In January, the UK Government reiterated its manifesto commitment to introduce a bill to regulate the most powerful AI models, arguing that it was ‘essential to act quickly’.
However, since then, there have been significant contextual shifts that raise questions for an AI bill – from the US administration’s vocal opposition to AI regulation, to the UK Government’s drive for economic growth.
Despite this, we argue that the UK’s AI bill remains urgent for global AI security. By developing pro-security legislation that doesn’t onerously interfere with innovation, the UK can create a regulatory blueprint that other nations can adopt with global impacts.
To do so, the bill must strike a pragmatic compromise between addressing the fast-developing security risks and navigating the political headwinds. But how?
In this post, we argue for an approach that focuses not only on intervening in how models get developed and deployed, but also on building a much broader strategy of preparedness for more dangerous models being widely available in the coming years.
The UK’s AI bill and a changing context
We’ve long advocated for a UK bill targeted at frontier AI security, as something we believe can increase the likelihood that any dangerous behaviours or vulnerabilities in AI models are caught and addressed as early as possible, and ideally before they are released.
However, our view is that the evolving geopolitical picture of the last six months means that approaches to AI security may need to change tack. This is for several reasons.
- The US has expressed strong opposition to a precautionary approach to frontier AI regulation, both in J.D. Vance’s remarks at the AI Action Summit in Paris in February, and in its ‘America’s AI Action Plan‘ in July, with increasing perceptions of a technological ‘race’ that must be won by the US.
- The UK Government is increasingly focused on pursuing adoption and economic growth, and is itself more tentative about the possibility of legally binding rules that might hamper frontier AI innovation.
- Innovation at Chinese frontier AI companies is increasingly significant, which brings additional emphasis to the challenge for the UK’s legislation to have the necessary extra-territorial influence.
Our principal conclusion from this picture is that it is now less politically feasible to seek to implement a strict set of conditions that must be met before a model can be released, also known as pre-deployment safety measures.
While some pre-deployment measures remain vital, and should be mandated where feasible and incentivised where not, this type of regime appears to be less feasible compared to 12 months ago.
Nonetheless, the proliferation of increasingly high-risk models is moving very fast.
The latest models are increasingly viewed by AI companies as on a path towards uplifting malicious actors in the development of biological weapons. They are are also being given greater autonomy to act in the world, unsupervised despite their emerging capacity to deceive their operators.
If, as Peter Kyle, the UK Secretary of State for the Department of Science, Innovation and Technology (DSIT), argues, we are only a few years away from transformative AI capabilities, we quickly need a new strategy for AI security that is fit for the political realities of 2025.
Moving towards preparedness
We propose that a ‘preparedness approach’ to AI security can provide this solution, being both more politically feasible while also clear-eyed about the technological trends.
This approach entails accepting that dangerous AI models could be widely proliferated within a few years, and undertaking a wide range of measures to anticipate, prevent, prepare and respond to this scenario.
A preparedness approach treats preventative safety measures – such as pre-deployment safeguards – as just one component of a much broader and more holistic strategy. As the Ada Lovelace Institute has recently argued, we need not only preventative measures, but also powers to respond to when those measures fail.
This type of strategy emulates the UK’s existing approach to biological security. We seek to prevent pandemics from occurring, but also plan for when they do. So too must we try to prevent dangerous AI models from being released but also plan for their impacts once they are.
Concretely, this means breaking down AI security into the following policy objectives:
- Anticipation: The UK Government should secure access to a sufficient breadth and quality of information about security risks from frontier AI to effectively develop and assess security risk scenarios.
- Prevention: The UK Government, frontier AI companies (FAICs) and the broader private sector should be incentivised to implement proportionate risk governance and management to mitigate security risks, noting this may only be partially successful, and critical sectors and infrastructures must be protected from catastrophic failures caused by advanced AI.
- Preparation: Plans for responding to major AI impacts and incidents should be regularly stress-tested and enhanced, with effective participation from FAICs, to avoid escalation into catastrophic impacts.
- Response: The UK Government should have sufficiently strong levers to direct fast, decisive action from FAICs to manage and contain an acute incident.
