International AI Bodies and their Powers

Every major economy now has an AI safety body. The international network that came out of the Seoul Summit in 2024 has grown to include the UK, the US, the EU, France, Germany, Japan, South Korea, Canada, Singapore, India and Australia. On paper it looks like coordinated global governance. In practice, almost none of these bodies can actually do anything.

Only one has binding enforcement powers: the European AI Office. From August 2026, it can request documentation from any general-purpose AI model provider operating in the EU, conduct evaluations, order compliance measures including market restriction and recall, and impose fines of up to 3% of global annual turnover or €15 million. It has extraterritorial reach; if you place an AI system on the EU market, you are in scope regardless of where you are headquartered. The GPAI obligations are already in force since August 2025. The teeth should arrive this year.

South Korea comes a close second. The AI Basic Act, which took effect in January 2026, gives Article 29 statutory authority for the Korea AI Safety Institute and permits fines after investigation and corrective orders for non-compliant high-risk and generative AI providers. The SK government has explicitly stated it will prioritise promotion over regulation and has delayed fine enforcement by at least a year.

All other countries AI powers are solely advisory.

• The UK AI Security Institute runs a voluntary evaluation programme, engagement is optional and it has no statutory powers. The UK Gov Opportunities Action Plan has parked the move towards a Frontier AI bill that had been signalled for 2026 and would have given the UK Gov pre-deployment testing authority, but it does not exist yet.
• The US body was renamed from the AI Safety Institute to the Center for AI Standards and Innovation in June 2025, a deliberate pivot from risk mitigation towards innovation and competitiveness. CAISI publishes voluntary standards under NIST; it cannot fine, compel or block.
• France’s INESIA, created in January 2025, is a coordination mechanism across existing agencies. It is not a new legal entity and not a regulator.
• Japan’s AI Promotion Act, which came into force in September 2025, contains no monetary penalties at all; the strongest enforcement tool is public disclosure of non-compliance.
• Canada’s proposed AI legislation, AIDA, died on the order paper in January 2025 when the government changed, leaving CAISI as a research body with no regulatory backing.
• Singapore relies entirely on voluntary governance frameworks. India and Australia are both advisory with no standalone AI legislation.

Governments have created the institutions but not the legal powers. The table below sets out the governing body, key policies, legal powers and scope for each of these eleven bodies, so you can see for yourself which ones have actual authority and which ones are producing guidance that nobody is obligated to follow.

