenges, effectively centralizing AI governance decisions in Washington.

For now, OpenAI and Anthropic face a choice between diplomatic friction with a key ally or friction with their home government. That choice likely resolves in favor of Washington, but it signals that international AI safety cooperation will remain subordinate to national regulatory interests.