Fostering global AI safety: EU AI Office participates in inaugural International Network of AI Safety

This milestone gathering marks a significant advancement in fostering global cooperation on AI safety. The meeting is organised around three tracks.

In Track 1, the Network will discuss initiatives to mitigate risks associated with synthetic AI generated content, focusing on digital content transparency techniques as well as safeguards at the AI model and system level to reduce and prevent harmful or illegal synthetic outputs such as child sexual abuse materials. Participants will discuss principles and initial practices. While technical methods are still evolving, best practices can foster stakeholder alignment on transparency and mitigation of risks arising from synthetic content and demonstrate trust to targeted audiences. Complementary approaches - including normative, educational, regulatory, and market-based measures - are essential.

Input from stakeholders during the meeting will refine this work, setting the stage for discussions at the Paris AI Action Summit in February 2025.

Track 2 will focus on evaluation and testing of foundation models. The objective is to foster a joint understanding of how to conduct evaluations to eventually reach a state of complementary testing among network members. AI Safety Institutes will present a prototype of a joint testing exercise which will lay the foundation for discussing how to expand and evolve this in preparation for the AI Action summit in France.

In Track 3, the Network will endorse a Joint Statement on Risk Assessment of Advanced AI Systems. This technical document, coordinated by the EU AI Office with the UK, outlines the evolving science and practice of risk assessment. It aims to establish a shared technical basis to support the network’s ongoing efforts to create comprehensive and effective risk assessment strategies for advanced AI systems.

The joint statement introduces six key aspects essential for robust risk assessment:

• Actionable insights
• Transparency
• Comprehensive scope
• Multistakeholder involvement
• Iterative processes
• Reproducibility

Track 3 will discuss how this work could further advance until the AI Action Summit in France in February.

Find further infomation about the International Network of AI Safety Institutes.