The UK’s Contribution to Developing the International Governance of AI Safety and How its Policy Took a Sharp Turn

[Dr Mando Rachovitsa is an Associate Professor in human rights law at the School of Law, University of Nottingham and the Deputy Director of the Human Rights Law Centre]

Although the UK’s ambivalence on adopting a national regulatory framework for AI has received considerable discussion, the UK’s policy and role in advancing the international governance of AI safety remains to be unpacked. In line with its 2023 International Technology Strategy, the UK plays a leading role developing different facets of governing AI safety, encompassing the political edifice and the scientific landscape. This post, firstly, recognises the significant contribution of the UK and, secondly, highlights that the UK’s policy has recently shifted with major global implications for AI safety. 

Driving International Cooperation and Building Political Consensus

The capabilities of general-purpose AI models called our understanding of the ensuing risks into question. The first international, high-level political process that focused on advanced AI models and AI safety was the 2023 AI Safety Summit at Bletchley Park. The UK government organised and hosted the summit and articulated a provisional grasp of risk areas (societal risks misuse, human autonomy). The Bletchley Declaration, endorsed by 29 states, including China, is a political commitment to international collaboration so as to identify AI safety risks of shared concern and build a shared scientific and evidence-based understanding of these risks. The summit turned into the AI summit series (followed by the AI safety summits in Seoul and Paris) with states and other stakeholders appreciating the flexibility and informality which still galvanise developments. 

The UK is a founding member of another new multi-stakeholder initiative: the Global Partnership on AI (GPAI) set up in 2020 to support the responsible and human-centric development and use of AI in a manner consistent with human rights, and democratic values. The GPAI was geared towards bridging the gap between theory and practice on AI. The GPAI had a close relationship with the OECD (Organisation for Economic Co-operation and Development) by design, with its secretariat being hosted in the OECD. Soon thereafter, this relationship turned into an integrated, project-based partnership. The G7, in turn, created a policy framework on advanced AI systems, the Hiroshima AI Process – which underscored the need for a variety of risks to be researched and mitigated – aligned with the GPAI and the AI safety summits. The reporting framework for monitoring how AI companies adhere to the Hiroshima Process International Code of Conduct for Organisations Developing Advanced AI Systems is operationalised under the auspices of the OECD. AI companies (many of which come from 16 companies that committed to identify, assess and manage AI risks via an announcement by the UK and Republic of Korea at the 2024 Seoul AI safety summit) submitted their inaugural reports. The fact that the OECD sits at the epicentre of integrating partnerships is hardly surprising, since it is a reliable forum to “translate” high politics and principles into actionable best practices and frameworks implementable in national regulation.

The UK’s policy extends to contributing to the formation of political consensus during the negotiations and drafting of the first treaty on AI, the Council of Europe’s Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law. The expertise of the Turing Institute was also mobilised to create a governance framework for carrying out AI impact assessments – which came to be known as the HUDERIA methodology. Having worked closely with the Council of Europe since 2020, HUDERIA was adopted by the organisation, with the Turing Institute currently supporting the ongoing work of the Council of Europe Committee on AI to pilot the framework.

Creating the Science of AI Safety and Forming Scientific Consensus

The idea of creating a scientific report of global appeal on AI safety was promulgated by the UK in the first AI safety summit. The 2025 International Scientific AI Safety Report sets out the first comprehensive, authoritative, shared scientific understanding of AI safety and risks’ mitigation. The report and its preparation were carefully crafted to establish scientific and political consensus, with the UK Department for Science, Innovation, and Technology providing operational support. The report is the culmination of work by 100 AI experts, chaired by Yoshua Bengio and it is supported by an international expert advisory panel of representatives from thirty countries (including the UK, China and the US) the UN, EU and OECD. The report acknowledges a broad spectrum of AI safety risks and divides them into three main categories, namely risks from malicious use (AI security, harm to individuals through fake content, manipulation of public opinion), risks from malfunctions (reliability, bias, loss of control) and systemic risks (e.g., labour market risks, market concertation and single points of failure, risks to the environment, risks to privacy and copyright infringement). 