This type of approach can provide a framework for improving AI security that is fit for today’s geopolitical and technical context.
A UK AI bill is needed to improve preparedness for AI security risks
There are many interventions to improve preparedness – that is, anticipation, prevention, preparation and response – for frontier AI’s security risks.
Some of these do not require legislation. For example, the UK AISI is already undertaking extremely impactful work in service of these objectives outside of a legislative framework.
Where policy objectives can be achieved through non-legislative measures and their iteration, we are supportive of doing so – new legislation is not the answer to everything. (We will be publishing more soon on the UK Government’s non-legislative policy options).
However, some legislation remains a necessary component for preparedness. In the table below, we provide a view on the viability of achieving each preparedness objective without legislation.
We strongly believe that, given the analysis outlined in the table, a set of legislative measures will be essential for an effective approach to AI security.
Preparedness and innovation can coexist
A preparedness approach broadens the focus for interventions beyond trying to alter what models are released and how – one key objection to AI regulation, particularly from the US.
- Anticipation can be supported through legal requirements for transparency that is aimed at enhancing the UK Government’s capacity to foresee AI’s impacts. These measures do not need to interfere with the development or release of frontier AI models, as indicated by the AI industry welcoming greater transparency.
- Prevention can avoid onerous intervention by: (i) focusing on incentives that encourage positive behaviour rather than only requirements with penalties; and (ii) considering intervention points other than model development, such as enhancing the security of the UK’s critical national infrastructure, or regulating the activities of suppliers (e.g. mandating nucleic acid screening).
- Preparation can be supported largely through non-legislative measures, and any legal duties that might be introduced (e.g. to participate in preparatory activities such as table-top exercises), if desirable, would not need to affect model development.
- Response can be supported by the introduction of new powers – such as the powers to direct FAICs in an emergency – that would only be used in extreme circumstances and not affect business as usual model release.
Together, this preparedness approach can help to enhance AI security without limiting attention to strict pre-deployment requirements, in response to the new political climate that we are now in.
The UK’s AI bill can have an urgently needed global impact
This approach can have implications that go far beyond the UK’s borders – in fact, we believe the UK has a crucial role to play in the global shift towards preparedness for transformative AI.
While the EU GPAI Code of Practice sets important new standards for some FAICs, its emphasis is not on preparedness. It is also not an easily replicable blueprint for legislation, as it is regarded by many as too onerous, especially the US.
The UK is therefore in a position to introduce a US-aligned blueprint for AI preparedness legislation that other nations – especially other middle powers and potentially the US itself – can adopt.
This is partly because the UK enjoys a ‘special relationship’ with the US, with President Trump reaffirming this with less punitive tariffs on the UK economy. It is therefore in a good position to align the US administration with a model of AI regulation that other middle nations can adopt, referred to by the Oxford Martin School as a ‘London effect’.
We’ve already seen the London effect on AI in action – the UK launched the world’s first AI Safety Institute (with similar institutions being set up around the globe), AI Summit (with the next in India, following the Prime Minister’s trade agreement), and AI Opportunities Action Plan (which was soon followed by the United States’ own AI Action Plan). There is, we believe, strong evidence for a London effect, and good reason to believe the AI bill will be the next step in that process.
The opportunity for the UK AI bill is therefore significant: set a global standard for improving global anticipation, prevention, preparation and response to the extreme impacts and risks we face from transformative AI. We believe the UK is on the path to ground breaking global leadership on AI governance once again.
Recommendations for DSIT
As DSIT drafts its AI bill, ahead of an introduction to Parliament slated for 2026, we offer the following recommendations:
- Apply a ‘preparedness framework’ to the bill’s approach to AI security, with laws targeted at achieving significantly enhanced capacity for anticipation, prevention, preparation and response to AI’s risks.
- Alongside this, develop a broader AI Security Strategy (modelled on the Biological Security Strategy) that also includes accompanying non-legislative measures, and clarify responsibilities for preparedness, including the risks for which DSIT is the Lead Government Department.
- Once passed, work to diffuse the principles of this approach internationally to improve international harmonisation, as well as the preparedness of other nations.