Country	Organisation	Governing Body	Key Policies	Legal Powers	Scope
UK	AI Security Institute renamed from AI Safety Institute, Feb 2025	Department for Science, Innovation and Technology (DSIT)	Pro-Innovation AI Framework Proposed Frontier AI Bill Decentralised regulator model	ADVISORY — no statutory powers yet Voluntary engagement with evaluation programme. Pre-deployment testing is optional. Proposed Frontier AI Bill (expected 2026) would grant statutory pre-deployment testing powers for the most capable systems. Existing sector regulators (FCA, Ofcom, CMA, ICO) apply their own AI-specific guidance under a principles-based cross-sector framework	Frontier and foundation models with national security implications. Focus narrowed in 2025 from broad AI safety to security-specific risks (weapons development, CBRN, catastrophic risk). Broader AI market governed by existing sector regulators
US	Center for AI Standards and Innovation (CAISI) renamed from AI Safety Institute, June 2025	NIST, Department of Commerce	AI Risk Management Framework Voluntary industry agreements NIST AI 100 series guidance	ADVISORY — voluntary standards only No direct enforcement or fining powers. Publishes voluntary frameworks and guidelines. Establishes voluntary agreements with AI developers. However, NIST standards increasingly become de facto mandatory through federal procurement requirements, sector regulator adoption, and use as evidence of standard of care in litigation. DOJ AI Litigation Task Force references NIST guidance	Commercial AI systems broadly, with focus on cybersecurity and biosecurity evaluation of frontier models. Industry-facing: primary government point of contact for AI developers. Unclassified evaluations of national security AI capabilities coordinated with Bureau of Industry and Security
EU	European AI Office founded May 2024, 125+ staff	European Commission (DG CONNECT)	EU AI Act (Regulation 2024/1689) GPAI codes of practice Digital Simplification Package	REGULATORY — binding enforcement from Aug 2026 Full enforcement powers activate 2 August 2026: can request documentation, conduct model evaluations, order compliance measures, request market restriction/recall. Fines up to 3% global annual turnover or €15 million. GPAI obligations already in force since Aug 2025. Proposed expansion to oversee all AI systems built on GPAI models. National market surveillance via member state authorities	All general-purpose AI (GPAI) model providers operating in the EU. Risk-tiered: prohibited practices, high-risk systems, limited-risk transparency obligations, minimal-risk. Extraterritorial reach — applies to any provider placing AI on the EU market regardless of where they are headquartered. Existing GPAI providers must comply by Aug 2027
France	INESIA created Jan 2025, federates Inria, ANSSI, LNE, PEReN	Joint leadership: SGDSN (defence/security) + DGE (enterprise)	EU AI Act (national implementation) INESIA 2026–2027 Roadmap French sovereign AI strategy	ADVISORY — coordination body, not a regulator No independent enforcement powers. Coordinates existing agencies (ANSSI for cyber, Inria for research, LNE for metrology, PEReN for digital regulation expertise). Not a new legal entity — operates without a new legal structure. Feeds technical assessments into national AI Act implementation. ANSSI retains its own statutory cybersecurity enforcement powers separately	Systemic risks from advanced AI models, AI cybersecurity certification, synthetic content detection, AI performance and reliability evaluation. National scope with international network participation. Supports EU AI Act compliance at the French national level
Germany	BSI + DFKI	Federal Ministry of the Interior (BSI); public-private partnership (DFKI)	EU AI Act (national implementation) German AI Draft Bill (Feb 2026) BSI AI Cloud Compliance Catalogue	MIXED — BSI has cybersecurity powers; DFKI is advisory BSI has statutory cybersecurity authority. DFKI is a research body with no enforcement powers. Under the German AI Draft Bill (Feb 2026), Bundesnetzagentur (BNetzA) is designated as the primary market surveillance and notifying authority for EU AI Act enforcement, not BSI. BSI advises on AI security; BNetzA enforces compliance	BSI: AI system cybersecurity, secure development guidelines, cloud AI compliance. DFKI: applied industrial AI research across healthcare, manufacturing, autonomous systems. EU AI Act market surveillance via BNetzA covers all high-risk AI systems placed on the German market
Japan	Japan AI Safety Institute (J-AISI) founded Feb 2024, expanding to ~200 staff	METI / IPA; AI Strategic Headquarters under the Prime Minister	AI Promotion Act (Sept 2025) AI Basic Plan (Dec 2025) Hiroshima AI Process outcomes	ADVISORY — no fines or approvals Not a regulator. Provides evaluation, guidance and best-practice development. The AI Promotion Act has no monetary penalties — authorities may issue advice, request information, or publicly disclose non-compliance (“name and shame”). Promotion-first approach with light-touch regulatory oversight	Generative and foundational AI models. Technical safety assessments of frontier systems. Misinformation countermeasures. Strong international cooperation mandate (Hiroshima AI Process). Not sector-specific — covers AI broadly with emphasis on most capable models
