AI Safety Institutes

In November 2023, the United Kingdom and the United States launched dedicated AI safety organizations following the Bletchley Park AI Safety Summit. Over the subsequent three years, national approaches diverged significantly across staffing, mandate, and statutory enforcement powers:

• The UK body became the AI Security Institute within the Department for Science, Innovation and Technology (DSIT), maintaining over 100 technical staff focused on model evaluations.
• The US body was reorganized under NIST as the Center for AI Standards and Innovation (CAISI), operating primarily on voluntary benchmarks and technical guidelines.
• The European Union established the EU AI Office under the European Commission, equipped with statutory audit, market surveillance, and fining powers that took effect on 2 August 2026.

Subsequent diplomatic forums in Seoul (May 2024) and Paris (February 2025) expanded the International Network of AI Safety Institutes to include Japan, Singapore, South Korea, Canada, France, Kenya, and Australia.

United Kingdom: Technical evaluation and multilateral reporting

Originating from the Frontier AI Taskforce, the UK AI Security Institute operates with an initial £100 million multi-year seed allocation and maintains over 100 technical researchers drawn from frontier AI laboratories and academic institutions (AISI Overview).

The institute's primary outputs comprise pre-deployment and post-deployment technical assessments:

• Empirical cybersecurity comparisons between open-weight and closed frontier models.
• Peer-reviewed research on automated persuasion in Science.
• Voluntary pre-deployment testing of commercial frontier models upon lab invitation (such as evaluations conducted on Anthropic and Moonshot AI models).
• Secretariat responsibilities for the International AI Safety Report 2026, chaired by Yoshua Bengio with contributions from nominated experts across 30 nations.

The UK body functions as an empirical evaluation laboratory rather than a statutory regulator; its evaluations inform policy but do not carry binding licensing or injunctive enforcement powers.

United States: Voluntary standards and industrial competitiveness

Following the reorganization of the initial US AISI into CAISI under the Department of Commerce (NIST CAISI), the American body focuses on voluntary risk-management standards, unclassified cybersecurity evaluations, and dual-use capability benchmarking.

CAISI's annual operating appropriations stand at approximately $15 million in FY 2026, supplemented by Technology Modernization Fund allocations (Institute for Progress Analysis). Key published outputs include comparative evaluations of international open models (such as DeepSeek V4 Pro and GLM-5.2) and technical standards on agentic AI security. CAISI does not possess statutory fining authority or pre-market clearance mandates.

European Union: Statutory market supervision and penalty mechanisms

The EU AI Office operates within the European Commission with over 125 staff across six specialized units (European Commission AI Office). Under the European AI Act, the Office possesses binding supervisory authority over General-Purpose AI (GPAI) model providers:

• Enforcement Scope: Effective 2 August 2026, the Commission gained powers to issue information requests, mandate technical file disclosures, order corrective measures, and restrict non-compliant model deployments.
• Financial Penalties: Violations of GPAI obligations carry statutory fines of up to 3% of total worldwide annual turnover or €15 million (whichever is higher) under Article 101.
• Open-Weight Criteria: While open-source models with freely accessible weights receive specific regulatory exemptions, systemic-risk thresholds and Article 50 transparency requirements remain legally binding regardless of open licensing terms.

Initial supervisory actions prioritise technical compliance dialogues and structured verification of adherence to the GPAI Code of Practice.

International safety bodies across Asia and the Commonwealth

1. Japan: Established the Japan AI Safety Institute within the Information-technology Promotion Agency (IPA) in February 2024, focusing on industrial benchmarking, ISO standards alignment, and interoperable safety guidelines.
2. Singapore: Reorganized the Digital Trust Centre at Nanyang Technological University and IMDA into an AISI in 2024, supporting technical research initiatives such as the Singapore Consensus on safety priorities.
3. India: Outlined an IndiaAI Safety Institute under the Safe & Trusted AI pillar of the IndiaAI Mission, with an initial allocation of ₹20.46 crore dedicated to academic research grants in explainability, bias mitigation, and machine unlearning (IndiaAI Mission Details).
4. South Korea & Canada: Operating distributed research consortia across academic centres and standards agencies rather than centralized administrative enforcement directorates.

Comparative governance structure