The Seoul Declaration And Frontier AI Safety Commitments

Following the Seoul Declaration For Safe, Innovative And Inclusive AI, we have joined the Korea AI Safety to further expand on how algorithms can fuel public, assistive and accessibility taxonomies in a safe manner, including both aspects of the models and related critical infrastructure.

Related

• The Bletchley Declaration And AI Safety

• AI Pact and the Next Steps for Implementing the AI Act: Access And Public Systems

• General-Purpose AI Code Of Practice (Working Groups 2, 3)

The Seoul AI Safety Declaration, adopted on May 21, 2024, during the AI Seoul Summit, builds upon the Bletchley Declaration from the previous year. Signed by leaders from 10 countries and the European Union, it emphasizes the need for international collaboration to ensure the safe, innovative, and inclusive development of artificial intelligence (AI). The declaration underscores the importance of human-centric AI that aligns with the rule of law. It also advocates for bridging digital divides and enhancing global cooperation to address AI's challenges and opportunities.

Key outcomes of the Seoul Summit include the creation of an international network of AI safety institutes. This network brings together organizations from the UK, US, Japan, France, Germany, Italy, Singapore, South Korea, Australia, Canada, and the European Union. Its goal is to promote collaboration on safety research, share best practices, and develop technical resources to advance AI safety.

Another outcome is the introduction of the Frontier AI Safety Commitments. Sixteen global AI companies, including Amazon, Google, Meta, Microsoft, Anthropic, and OpenAI, have pledged to define and disclose the "intolerable risks" posed by advanced AI systems. These commitments focus on key areas like cybersecurity, model autonomy, and societal disruption. The companies have promised to publish safety frameworks that outline their strategies for managing risks, which include increasing transparency, conducting internal and external red-teaming, and encouraging third-party vulnerability reporting. These efforts represent a major step toward proactive risk management and the establishment of global standards for safe AI development.

Frontier AI Safety Commitments

The declation is complemented with Frontier AI Safety Commitments - a set of voluntary pledges from leading AI organizations to manage the severe risks associated with powerful, general-purpose AI models

• Rigorous Risk Assessment: Companies commit to evaluating potential harms before and during training and before deploying their AI models. This involves both internal testing and external evaluations.

• Defining and Managing "Intolerable" Risks: A crucial commitment is to establish clear thresholds for severe, unacceptable risks. If these thresholds are met, and mitigations are insufficient, the organizations pledge not to develop or deploy the model at all.

• Robust Governance and Transparency: This involves setting up strong internal safety frameworks, publicly reporting on progress, engaging with external experts for "red-teaming" (stress-testing AI), boosting cybersecurity, and selectively sharing information about model capabilities and risks.

AI Safety And Public Systems

Out input and participation in a public discourse included how Frontier AI Safety Commitments can be further refined to align with a broad protocols and taxonomy of public and assistive systems, focusing on both foundational and applied AI models such as Large Language Models (LLMs), Small Language Models (SLMs), Vision-Language Models (VLMs), and emerging 3D Foundation Models, multimodal capacities. These models are critical for assistive technologies, healthcare, and public contexts. AI Safety Institutes must structure their efforts and capacity to encompass a wide range of systems and interfaces, from assistive robotics to accessibility tools. Special attention should be given to regulatory sandboxes and testbeds that simulate real-world conditions, ensuring AI models are safe and accessible. For example, assistive AI systems, such as voice-enabled devices or mobility aids, must undergo compliance testing in these environments to ensure they operate safely, especially when interacting with users who have varied cognitive, sensory, or physical impairments.

Frontier AI models should be evaluated based on critical technical factors like data requirements, computational power, electricity, and cyber resilience across sectors. This evaluation must also consider the practical impact on high-risk public areas such as workplaces, healthcare and education. Risks and thresholds should be defined by real-world applications, not just theoretical capacities. For instance, deploying a 3D foundation model in assistive technologies or healthcare requires rigorous assessment for robustness, avoiding misuse. Finally, echoing the public input, AI developers should disclose both pre- and post-mitigation safety evaluations to provide a clearer picture of the effectiveness of their risk management strategies.

• • •

References

¹ UK Government. "Seoul Declaration for Safe, Innovative and Inclusive AI by Participants Attending the Leaders' Session: AI Seoul Summit, 21 May 2024." GOV.UK. May 21, 2024.

² UK Government. "Seoul Statement of Intent toward International Cooperation on AI Safety Science, AI Seoul Summit 2024 (Annex)." GOV.UK. May 21, 2024.

³ UK Government. "Frontier AI Safety Commitments, AI Seoul Summit 2024." GOV.UK. May 21, 2024.

⁴ UK Government. "Seoul Ministerial Statement for advancing AI safety, innovation and inclusivity: AI Seoul Summit 2024." GOV.UK. May 22, 2024.

⁵ Center for Strategic and International Studies. "The AI Seoul Summit." October 15, 2024.

⁶ Korea Economic Institute of America. "Major Commitments Launched at the AI Seoul Summit." May 29, 2024.