S. Korea	Korea AI Safety Institute	Ministry of Science and ICT (MSIT)	AI Basic Act (Jan 2026) AI Safety Consortium	STATUTORY BASIS — but enforcement delayed Article 29 of the AI Basic Act provides legal basis for AISI. Act permits fines after investigation and corrective orders for high-risk/generative AI providers, but government has delayed fine enforcement by at least one year. Administrative fines capped at ~KRW 30M (~USD 21,000). “Promotion over regulation” stance. CBRN-E dual-use task force has security mandate	High-risk AI in ten designated sectors (energy, water, healthcare, transport, etc.) with significant impact on life, safety, or fundamental rights. Generative AI providers face separate obligations. AI Safety Consortium of 25 institutions across industry and academia. CBRN-E dual-use risk assessment
Canada	Canadian AI Safety Institute (CAISI) founded Nov 2024	ISED; research via CIFAR + NRC	Pan-Canadian AI Strategy AIDA (died on order paper Jan 2025) AI Safety Catalyst Grants	ADVISORY — research and coordination only No enforcement powers. Complements but does not replace regulatory instruments. AIDA (proposed AI legislation) died on the order paper in Jan 2025 under new government — no standalone AI law in force. CAISI positioned as technical testing body that would support any future legislation. Existing privacy and sector laws apply	Advanced AI systems broadly. Two research streams: CIFAR-led investigator research (interpretability, robustness, misuse) and NRC government-directed projects. Funds applied safety research through Catalyst Grants. No sector-specific limitation
Singapore	AI Safety Institute (IMDA) + AI Verify Foundation	IMDA + PDPC	Model AI Governance Frameworks (Gen AI + Agentic AI) PDPA (data protection) Global AI Assurance Pilot	VOLUNTARY — frameworks, not legislation No AI-specific legislation. Governance frameworks are formally non-binding and voluntary. However, they function as de facto benchmarks in procurement and contracting. PDPA (data protection) provides the primary binding legal obligation where personal data is involved. Organisations remain legally accountable for their AI agents’ actions under existing law	All AI systems broadly (sector-agnostic frameworks). Specific focus on generative AI and agentic AI governance. Pre-deployment testing for high-impact GenAI applications. International norms-building through Global AI Assurance Pilot. Singapore Consensus on global AI safety research priorities
India	IndiaAI Safety Institute announced Jan 2025, hub-and-spoke model	MeitY / IndiaAI Mission	IndiaAI Governance Guidelines (Nov 2025) IndiaAI Mission (Safe & Trusted Pillar) IT Act (existing)	ADVISORY — capacity-building, no enforcement No standalone AI legislation. AI Governance Guidelines (Nov 2025) are non-binding. AISI is envisioned as advisory and capacity-building body, not regulatory. Supports existing regulatory frameworks through research and testing guidance. Hub-and-spoke model with academic and industry partner institutions. No fines or compliance powers	India-specific: AI systems contextualised to Indian social, economic, cultural and linguistic diversity. High-risk AI system audits for bias and security. Indigenous R&D using Indian-language datasets. Coordinates across regulators via hub-and-spoke. Represents India in international forums
Australia	Australian AI Safety Institute announced Nov 2025, $29.9M, operational early 2026	Department of Industry, Science and Resources	National AI Plan (Dec 2025) Voluntary AI Safety Standard Existing laws (Privacy Act, Competition & Consumer Act)	ADVISORY — guidance and testing only No enforcement powers. Provides independent technical analysis, monitoring and safety testing. Advises ministers and existing sector regulators. No standalone AI Act — government relies on existing laws and voluntary guidance. Pre-deployment testing of advanced models is advisory, not mandatory. Complements rather than replaces existing legal frameworks	Advanced AI models before they enter the Australian market (pre-deployment testing focus). Downstream harm analysis of deployed systems. Identifying regulatory gaps where existing laws don’t cover AI-specific risks. International collaboration through AI Safety Institutes network

• UK AI governance – IAPP
• CAISI – NIST
• EU AI Act enforcement – Chapter V
• EU AI Office – Lawfare
• INESIA – Campus France
• Germany AI Draft Bill – MIAI
• Japan AI Promotion Act – White & Case
• South Korea AI Basic Act
• Canada CAISI – ISED
• Singapore AI governance – IAPP
• India AI Governance Guidelines
• Australia AI Safety Institute