137AI > AI Governance

AI Governance

AI governance is the layer of law, regulation, and policy that defines what AI agents are allowed to do, who is responsible when something goes wrong, and which institutions have authority to enforce the rules. The autonomous and ambient agents that have come into the field over the past several years sit in unsettled territory across nearly every governance dimension. Vehicle safety law assumed a human driver. Industrial machinery law assumed a stationary or fixed-path machine. Personal data law was written for transactions and accounts, not for continuous ambient capture. Criminal law was written for human actors. Each of these regimes is being adapted to handle autonomous and ambient agents, and the adaptation is uneven across jurisdictions, sectors, and agent categories. The work covered here surfaces what exists, what is changing, and where the gaps remain consequential.

Regulatory Frameworks

The regulatory landscape for AI agents is a patchwork of horizontal AI legislation and sector-specific rules that predate AI entirely. The EU AI Act is the most comprehensive horizontal framework, classifying AI systems by risk tier and imposing conformity assessment, transparency, and oversight obligations that scale with the classification. In the United States, the framework is sectoral: NHTSA and state DMVs regulate autonomous vehicles, FMCSA regulates autonomous trucks, FAA regulates drones and uncrewed aerial systems, FDA regulates AI as a medical device, and state robotics and AI laws fill some of the gaps with consumer protection and recording-consent requirements. Asia-Pacific jurisdictions have their own approaches, with China's algorithmic recommendation regulations and generative AI rules among the most developed. Regulatory Frameworks covers the EU AI Act and its conformity assessment regime, the US sectoral landscape including NHTSA, FMCSA, FAA, and FDA, the UN-R 155 cybersecurity regulation for connected vehicles, and the emerging state-level robotics and AI legislation across major US jurisdictions.

Liability and Product Law

When an autonomous agent causes harm, the chain of liability runs through the operator, the manufacturer, the software vendors that supplied components of the agent's stack, and in some cases the user. Conventional product liability law allocates responsibility based on a defect in design, manufacture, or warning. AI agents complicate the analysis because the behavior at issue often emerged from training on data the manufacturer did not author, in response to inputs the manufacturer did not anticipate, executed by a model the manufacturer may not have fully understood. Strict liability regimes that hold operators or manufacturers responsible regardless of fault are being proposed and adopted unevenly. Insurance markets are still developing products that handle autonomous agent exposure, with no settled approach for humanoid criminal exploitation, fleet-scale cyber-physical compromise, or third-party harm from coordinated multi-agent misuse. Liability & Product Law covers the product liability framework, the strict liability proposals, the insurance market response, and the contractual allocation patterns operators are using while the law catches up.

Criminal Law and Unsettled Categories

Criminal law was written for human actors with the capacity to form intent. Autonomous agents introduce categories that the existing statutes do not cleanly address. When a humanoid under operator direction removes merchandise from a store, the applicable charge is unsettled across jurisdictions: robbery presumes force or threat by a human agent, theft presumes a human taking without force, and the humanoid's physical capability sits somewhere between the two. When a robotaxi is used to move contraband, the operator's culpability is clear but the vehicle itself is treated inconsistently across jurisdictions, sometimes as a tool, sometimes as a co-defendant in civil forfeiture, sometimes as evidence. When a software agent makes an unauthorized purchase under prompt injection, the user, the operator, the agent developer, and the upstream content publisher all have arguable roles, and case law assigning responsibility is sparse. Criminal Law & Unsettled Categories covers the open questions across property crime, fraud, harassment, and weapons law as they apply to autonomous and ambient agents.

