Artificial Intelligence (AI) is advancing. So, too, is international collaboration to ensure that advances are made in a safe and responsible way. In May, ten countries and the European Union (EU) met in South Korea and signed the “Seoul Statement of Intent toward International Cooperation on AI Safety Science,” which establishes an international network of AI safety institutes. This agreement builds on measures that several of its signatories have taken on AI safety since the Bletchley Park summit in November 2023. Since November, for example, the United Kingdom, the United States, Japan, and Singapore have established AI safety institutes, and the EU has set up an AI office with a unit dedicated to safety.
The Statement of Intent also builds on existing bilateral agreements. At the EU-US Trade and Technology Council meeting held in April 2024, the EU and the United States announced that the AI Office and the US AI Safety Institute would work together to develop the tools needed to evaluate AI models. Additionally, ahead of the Seoul Summit, the US AI Safety Institute signed a memorandum of understanding with the United Kingdom’s AI Institute, also aimed at building out a shared approach to AI safety, with an emphasis on developing testing and evaluation metrics.
The Statement of Intent signed at the Seoul Summit represents an important step forward in the AI safety conversation. It demonstrates both increasing international interest in and commitment to advancing the science needed to promote AI safety. To be successful in its implementation, however, the signatory countries will need to prioritize the most pressing areas of need for scientific practices, deepen their engagement with international standards-setting bodies, and collaborate with stakeholders across the AI ecosystem.
Why the Seoul Summit statement matters
The Statement of Intent matters for several reasons. First, it will help foster not only a common understanding of key AI safety concepts but also help advance common approaches to testing models (or otherwise ensure that approaches are interoperable). Indeed, there remains a lack of consensus around terms and taxonomy specific to AI safety—for example, the specific difference between a “frontier” AI model, an “advanced” AI model, and a “general purpose” AI model. There is also a lack of consensus on what constitutes red-teaming for AI, which is typically understood in cybersecurity as a process by which a team within an organization (a “red team”) attacks a system to expose vulnerabilities or weaknesses. There is also, in the context of AI, a lack of consensus about whether additional testing, evaluation, validation, and verification (TEVV) techniques beyond red-teaming are required to appropriately evaluate a system’s capabilities and risk. The Seoul Summit statement enables cross-border information sharing, allowing the institutes to learn from each other, share empirical findings, and identify best practices.
Second, the statement helps lay the groundwork for the institutes to share resources, such as expertise, datasets, and infrastructure. Indeed, beyond questions around red-teaming and testing and evaluation, there remains an important need to develop consistent metrics and criteria for risk evaluation, guidance related to transparency, and benchmarks for safety, reliability, and performance. Setting out a collaborative vision for how institutes intend to leverage joint resources can help to progress the research needed to advance work on these critical topics.
Finally, such an agreement can potentially contribute to technical capacity building and talent development. While the primary objective of these institutes is to advance the science of AI safety, they could play an important role in developing specific training curricula for practitioners, ensuring that guidance is practical and can be implemented.
How the AI safety institutes can turn intent into action
It’s early days for the network of AI safety institutes, and it remains to be seen exactly how collaboration will unfold. But to make meaningful progress on their stated objectives, there are three important objectives that will be vital for the institutes to focus on.
The AI safety institutes should:
- Prioritize collaborating to develop metrics for testing and evaluation of advanced AI models and ways to quantify emerging risks. While there are many areas worth exploring in the field of AI safety, the institutes should focus their immediate efforts on establishing collective practices that are most necessary to advance the development of trustworthy AI systems. One area the institutes should focus on is the testing and evaluation of advanced AI models. This is unsurprising given the ongoing focus on the role of red-teaming in AI risk management, but it is important that the institutes also seek to make progress on TEVV more broadly, including pre-deployment TEVV to assess and manage known and emerging risks associated with advanced AI systems. Indeed, stakeholders have agreed that assessing and understanding risks presented by such models is important to mitigating them, but there remain challenges in consistently measuring and evaluating those risks, particularly in cases where those risks are still emerging. The institutes should also seek to develop consistent ways to measure and report on model and system performance.
- Prioritize engagement in standards bodies that are undertaking standards development activities. International technical standards, and the work of bodies such as the International Organization for Standardization and International Electrotechnical Commission’s Subcommittee 42 on AI, will be critical to consistently implementing AI safety best practices. For example, bringing the institutes’ work on testing and evaluation of advanced AI models to international standards bodies will help promote international alignment and adoption. Conversely, engaging with international standards bodies may also help inform the safety practices and guidance that the institutes are seeking to develop.
- Seek to collaborate with stakeholders from across the AI ecosystem. Announcing a network of publicly backed institutions is important. But it is equally important that the AI safety institutes prioritize engagement with stakeholders from across the AI ecosystem. This includes the private sector, civil society, and government officials and policymakers. In particular, the institutes should consider how to enable cooperation between affiliated consortiums. For example, the AI Safety Institute Consortium within the National Institute of Standards and Technology (NIST) is comprised of stakeholders from multiple parts of the ecosystem, and NIST has sought to engage the Consortium on initial AI trust and safety work under the Executive Order on Safe, Secure, and Trustworthy AI. To the extent that other institutes pull together similar groups, it will be imperative that there is understanding of exactly how the expertise of those groups will feed into the wider network.
The Statement of Intent is a necessary first step in fostering international collaboration on AI safety. By focusing on making concrete progress on metrics for AI safety, engaging with standards-setting bodies, and ensuring that stakeholders from across the AI ecosystem are involved, countries can build on the Statement to make concrete advances in AI safety.
Courtney Lang is a nonresident senior fellow at the Atlantic Council’s GeoTech Center.
Further reading
Fri, Nov 3, 2023
Dispatch from Bletchley Park: Where does transatlantic AI cooperation stand?
New Atlanticist By Mark Boris Andrijanič, Nicole Lawler
Politicians and business leaders just met outside of London for a summit on how to regulate artificial intelligence. Here’s how to take the collaboration to the next level.
Fri, Jun 21, 2024
The sustainability questions policymakers should be asking about AI
GeoTech Cues By
Focusing on the sustainability of the AI industry offers an opportunity to steer entire industries toward contributing to a positive future.
Mon, Jun 10, 2024
Generative AI provides a toolkit for decarbonization
EnergySource By
Artificial intelligence models have long provided niche tools for energy a climate technologists. With the unique capabilities of generative AI, spanning applications in strategy, regulation, and finance, opportunities (and responsibilities) have emerged for all decarbonization stakeholders.
Image: Themenfoto,Symbolfoto : Kuenstliche Intelligenz,KI *** Theme photo,Symbol photo Artificial intelligence,AI