The Trump administration must publish a voluntary AI framework by Aug. 1 under an executive order, and debates in Washington about what should be covered are unfolding as European and UK testing approaches inform those discussions. EU and U.K. experiences emphasize scenario-based testing, benchmarking tied to concrete harms, and independent verification.
Why this matters
U.S. allies have been wrestling for years with many of the same AI safety questions now facing the Trump administration. Whatever shape the U.S. framework ultimately takes, architects of the EU and U.K. approaches say it will likely be only a first step toward broader oversight.
Political context
President Trump began his second term by undoing his predecessor’s AI strategy and favoring a less regulatory stance. Yet as AI models have rapidly increased in capability, the administration has found itself building an oversight framework and deciding what to include.
Under the executive order, the administration is due to release voluntary AI guidance by Aug. 1. The government is considering a unified proposal from OpenAI and Anthropic that would create "an equal playing field" in which highly capable models are covered regardless of whether they are open- or closed-source, according to a person familiar with the discussions.
The EU and U.K. approach
The European Union has treated AI safety as a sustained technical and policy project for several years, using a flexible method that lets past incidents and specific cybersecurity use cases inform future rules.
One lesson from Europe is to start with specific threat scenarios. Chris Canal, CEO and co-founder of AI evaluation firm EquiStamp, said researchers begin with historical data on threats — for example, related to biological weapons — then measure how much AI would increase a bad actor’s capabilities and whether that rise creates a systemic risk to society.
Other core elements of the EU process include:
- Tying benchmarking — standardized evaluations of model capabilities and risks — to particular scenarios such as bioweapons or hacks of government databases.
- Emphasizing multiple, independent verifications rather than relying solely on companies’ own assessments.
Canal has worked with EU and U.K. governments to help shape these approaches.
Scientific grounding and talent flow
The EU aims to secure buy-in from the scientific community, grounding AI risk thresholds in peer-reviewed science before embedding them in policy and encouraging researchers to publish their methodologies. By contrast, the U.S. mostly relies on internal or industry-driven standards.
Technical experts are also gravitating toward the private sector, especially in the U.S. and U.K. Canal noted that there have been multiple instances where government staff gave short notice and then took attractive roles at firms like OpenAI.
Bottom line
Governments are converging on certain cyber risks, but the U.S. and its allies remain in the early stages of mapping AI’s broader societal challenges. The Aug. 1 deadline and proposals from entities such as OpenAI and Anthropic will shape how much of the European-style, scenario-based testing and independent verification is adopted in U.S. guidance.
