How summits in Seoul, France and elsewhere can galvanize international cooperation on frontier AI safety
Last year, the UK government hosted the first major global summit on frontier AI safety at Bletchley Park. It drew global attention to rapid advances on the frontier of AI development and took concrete international actions to address potential future risks, including the Bletchley Declaration. New AI Safety Research Institute. and the International Scientific Report on the Safety of Advanced AI.
Six months after Bletchley, the international community has an opportunity to build on its momentum and galvanize further global cooperation with this week’s AI Seoul Summit. We share some thoughts below about how this and future summits can foster progress toward a common global approach to frontier AI safety.
AI capabilities continue to evolve at a rapid pace
Since Bletchley, we’ve seen strong innovation and progress across the field, including Google DeepMind. AI continues to drive breakthroughs in important scientific fields, and our new AlphaFold 3 model predicts the structure and interactions of every molecule in life with unprecedented accuracy. This research will help change our understanding of the biological world and accelerate drug discovery. At the same time, our Gemini family of models is already making products used by billions of people around the world more convenient and accessible. We’re also working to improve the way models perceive, reason, and interact, and recently shared our progress building the future of AI assistants with Project Astra.
This advancement in AI capabilities promises to improve the lives of many people, but it also raises new challenges that need to be addressed collaboratively in many key safety areas. Google DeepMind is working to identify and solve these challenges through pioneering safety research. In the past few months alone, we have developed a comprehensive assessment of the safety and liability of advanced models, including initial work assessing critical features such as deception, cybersecurity, self-propagation, and self-propagation. We have shared an evolving approach to doing so. inference. We also released an in-depth study on aligning future advanced AI assistants with human values and interests. Beyond the LLM, we recently shared AlphaFold 3’s approach to biosecurity.
This effort is driven by the belief that we must innovate in safety and governance as quickly as we innovate in functionality, and that both must be done in parallel, continuously informing, and mutually reinforcing each other. It is promoted on the basis of
Building a global consensus on frontier AI risks
To maximize the benefits of advanced AI systems, build international consensus on key frontier safety issues, such as anticipating and preparing for new risks beyond those posed by current models. Must be. However, given the high degree of uncertainty about these potential future risks, there is a clear need for policymakers to provide independent, evidence-based views.
The launch of the new Interim International Scientific Report on the Safety of Advanced AI is therefore a key element of the AI Seoul Summit, and we look forward to presenting evidence of our findings later this year. Over time, this type of effort could become a central input into the summit process and, if successful, could loosely model the functioning of the Intergovernmental Panel on Climate Change and lead to a more permanent status. We believe that it should be given. This will be a vital contribution to the evidence base needed by policy makers around the world to inform international action.
We believe that these AI summits can provide a regular forum dedicated to building a common and coordinated approach to international agreements and governance. Maintaining a unique focus on frontier security will also ensure that these conferences are complementary to, rather than duplicative of, other international governance activities.
Establishing best practices in evaluation and a consistent governance framework
Evaluation is a key element needed to inform AI governance decisions. These allow the functionality, behavior, and impact of AI systems to be measured, providing critical information for risk assessment and designing appropriate mitigations. However, the science of frontier AI safety assessment is still in its early stages of development.
The Frontier Model Forum (FMF), launched by Google with other leading AI research institutes, is collaborating with AI Safety Institutes and other stakeholders in the US and UK on best practices for evaluating frontier models. This is why. The AI Summit could help scale this effort internationally and avoid a patchwork of overlapping or mutually contradictory national testing and governance regimes. It is important to avoid fragmentation that can inadvertently compromise safety and innovation.
The U.S. and U.K. AI Safety Associations have already agreed to develop a common approach to safety testing, an important first step toward stronger collaboration. We believe that over time there is an opportunity to build this towards a common global approach. The first priority of the Seoul Summit may be to agree on a roadmap for a wide range of stakeholders to collaborate on the development and standardization of frontier AI assessment benchmarks and approaches.
It is also important to develop a shared framework for risk management. To contribute to these discussions, we recently introduced the first version of the Frontier Safety Framework. It is a set of protocols to proactively identify future AI capabilities that have the potential to cause serious harm and introduce mechanisms to detect and mitigate them. We look forward to significant advances in this framework as we learn from its implementation, deepen our understanding of AI risks and assessments, and collaborate with industry, academia, and government. Over time, we hope that sharing our approach will facilitate work with other collaborators to agree on standards and best practices for assessing the safety of future generations of AI models. I hope that you will.
Frontier Towards a global approach to AI safety
Many of the potential risks that may arise from advances on the AI front are global in nature. In the lead-up to the AI Seoul Summit, and with an eye toward future summits in France and other countries, we are excited about the opportunity to advance global collaboration on frontier AI safety. It is our hope that these summits will become dedicated forums to move towards a common global approach. Getting this right is a critical step in unlocking AI’s tremendous benefits for society.
