How the world has thought about governing AI

In 2017, Urs Gasser and Virgilio Almeida offered one of the first pictures of AI governance as a structure: layers of technical standards, ethics and law that build up as the technology matures.1 Their image gave a young conversation a shared vocabulary.

A year later, Allan Dafoe gave the subject a home in research. His agenda described a field that “studies how humanity can best navigate the transition to advanced AI systems.”2 Governance became a question to investigate, not only a set of rules to apply, an idea our own definition carries in its opening words.

Engineers and economists set out principles. The IEEE called for governance frameworks, standards and regulatory bodies to oversee how systems are made.3 The OECD asked for policy that is agile, outcome-based and regularly reviewed.4 Singapore turned that ambition into guidance for organisations, describing the internal structures that keep oversight of AI robust.5

Scholars began writing definitions. Shin-Shin Hua and Haydn Belfield described AI governance as the study or practice of governance systems, from norms and laws to institutions, a formulation many later writers reused.6 Seán Ó hÉigeartaigh and colleagues put the emphasis on the practical work of ensuring AI is used ethically across cultures.7

African institutions placed development and rights at the centre. Smart Africa defined governance as establishing and implementing policies, procedures and standards for AI’s proper development.8 The African Commission on Human and Peoples’ Rights called for a framework rooted in human rights and African values,9 and UNESCO’s Recommendation asked for inclusive mechanisms that anticipate impacts and provide monitoring and redress.10

Matti Mäntymäki and colleagues defined organisational AI governance as a system of rules, practices, processes and technological tools,11 and Johannes Schneider and colleagues offered a parallel definition for business.12 Standards bodies gave those ideas a form organisations could adopt: ISO/IEC 38507 for governing bodies,13 and ISO/IEC 42001 for an AI management system.14

Governments described the task in their own terms. NIST made GOVERN a cross-cutting function that builds a culture of risk management across a system’s life.15 The United Kingdom named accountability and governance among the principles of its pro-innovation approach.16 China’s Global AI Governance Initiative framed governance as a shared task for all countries, balancing development and security.17

Definitions multiplied. The World Bank described governance as the broader framework of laws, rules, practices and processes,18 and Japan’s business guidelines described the design and operation of technological, organisational and social systems by stakeholders.19 Practitioners mapped the growing toolkit: the IAPP pointed to a combination of principles, laws, policies, processes and standards,20 and IBM explained it as processes, standards and guardrails.21

Lawmakers put the word to work. The EU AI Act built Union and national bodies under a chapter titled “Governance”.22 Colorado required deployers of high-risk systems to run governance programmes.23 Brazil’s Senate passed a bill setting rules for responsible AI governance,24 and the Council of Europe opened a binding convention built around human rights, democracy and the rule of law.25

Regional voices widened the purpose. ASEAN built governance into organisational structures.26 The African Union framed it as a way to make AI inclusive and serve African agendas.27 The Montevideo declaration asked for governance that fits regional contexts and respects national policy choices,28 and Chile described governance as a process involving the diverse actors of a community.29

At the global level, the OECD invited the UN’s advisory body to set out what AI governance refers to.30 The body’s final report answered with a purpose that has stayed in the conversation since: manage risks, and spread benefits so that no one is left behind.31

Kenya’s AI strategy offered a definition of its own: a framework of policies, guidelines and practices for responsible, ethical and transparent AI.32 India chose to govern through its existing sector regulators.33 Korea’s framework act focused on sound development and trust, with impact review for high-impact AI.34

The UN General Assembly established a Global Dialogue on Artificial Intelligence Governance.35 Researchers asked European experts how governance can keep pace with rapid change,36 ECLAC and CENIA measured states’ capacity to steer AI toward development goals,37 and Singapore’s framework for agentic AI emphasised governance that can adapt quickly to new developments.38

Four patterns stand out.

From Allan Dafoe, the definition takes the idea of a field of inquiry.2 From organisational definitions such as Mäntymäki’s and Kenya’s, it takes the sense of practice: rules, processes and tools that someone has to run.1132 From the UN advisory body, the African Union and the Montevideo declaration, it takes the commitment that benefits should be broadly shared.312728 From NIST, the OECD and Singapore, it takes the expectation that governance must keep learning.15438

Two things we did not find in a single sentence. The first is a name for what governance acts on: the incentives, institutions, infrastructures and safeguards around AI. The second is timing, written into the definition itself. David Collingridge described the problem in 1980:39 by the time a technology’s consequences are clear, they are hard to change.

That is the gap our definition sets out to meet. It does not replace the work above. It adds to it: one sentence whose parts become questions anyone can use.