SB 833 mandates human review, annual assessments, and adverse event reporting
California lawmakers have introduced Senate Bill 833 (SB 833), aimed at strengthening oversight of AI and automated decision systems used in critical infrastructure. The legislation requires state agencies that deploy AI in sectors like energy, transportation, emergency services, and communications to implement real-time human monitoring, annual risk assessments, and strict adverse event reporting protocols.
The bill, introduced by Senator McNerney and recently amended in both the Senate and Assembly, responds to recommendations from California’s Joint Policy Working Group on AI Frontier Models and Executive Order N-12-23, which outlined the state’s governance priorities for AI.
Under SB 833, any state agency operating critical infrastructure must establish a human oversight mechanism to:
The California Department of Technology will provide specialized training in AI safety protocols and risk management for designated oversight personnel.
Annual AI system assessments will be required, evaluating:
Summaries of these findings must be submitted to the Department of Technology.
SB 833 introduces one of the first state-level AI adverse event reporting systems in the U.S. Entities that use or operate AI in critical infrastructure must report events involving:
Depending on severity, entities must report incidents within 4 to 72 hours, with lesser events requiring reporting within 14 days. A civil penalty of $500 per week applies to missed deadlines. Reports must detail the AI system used, incident description, training data sources, and mitigation steps.
The Office of Emergency Services will manage reporting, share aggregated insights, and protect sensitive or exempt information from public disclosure.
California’s move follows a wave of national concern about the safety, transparency, and accountability of advanced AI systems. The bill echoes global efforts to regulate high-risk AI applications, particularly in sectors where malfunctions could trigger physical or economic harm.
As of mid-2025, no U.S. federal law mandates adverse event reporting for AI, making California’s proposed framework a potential model for other states and federal agencies.
