Abstract
The UK’s convening power, close US relationship, and world-leading state AI technical capability give it outsized potential to shape global frontier AI governance. With the promised frontier AI Bill repeatedly delayed, the authors set out three non-legislative pathways for cementing that leadership: strengthening AISI’s international engagement, advancing international alignment on risks and mitigations, and influencing frontier AI companies’ behaviour.
Framing
- AI Minister Kanishka Narayan has stressed implementing “an AI policy which can double-down on our strengths in key areas” to ensure the UK has “a seat at the table” as capabilities develop.
- Government focus so far has been on investing heavily in sovereign UK AI infrastructure to maintain geopolitical relevance.
- The authors argue the UK is also particularly well positioned to influence global frontier AI governance conversations, given that it:
- serves as Network Coordinator for the International Network for Advanced AI Measurement, Evaluation and Science (formerly the International Network of AI Safety Institutes);
- convened the inaugural international AI safety summit;
- was the first country to establish a government institute dedicated to frontier AI research and model testing.
- The legislative route has stalled. Despite promises from two successive governments, no targeted AI regulation has materialised. Technology Secretary Liz Kendall recently indicated the Government is unlikely to bring forward new regulation beyond extending the online safety regime.
- The authors call this an opportunity missed: failing to regulate leaves the UK insufficiently equipped to address extreme domestic risks and forgoes a chance to cement global leadership. But an AI Bill is not the only route to a seat at the table.
The UK’s “AI soft power”
Three strategic advantages underpin the UK’s outsized potential influence:
- Strong convening capability. The 2023 AI Safety Summit brought US and Chinese representatives together for the first time to discuss AI safety and security research. The resulting declaration led to the international AI safety report and the international AISI network, which the UK now coordinates.
- Close UK–US relations. The relationship has held despite administration changes in both countries. September’s technology-focused Memorandum of Understanding recognised the two as “most trusted security and defense partners” and committed to promoting “secure AI innovation,” including joint testing and standards development through the AISI–CAISI partnership.
- Best-in-class state AI technical capabilities. In little over two years, the UK AI Security Institute has become a world-leading example of public sector frontier AI safety and security research and model testing. Its stable funding, research credentials, and technical talent have enabled trusted voluntary partnerships with developers, compounding UK international credibility.
Pathway 1: Strengthening AISI’s international engagement
AISI should systematise ways to diffuse its knowledge internationally, using the UK’s new Network Coordinator role to reinvigorate the international coordination mechanism.
- Low-hanging fruit: use the Network Coordinator role to establish strong communications channels and regular meeting structures for closer collaboration and information-sharing; use the leadership position to raise the salience of emerging extreme AI risks identified by the network in other international political fora, acting as an early warning mechanism for governments.
- Higher-ambition intervention: institute a separate, private information-sharing partnership between AISI and the US CAISI, building on the Memorandum of Understanding — an early warning mechanism for national security-relevant risks emerging from UK AISI testing of US models, increasing joint strategic awareness and enabling appropriate US responses.
Pathway 2: Advancing international alignment on risks and mitigations
The UK should leverage convening power and its US relationship to build mutual understanding of the evolving risk landscape and coordinate mitigations, including proposing baseline international agreements — for instance defining red lines for governments’ use of frontier AI capabilities — at fora such as the UN, NATO, and Five Eyes.
- Low-hanging fruit: champion best-practice regulation and research in other jurisdictions — such as the Safety & Security chapter of the EU AI Act’s GPAI Code of Practice and the Singapore Consensus on Global AI Safety Research Priorities — to inform forthcoming US AI policy interventions and facilitate high-level multilateral alignment on a baseline set of mitigations.
- Higher-ambition intervention: work with AISI’s research and strategic awareness teams to identify extreme risks requiring dedicated global governance institutions, and secure buy-in from key partners such as the US to create new cooperation mechanisms — one example being an institution dedicated to international alignment on serious AI incident monitoring and reporting frameworks.
Pathway 3: Influencing frontier AI companies’ behaviour
AISI’s technical expertise plus UK convening power make the UK comparatively well placed to align developer incentives toward effective risk mitigation, strengthening industry partnerships and pushing for new voluntary commitments that raise the floor on best practice.
- Low-hanging fruit: build on AISI’s trusted relationships to incentivise greater transparency and information-sharing proportionate to the speed of capability development; coordinate privately with companies to secure buy-in for extending the scope of voluntary international commitments on frontier AI risk transparency and management.
- Higher-ambition intervention: identify areas of extreme AI risk not adequately covered by existing regulatory regimes globally — such as risks from the race between frontier companies to achieve significant recursive self-improvement capabilities — and design a voluntary code of practice to influence responsible industry practices.
Status
The authors describe the post as the start of their thinking rather than the end, and invite input from others working on similar issues to refine the recommendations.