An export-led approach to AI governance for middle powers
How countries can influence frontier AI without building it
Dr Imogen Stead, Dr Jess Whittlestone & Hamish Hobbs
Recent months have seen an emerging debate about the role that middle powers should play in navigating and developing advanced AI. Some have suggested that middle powers - which we define here as countries with significant international influence but which wouldn’t be classed “great powers” - should pool resources to compete directly on frontier AI, while others have focused on adapting and increasing resilience to AI-driven change. In this blog, we present a third, complementary perspective: middle powers can play a valuable role in global AI governance by “exporting” relevant information, capabilities, analysis and governance approaches to frontier AI nations.
Competing directly to build frontier AI isn’t the only way for powers to influence its development
A number of different positions on the role of middle powers in AI governance have emerged in recent months.
Some have suggested countries join forces to try and compete at the frontier of AI development. A paper from last November led by the Oxford Martin AI Governance Initiative suggests pooling compute, talent and data across nations to build advanced AI systems. More ambitiously, a group of authors from Conjecture and Control AI suggest that a large enough consortium of middle powers could aim to centralise all advanced AI development, and in doing so control its safety.
Others take a different view, suggesting middle powers should accept they cannot realistically influence frontier AI development, and instead focus on their own national strategies. For example, the Tony Blair Institute focuses on how middle powers can capture the economic benefits of AI by building open-source ecosystems, while Anton Leicht emphasises the importance of building national resilience to the changes that transformative AI will bring.
These perspectives identify real challenges and important priorities. Middle powers’ dependency on the US and China creates vulnerabilities, and it is reasonable for these countries to try to have a stake in frontier AI development. But competing at the frontier may be an unrealistic goal for many middle powers at this point, and focusing on adaptation and resilience might actually bring more domestic benefits. This can make it seem like middle powers have a choice between a long-shot bet at influencing frontier AI development, and “giving up” on influencing the most important parts of AI development entirely.
We suggest an additional option which strikes a middle ground between those already discussed: middle powers can influence frontier AI development, but may be most effective at doing so indirectly via “exporting” relevant information, capabilities, analysis and governance approaches to frontier AI nations.
How middle powers can influence frontier AI via an export-led approach
It is possible to influence how frontier AI is developed, used, and governed without building a domestic frontier AI lab. A middle power can produce and share information, analysis, innovations, services and governance approaches with those actors who are developing frontier AI.
Not every middle power will be able to exert this kind of influence, but many will. What’s important is that, in order to influence AI governance in this way, middle powers don’t necessarily need to build frontier AI systems or have military or economic dominance. Instead, they need some kind of unique expertise or selling point related to developing and governing secure and beneficial advanced AI, which can be recognised internationally. We’ll discuss some examples of this more below.
1. Exporting information and expertise
Decisions about how frontier AI is developed and deployed depend heavily on access to reliable, up-to-date information and research on emerging capabilities and risks. Middle powers can influence these decisions by producing credible, independent analysis that frontier developers and governments can use to inform their thinking.
The UK’s AI Security Institute (AISI) is an excellent demonstration of this model. Its evaluations of model capabilities and risks are shared with both the US Centre for AI Standards and Innovation (US CAISI) and with frontier companies themselves. When UK AISI identifies important capabilities, risks, or limitations of AI models, this feeds directly into development and deployment decisions by other actors.
A related example is the AI Alignment project, a global fund seeking to advance the field of AI alignment that is supported by the UK, Canadian and Australian governments, alongside other industry and philanthropic partners. The project has attracted significant philanthropic support and private funding, including from frontier labs such as Anthropic and OpenAI. This approach gives the UK, Canada and Australia an opportunity to contribute to the frontier of AI alignment science, while simultaneously strengthening their renowned research institutions and ability to attract top technical talent.
