Agile Multilateral AI Governance Starts with Foundation Models

· 5 min read
By , and

The need to act yet remain agile 

Artificial intelligence (AI) is on everyone’s mind – most sectors will be transformed and policymakers seek solutions. At the international level, however, there are significant risks to acting too fast: launching institutions that cannot keep up with technological developments; proposing ideas that alienate low- and middle income countries (future majorities); or competing with more urgent national-level regulatory processes. While the current political attention opens a window of opportunity for AI governance at the multilateral level, we recommend international policymakers to avoid hasty actions that could reduce progress on national regulation.

This stepping stone approach to building up international AI governance should focus on risks just as much as opportunities. The benefits of AI will be reaped by tackling both the risks of social injustice and potential catastrophes equally, as well as using AI to reduce the digital divide and societal inequalities more broadly. The key to efficient AI governance is to focus on the most powerful AI systems – so-called “foundation models”. Similar to Apple or Microsoft with operating systems, big AI labs develop general-purpose software. Instead of “macOS” and “Windows”, they’re called “GPT” or “PaLM” and instead of “versions” (e.g. Windows 95 or Vista), they’re called “models” (e.g. GPT-3 and GPT-4).

Foundation models, as their name implies, act as the building blocks for a broad range of AI applications, and can be adapted to execute a variety of tasks. For example GPT-4 serves as the foundation of the well-known ChatGPT but also serves as the foundation for tools in financial services and governments. Foundation models are trained on vast amounts of data to autonomously learn useful skills and behaviours, and are then iteratively fine tuned through human feedback. As this development process requires extraordinary amounts of human talent, hardware, and data, foundation models can currently be developed by just a few laboratories – notably OpenAI, Google DeepMind, and Anthropic – powered by three big cloud providers: Microsoft Azure, Google Cloud, and Amazon Web Services.

Ideally, foundation models would only be accessible via a secure online interface, allowing developers to retain full control and continually monitor evolutions in model behavior.  Making foundation models open-source and freely available to all would mean that anybody could fine-tune a copy to their liking. This is dangerous not only because the AI could be instructed to achieve illegal ends, but because the models are mostly black boxes – leaving us with no clear understanding of their decision-making processes or any assurance that they align with human interests. Examples of well-intentioned AI training gone awry exist aplenty.

As AI research and development moves faster than the multilateral system can understand and respond to, the UN should first seek to act as a guide and mediator. It should focus on regulating the development and deployment of foundation models, managing the distribution of profits accrued by their developers, and facilitating discussions among key stakeholders. This way, the multilateral system can (a) help consolidate and diffuse effective national and regional regulation, and (b) lay the stepping stones for a multilateral regime complex for the safe development and deployment of foundation models.

The goal of multilateral activity in the next few years should be to lay the stepping stones toward an agile regime complex for AI. By avoiding hurried moves and rushed proposals for new institutions, the multilateral system can maintain the flexibility to adapt to technological changes, geopolitical shifts, and national regulatory efforts as they arise. A regime complex of this nature might include institutions similar to an IPCC, IAEA, CERN or ICAO – with varying degrees of enforcement, verification, or standardization powers. But prematurely advancing complex proposals while national legislators are making up their minds is not going to help increase the focus of discussion.

Longer-term, a secondary goal of multilateral activity might be to encourage the development of a treaty for responsible AI development in collaboration with key nation-states, aiming for an initial agreement between the two great powers: China and the US. Given the low likelihood of a successful agreement between great powers, a parallel process, including especially the EU and India as crucial mediating parties, might create the necessary pressure and precedent to eventually consolidate an international agreement and entity for the responsible development of AI.

The shape of international institutions should not be decided too early, and certainly not without the buy-in of the global leaders in AI technology. An inclusive and well-mediated stakeholder process will probably lead to the best outcome.

Proposals for the 2023 ministerial meeting

The 2023 preparatory ministerial meeting of the Summit of the Future offers an opportunity to announce multilateral AI governance proposals. Following the considerations above, we recommend the ministerial announce three stepping stones.

  1. The appointment of a high-level advisory board on foundation model governance, tasked with analyzing AI-driven catastrophic risks that would irreversibly shape the course of humanity, and sharing these insights with key multilateral actors.
  2. The development of an inclusive process to develop standards for a comprehensive monitoring system of AI R&D to foster the inclusive and safe development of AI, balance risks and opportunities, as well as ways to deploy unbiased foundation models.
  3. The establishment of a forum to promote the participatory governance of AI developments, including the necessary capacity-building support for low- and middle-income countries to develop national expertise on the topic.