International AI Treaties and Summits: Where Global Governance Actually Stands

The starting point: Bletchley

The first global AI Safety Summit, held at Bletchley Park, UK in November 2023, produced the first document of its kind, committing signatories to human-centric, safe, and responsible AI development. Twenty-eight countries plus the European Union signed, later joined by New Zealand. Notably absent: Russia was not invited, and much of Africa and Latin America beyond a handful of states did not participate.

What it did and didn’t do

The Bletchley Declaration is a political statement of shared concern, not a binding treaty — it imposes no enforceable obligations on governments or companies. Its practical effect was establishing a recurring forum (the AI Safety Summit series) and a shared vocabulary for discussing frontier AI risk across otherwise rival governments, including the US and China at the same table.

What followed

Subsequent summits (Seoul 2024, Paris 2025) produced the International AI Safety Report — a scientific consensus document led by Yoshua Bengio and roughly 100 experts from 30 countries — and additional voluntary commitments from labs on safety testing and incident reporting. The EU AI Act, developed in parallel, is the most binding regulatory instrument to date, with real financial penalties, but it only has direct jurisdiction over the EU market.

The honest assessment

As of this writing, no binding, verifiable, multilateral AI safety treaty exists. What exists is a shared diagnostic vocabulary, a recurring summit process, and one binding regional regulation. Whether that scaffolding hardens into real verification and enforcement, the way nuclear arms-control regimes eventually did, remains an open and consequential question.