Middle power AI safety and security institutions in the International Network for Advanced AI Measurement, Evaluation and Science (formerly the international network of AISIs) have the opportunity to act as more neutral evaluators than entities with direct commercial or geopolitical stakes. Crucially, this kind of informational influence is very different from political influence, in that it enables well-informed decisions without overtly saying what those decisions should be. Reliable information and analysis can materially affect global AI policy in major jurisdictions, as well as developer decisions in frontier AI companies, without triggering the credibility concerns that overt advocacy might.
2. Exporting innovations and services
Middle powers can also develop innovations or services that can be used to support secure and beneficial AI development globally. This can include producing proof-of-concept innovations that can then be adopted elsewhere, often with positive spillover effects on AI governance, as well as providing key services that can be exported, such as frontier AI assurance services.
Consider, for instance, defensive capabilities for monitoring AI-related national security threats. If a middle power develops effective methods for detecting AI-enabled cyber operations or AI-enabled biological threats, these techniques have value for any country facing similar threats - including those developing frontier AI. Demonstrating that something works is often more persuasive than arguing it should be done. Building an industry providing these services can increase security globally, while providing domestic economic benefits.
One example of an exported innovation for AI security is the Inspect framework, developed by the UK AISI as an open-source tool to support frontier model evaluations. This framework has been used to build evaluations by frontier AI developers and evaluation organisations, including Anthropic. Middle power innovations and services can help to develop testing and verification capabilities that can be used by private sector actors to improve oversight and governance of their systems, supporting global risk mitigation.
3. Exporting governance and regulation
Finally, there is scope for some countries to govern the development of new technologies without controlling their production. A clear example is the EU AI Act, which leverages the EU’s large market to govern frontier AI companies who wish to sell products on this market. The EU AI Act is likely to meaningfully influence frontier AI development towards better risk management practices. While implementation remains ongoing and enforcement still uncertain, most frontier developers have already signed up to the AI Act’s Code of Practice for General-Purpose AI, which includes substantive commitments on safety and security practices that align with the regulatory principles in the AI Act. Companies face strong incentives to comply because the EU represents a major market and maintaining different development practices for different jurisdictions would be operationally complex and costly. However, the impact of the EU AI Act will depend upon the effective interpretation, implementation and enforcement, which will require technical capability and sustained political will. Similarly, European data protection law fundamentally changed how global technology companies handle user information, creating practices that extended well beyond EU borders. California’s vehicle emissions standards influenced automotive design worldwide, as manufacturers found it more efficient to meet the strictest standard everywhere rather than maintaining multiple production lines.
It is also possible to export governance and regulation without relying upon the leverage of a large market. Australia’s social media ban for users younger than 16 appears to have triggered pushes for similar regulations in a variety of other countries. Beyond technology policy, New Zealand’s move to setting an inflation target for its central bank in 1989 was widely copied and has now become the dominant approach to monetary policy globally. Where middle powers identify and adopt governance or regulatory innovations related to frontier AI, they may be able to catalyse wider global adoption through a similar mechanism.
Why this matters
AI development is likely to have profound impacts globally. Ensuring a wide range of countries can influence advanced AI development enables those with a stake in the outcomes to help shape the process of development and diffusion. This is not just an opportunity to strengthen one country’s global influence, it is an opportunity to ensure global risks are managed globally and global benefits are shared. AI middle powers have a direct interest in exporting ideas, innovations, services and governance approaches to support these goals. Maintaining this level of expert, innovative and governance capability also ensures that key AI-related capabilities are not purely concentrated within a single state or actor. The strategies outlined here are not only routes to positive influence over frontier AI development, but also potential mechanisms for maintaining the distributed capacity that makes dangerous concentrations of power less likely.
A modest suggestion. The two nations of France and the UK together have more Fields medals than any other country. They certainly form the nucleus of a European superpower in mathematics. Surely rather than trying to emulate what other richer countries do, within a few years there could be a significant breakthrough in making AI more efficient and cheaper: the DeepSeek approach.