Britain wants to test AI models before they launch

While AI labs race to train ever-larger models behind closed doors, Rishi Sunak wants the UK to become the world's lead safety inspector.

Speaking in London ahead of the UK's global AI safety summit, the British Prime Minister announced that his government will establish the world's first AI Safety Institute. The goal is straightforward: test and evaluate frontier systems before they are released to the public.

Backstage passes to frontier models

The new body will build directly on the UK's existing Frontier AI Taskforce. Crucially, Sunak noted that the taskforce has already secured privileged access to models developed by industry heavyweights, including OpenAI, Google DeepMind, and Anthropic.

Sunak said the institute will investigate risks spanning from everyday harms like algorithmic bias and disinformation all the way to extreme, catastrophic threats. The plan is to make the institute's research globally available.

That transparency comes with an asterisk, though. A UK official familiar with summit negotiations told POLITICO that while baseline research will be published, sensitive national security intelligence will remain restricted to a smaller circle of like-minded allies.

An IPCC for artificial intelligence

Sunak is also pushing for a new multilateral body modeled on the UN's Intergovernmental Panel on Climate Change (IPCC).

The proposed expert panel would be nominated by countries and organizations attending the summit, with a mandate to publish regular state of AI science reports.

Unlike existing Western-leaning coalitions, such as the Franco-Canadian GPAI, Britain plans to include Chinese representation. Sunak defended the move bluntly, arguing that no serious AI strategy can exist without engaging all of the world's leading AI powers.

Why it matters

For anyone building or using frontier AI, this signals a massive shift. The era of pure corporate self-policing is coming to an end, and pre-deployment government red-teaming is taking its place.

Geopolitically, the move looks like a calculated bid by the UK to carve out relevance between the US and the EU. While Brussels bets on heavy compliance regulations and Washington leans on voluntary pledges, London wants to position itself as the world's technical testing ground.

We think getting genuine internal access to proprietary frontier models is a smart, concrete mechanism that goes beyond empty political posturing. Bringing China to the table is also a pragmatic reality check that Western policymakers often dodge.

However, the real question is leverage. Independent safety testing sounds great, but if access terms remain beholden to tech giants, or if sensitive intelligence is secretly funneled to select allies, the institute will struggle to be seen as a truly neutral global authority.