Personal Data and Surveillance Law

Personal data law evolved around the model of discrete transactions and identifiable records held by data controllers with clear processing purposes. Personal and ambient agents do not fit that model cleanly. A smart glasses recording captures audio and video of bystanders who never consented to processing. A robotaxi cabin recording accumulates rider data continuously across trips. A smart home assistant builds a long-term profile of household conversations. A cabin AI in a connected vehicle watches drivers and passengers across years of ownership. GDPR addresses some of this with its consent, minimization, and retention requirements, but enforcement against ambient capture has been uneven. US state privacy laws are catching up unevenly. Recording-consent laws designed for human-held recording devices apply imperfectly to always-on AI agents. Biometric privacy laws in Illinois, Texas, and elsewhere reach some of the capture surface but not all. Personal Data & Surveillance Law covers the GDPR application, the US state privacy landscape, recording-consent law, biometric privacy law, and the gaps in coverage for continuous ambient capture.

International Coordination

Autonomous and ambient agents cross borders. A robotaxi platform deployed in one country runs on training data sourced globally and models updated from a cloud provider in a third jurisdiction. A humanoid manufacturer in one country ships hardware to facilities in dozens of others. A software agent platform serves users across most of the world from a backend in a single region. Coordination across jurisdictions matters because divergent rules create operational complexity and regulatory arbitrage, and because some risks, particularly cyber-physical compromise of fleet infrastructure and supply-chain-of-updates attacks, are inherently cross-border. The G7 AI principles, the Council of Europe AI Convention, OECD AI guidelines, and bilateral arrangements between major jurisdictions are the main coordination vehicles, with varying degrees of binding effect. International Coordination covers the major multilateral instruments, the bilateral arrangements, and the practical interoperability of conformity assessment and incident reporting across jurisdictions.

Standards Bodies

Standards bodies sit between regulators and operators, developing the technical specifications that regulators reference and operators implement. ISO and IEC have produced standards for AI management systems, AI risk management, and AI quality. IEEE publishes ethical and technical standards for autonomous systems. SAE has long-standing standards for vehicle systems that are being extended to autonomous variants. UL has published UL 4600 as a safety case methodology for autonomous products. NIST develops voluntary frameworks including the AI Risk Management Framework and supporting profiles. The standards landscape is voluminous and uneven in adoption, with some standards approaching de facto requirement status through procurement and contractual references, and others remaining largely aspirational. Standards Bodies covers ISO/IEC, IEEE, SAE, UL, NIST, and the sectoral standards organizations relevant to autonomous and ambient agents.

Incident Reporting and Registries

Incident reporting is the mechanism by which the field learns from what has gone wrong. Aviation reporting has been the model for decades: mandatory reporting of accidents and incidents, with independent investigation, public findings, and feedback into design and operating practice. Autonomous vehicle reporting is partial: NHTSA collects standing general orders on autonomous vehicle incidents, but the data is incomplete and not always public. AI incident reporting more broadly is still informal, with voluntary databases maintained by research organizations and incident reporting requirements emerging in EU AI Act high-risk obligations. The registries that exist are uneven in scope, coverage, and verification, and the field does not yet have a canonical incident reference that journalists, researchers, policymakers, and operators can rely on. Incident Reporting & Registries covers the existing reporting regimes, the gaps, and the emerging structures for autonomous and ambient agent incident data.

Critical Infrastructure Policy Intersection

Critical infrastructure has its own substantial policy framework: PPD-21 in the United States, CISA's coordination role, sector-specific risk management agencies for the energy, water, transportation, communications, and other sectors. AI agents intersect this framework in two ways. First, AI is being deployed inside critical infrastructure operations for forecasting, predictive maintenance, optimization, and decision support, raising questions about how existing infrastructure policy handles AI components. Second, AI agents deployed in consumer and commercial environments can be turned into attack vectors against critical infrastructure through sensor compromise, telemetry tampering, and orchestration-layer attacks. The governance vacuum at this intersection is acute because AI regulators do not have jurisdiction over critical infrastructure operations, and critical infrastructure regulators do not have deep AI expertise. Critical Infrastructure Policy Intersection covers CISA's role, sector regulator engagement with AI, and the proposals for closing the governance gap at the intersection.

Related Coverage

Risks & Management | Security & Trust | Compliance & Conformity | Controls