Newsom wants independent verification of AI emergency shutdowns

Gavin Newsom, governor of California, signed an executive order on September 18, 2026 that accelerates the establishment of independent AI oversight and calls for an examination of emergency shutdowns for the most advanced models. Order N-9-26 calls for studying a requirement accompanied by continuous verification of shutdown effectiveness by an independent body.

The order sets instructions and deadlines for the Government Operations Agency, the state agency responsible for preparing the recommendations. These administrative instructions take effect. The additional provisions aimed at laboratories are proposals for legislative changes, with no new mandatory shutdown mechanism established immediately.

California already has transparency rules for the major developers covered by them. SB 53 requires them to publish safety frameworks and provides for reporting the specified critical incidents to the Governor’s Office of Emergency Services, the state’s emergency services agency.

SB 813 establishes the process for designating independent verification bodies. AB 1405 creates a registry of auditors and sets rules for their independence, transparency and integrity. Neither law currently requires any laboratory to use these bodies, as CalMatters notes. The executive order accelerates the administrative implementation of this framework.

Verifiers inside laboratories

The recommendations must examine the presence of designated independent bodies within the laboratories of major frontier model developers. They would conduct periodic on-site audits and assessments. Safety frameworks, transparency reports and risk assessments could also be subject to independent verification.

For emergency shutdowns, Newsom asks for an assessment of the technical feasibility and potential effectiveness of a requirement accompanied by continuous oversight. Such verification would seek to establish that the shutdown works, a separate question from publishing a safety framework. If these proposals became mandatory and the checks were reliable, users and government bodies purchasing AI could have additional support in assessing an AI system’s shutdown capability.

In his official announcement on September 18, Newsom cites recent incidents as grounds for the executive order. The administration must also examine a broader definition of reportable incidents to include losses of control. Developers would face new obligations and oversight costs if these changes were adopted.

Recommendations due by November 16

The Government Operations Agency must submit its recommendations to the governor’s office by November 16, 2026. It will develop them with national experts, in consultation with the state’s emergency services agency. The requested work concerns the technical feasibility and potential effectiveness of the legislative changes.

By May 1, 2027, the agency must complete the requirements of section 8898.1 and publish the application conditions, procedures and criteria for verification bodies.

By December 1, 2027, it must fulfill the requirements of section 11549.82(a), concerning the auditor registry, fees and the reporting of violations, and begin the actions set out in its subsection (b). These 2027 deadlines establish the administrative framework for oversight. They do not set any date for the mandatory installation of an emergency shutdown mechanism.

The executive order leaves the shutdown’s technical architecture, testing protocol and detailed activation procedure to be defined. The authority empowered to activate it and the models, systems or deployments covered still need to be specified. Preventing abusive shutdowns and ensuring verifiers’ effective independence are also open questions.

The recommendations due by November 16 will be the next piece of information for examining the proposed scope and changes to the law. Effective protection would then require demonstrating the reliability of the shutdown.