A second major contribution of the UK consists of forging a novel institutional governance model: the AI Safety Institutes (AISIs). In the aftermath of the 2023 AI safety summit, the UK established the UK AISI – the world’s first state-backed body dedicated to evaluating the safety of advanced AI models. In the last 2 years, 7 more AISIs have been set up (US, Japan, Singapore, South Korea, Canada, France, India). AISIs rationale is to equip governments with an independent scientific understanding of AI risks by conducting research and testing. The fact that evaluations of AI models in AI labs are not sufficient for proper oversight gave rise to the pressing need for states to have access to credible, external evaluations. In this sense, AISIs have a unique role in AI safety governance functioning as an institutional interface or a “bridge” between scientific assessment and policy.

The UK AISI was entrusted with a broad mandate and it is the only well-resourced institute. The UK AISI (followed by US and Singapore) has been successful in strategically establishing partnerships with leading AI labs to ensure access to AI models and it has the world’s largest team invested in leading the development of the science of AI safety, testing AI models and building scientific consensus. The institute has conducted extensive testing ranging from chemical, biological, and cyber risks to AI models’ ability to manipulate, safeguards, and impacts on society. Finally, the institute holds an important role (co-)leading many of the activities of the International Network of AI Safety Institutes, which was created in 2024, with the mission to accelerate the advancement of the science of AI safety by sharing resources and conducting joint evaluations.

Castles Made of Sand? The UK Policy’s Sharp Turn

International affairs caution us that ambitious vision and its consistent implementation remain fragile and may be subject to (temporary) change. Despite its long-standing commitment to AI safety in 2025, the UK shifted its policies away from AI safety, aligning with the US’s new priorities. This shift became prominent when the UK and US refused to sign the 2025 Paris Statement on Inclusive and Sustainable AI. 

The UK moved on to rebrand its AI Safety Institute to the ‘AI Security Institute’ followed, in June 2025, by the US renaming its AI Safety Institute to the ‘Center for AI Standards and Innovation’ (CAISI). Both organisations now focus only on security risks (see also here, here and here). AI security consists of protecting AI models, systems and infrastructure from external threats and malicious actors, whereas AI safety is engaged with the full spectrum of (un)intentional risks to systems, society and individuals. 

Notably, the UK, hand-in-hand with the US, are in the process of erasing the vernacular and expertise of AI safety altogether. The AI Security Institute distanced its public-facing materials from references to ‘societal impacts’ (replaced by the obscure ‘societal resilience’), ‘unequal outcomes’ and ‘harming individual welfare’, as well as ‘public accountability’ (replaced by ‘public safe and secure’). From the other side of the Atlantic, an executive order emphasised that the US ‘develops AI systems that are free from ideological bias or engineered social agendas’ and ‘censorship and regulations’. Both the UK and US administrations have provided pointers and instructions in their administrative apparatus to exclude AI safety expertise. The former UK Technology Secretary questioned the role of AISI staff to research and assess freedom of expression, bias or discrimination. The US National Institute of Standards and Technology issued instructions to scientists who partner with CAISI to eliminate references not only to ‘AI safety’ but also to ‘responsible AI’, and ‘AI fairness’ from the skills expected of its members. Erasing this political and regulatory vocabulary undermines not only AI safety but also fundamental tenets of responsible AI ethics principles, (principles of fairness and non-discrimination and accountability) as enshrined in the OECD Recommendation on AI and the UNESCO Recommendation on AI Ethics.

This highly politicised climate found its way into the International Network of AISIs too, when the US and UK suggested to rename the Network – a proposal that was rejected. It is clear that the US has distanced itself from the Network’s activities by not participating in the second and third joint testing exercises. The UK did participate and co-lead the last two testing exercises, but it remains to be seen whether it will continue to participate only insofar AI security is concerned. 

‘Safety Risks Don’t Suddenly Disappear’ 

‘We are doing science – and safety risks don’t suddenly disappear’ was the response of the Director of South Korea’s AISI  to the tensions that the new UK and US policies have caused. The US now focuses on the so-called ‘demonstrable risks’ suggesting that other than AI security related risks (e.g., societal, psychological) are hypothetical or are not subject to proof, despite evidence of documented harms in current deployments. However, unless risks are researched, identified, and assessed they will remain speculative leaving ‘a whole range of harms to people and society unaddressed’. The UK AISI’s new research agenda seems to leave little leeway for also researching risks related to societal resilience and human influence.

It is expected that the UK will continue to contribute to various corners of the international governance of AI safety. However, there is no doubt that the UK’s shifting policy leaves a serious gap in international policymaking and the scientific assessment of AI safety risks, begging the question of whether other states will step up to the challenge.