AI Governance

AI governance refers to the full set of mechanisms through which society shapes the development and use of AI. This includes internal corporate policies, professional norms, national regulations, international agreements, technical standards, and liability frameworks.

The governance challenge is distinctive because AI systems are general-purpose technologies that develop rapidly and have effects across almost every domain. This makes domain-specific regulatory approaches (medical device rules, financial regulations) necessary but insufficient - the same underlying model can be deployed in contexts with entirely different risk profiles.

Several approaches to AI governance have emerged. Risk-based approaches, exemplified by the EU AI Act, classify AI applications by the severity of potential harm and apply proportionate requirements. High-risk applications face mandatory testing, documentation, and human oversight; general-purpose models face transparency requirements. In the US, a more fragmented approach relies on sector-specific agencies, executive orders, and voluntary commitments from developers.

Frontier AI governance - governing the most powerful models under development - has emerged as a particular focus. This includes discussions of compute thresholds (how much computational power triggers oversight requirements), pre-deployment evaluations (testing for dangerous capabilities before release), and international coordination to prevent races to the bottom on safety standards.

The technical and policy communities have increasingly converged on the importance of evaluations as a governance tool: a way to translate abstract safety concerns into measurable requirements that regulators can mandate and developers can test against.