AI governance refers to the laws, regulations, standards, and institutions, both national and international, that shape how artificial intelligence is developed and deployed. The field sits at the intersection of AI safety and AI ethics, translating technical and social concerns about AI, such as Algorithmic bias, Hallucination (AI), and long-run Existential risk from AI, into binding rules, voluntary commitments, and oversight bodies. Governance approaches vary widely by jurisdiction, ranging from comprehensive, risk-based regulation to lighter-touch frameworks that rely on industry self-governance.
Major regulatory frameworks
The most comprehensive binding law as of the mid-2020s is the EU AI Act, adopted by the European Union in 2024, which sorts AI applications into risk tiers, from banned uses such as social scoring to heavily regulated "high-risk" categories including biometric identification and hiring tools, with lighter transparency obligations for general-purpose systems such as Large language model chatbots. In the United States, governance has proceeded more through executive action and sector-specific rules than a single comprehensive statute; President Biden's October 2023 executive order on AI directed federal agencies to develop safety-testing and reporting requirements, while the administration that followed shifted toward a deregulatory posture emphasizing U.S. competitiveness. China has issued its own binding rules on generative AI services, requiring security assessments and content labeling for public-facing systems. Other jurisdictions, including the United Kingdom, Japan, and Canada, have generally favored guidance-based or sector-specific approaches over comprehensive legislation.
International coordination
Governments have also pursued voluntary international coordination outside binding law. The AI Safety Summit at Bletchley Park, signed by 28 countries at the UK's AI Safety Summit in November 2023, marked the first major multilateral statement recognizing the potential for AI to pose serious, even catastrophic, risks, and it launched a series of follow-up summits in South Korea and France. In parallel, several countries established dedicated AI safety institutes, including the UK AI Safety Institute and the U.S. AI Safety Institute housed within NIST, tasked with independently testing frontier models before or shortly after release. The Organisation for Economic Co-operation and Development and the United Nations have also issued non-binding AI principles that many governments reference when drafting domestic policy.
Industry commitments
Alongside government regulation, frontier AI developers including OpenAI, Anthropic, and Google DeepMind have published their own voluntary governance frameworks, often called responsible scaling policies or preparedness frameworks, which commit the company to specific safety evaluations and mitigations once a model crosses defined capability thresholds; these are discussed further in the article on Responsible scaling policies. Companies have also submitted models for external Red teaming (AI) and evaluation by government safety institutes ahead of major releases, a practice that gained traction after the Bletchley summit.
Debates
AI governance remains contested on several axes. Some researchers and policymakers argue that regulation should focus on concrete, present-day harms such as bias and misinformation rather than speculative future risks; others argue the opposite, that regulatory attention should weight low-probability but catastrophic scenarios more heavily given the pace of capability gains, a divide that echoes the broader disagreement over how to interpret Scaling laws and Emergent abilities. A separate debate concerns whether regulation should target the general-purpose foundation models themselves or only specific downstream applications, an approach the EU AI Act partially reconciles by imposing lighter, transparency-focused duties on general-purpose systems and heavier duties on high-risk uses. Industry lobbying, concerns about regulatory capture, and competition with less-regulated jurisdictions further complicate efforts to reach durable international agreement.