AI Safety Institutes and the international AI safety summits

Short answer: AI Safety Institutes (AISIs) are government bodies that study and test the safety of advanced ("frontier") AI models. They emerged in the slipstream of the international AI safety summits, starting with the summit at Bletchley Park (2023). Together they form an international network that shares knowledge and testing methods. Importantly, this is a complementary, largely non-binding layer alongside binding legislation such as the EU AI Act, not a replacement for it.

What an AI Safety Institute does

An AISI focuses on the most capable models. It develops evaluations and tests for risks such as misuse (for example cyber or bio risks), loss of control and societal harm. It builds technical expertise within government, advises policymakers and works with model providers and other institutes. The emphasis is on measuring and understanding — science and testing capacity — more than on enforcement.

The international summit series

The series began with the AI Safety Summit at Bletchley Park in 2023, where a large number of countries signed a joint declaration on the risks of frontier AI. Later gatherings continued it, each with its own emphasis: from safety and international cooperation towards a stronger focus on innovation, opportunity and broad access. The shift in tone — from risk management towards growth — mirrors the wider policy debate on balancing safety and competitiveness.

A network, not a global regulator

The institutes work together in an international network to align testing methods and share findings, so that a model is not assessed from scratch in every country. But an AISI is not a regulator with the power to fine: it can evaluate a model and flag concerns, not ban it. Binding obligations remain the domain of national and regional legislators. The EU approach for general-purpose AI references safety testing, but the enforceable obligations sit in the AI Act itself.

Place on the world map

AISIs form the cooperation layer above varied national regimes. Whether a country regulates with binding law like the EU and South Korea, or voluntarily like Singapore, the Gulf region and Canada after AIDA stalled: the safety network provides a shared technical language. See also our comparison of international AI governance and the state of AI regulation.

What this means

• Follow the evaluations, not just the laws: AISI methods are becoming the de facto reference for testing advanced models.
• Expect no global regulator: enforcement stays national/regional; the network shares knowledge, not sanctions.
• Build on the shared grammar: testing approaches align with the NIST framework and the EU's risk-based logic.