What was agreed at recent international AI safety summits?

Dialogue and Declarations, Not Binding Treaties

In recent years, several international AI safety summits have brought together government officials, representatives from major AI companies, researchers, and civil society groups from multiple countries to discuss risks associated with advanced AI systems and potential approaches to governance. These gatherings have generated considerable public attention and have produced various declarations and statements. It’s important to be precise, however, about what these outputs actually represent: generally, non-binding expressions of shared intent, principles, and in some cases voluntary commitments, rather than binding international treaties with enforceable legal obligations.

This distinction matters because summit declarations are sometimes reported or perceived as more binding or consequential than they formally are, which can create a misleading impression of how much concrete, enforceable international governance currently exists for AI.

Common Themes Across Summit Outputs

While the specific content of any given summit’s declaration or outcome statement varies, some recurring themes have appeared across multiple such gatherings. These generally include expressions of shared concern about risks associated with the most advanced, or “frontier,” AI systems; commitments to greater international cooperation and information-sharing on AI safety research; and calls for continued dialogue among governments, industry, and researchers on how to jointly approach AI governance challenges going forward.

Some summits have also included voluntary commitments specifically from participating AI companies, such as agreeing to certain safety testing or evaluation practices for their most advanced models before public release, or committing to share information about model capabilities and risks with participating governments under certain circumstances. These company-level commitments, like the broader summit declarations, have generally been voluntary rather than legally enforceable.

Why the Outputs Have Stayed Non-Binding So Far

The non-binding nature of these summit outputs reflects the genuinely early stage of international AI governance as a coordinated field. Reaching binding international agreement typically requires extensive negotiation, alignment of vastly different national interests and legal systems, and often ratification processes that take considerable time — none of which happens quickly, particularly for a technology as economically and strategically significant, and as rapidly evolving, as AI. Many observers view these summits as an important, incremental part of a longer-term process of building the trust, shared understanding, and diplomatic groundwork that might eventually support more binding forms of international AI governance, rather than expecting binding agreements to emerge from any single summit.

Bottom Line

Recent international AI safety summits have generally produced non-binding declarations, shared statements of intent, and in some cases voluntary company commitments focused on themes like international cooperation and frontier AI risk assessment — meaningful steps in building international dialogue, but not binding treaties or enforceable global rules, reflecting the still-early stage of coordinated global AI governance.

Go deeper

• What is an ai bill of rights and has any government actually adopted one?
• What does an AI safety job actually involve?