The Pulse
California report proposes mandatory independent checks for frontier AI
A California report recommends requiring independent evaluations of catastrophic risks from frontier AI models. The proposals reflect several recommendations from the nonprofit Fathom and would need legislative action and rulemaking.

California is proposing mandatory independent assessments of catastrophic risks from frontier AI models, replacing a voluntary approach with oversight by state-licensed evaluators. Governor Gavin Newsom released the implementation roadmap on November 18, according to a statement from Fathom, a nonprofit that recommended several of the measures.
The recommendations are not themselves new requirements for developers. They set out a proposed approach to carrying out California's recently passed AI laws and call for further legislative action and rulemaking.
Independent checks
Under the proposal, large frontier developers would have to let independent verification organizations, or IVOs, assess the risks created by their models. The state's Government Operations Agency would license and continually oversee the evaluators, with the power to revoke the licenses of those that fall short.
The evaluators would report their findings to the state, which would decide what action to take. California would define acceptable risk in law, while the evaluators would develop and update the tests used to measure it.
Fathom said the report reflects its proposed legal test: a frontier model should be unlikely to increase the risk of catastrophic harm above the level that would exist without it, and its benefits should outweigh its remaining risk. The nonprofit had recommended that the state set the safety outcomes and leave technical assessments to the evaluators.
Tracking AI research
The report also recommends making assessment of automated AI research and development the evaluators' first task. Fathom had urged the state to prioritize recursive self-improvement, in which AI systems can improve their own capabilities.
In its earlier submission to California officials, the nonprofit proposed that evaluators have ongoing access to AI systems during development and internal use. It also recommended that their work address internal-use governance, pre-registered capability thresholds and verification of control measures.
Fathom suggested that the state initially assign evaluators from a pre-designated pool, with payment based on time and materials. Under that proposal, developers could choose evaluators as the pool grows, while the state would retain powers including the ability to veto a choice, require rotation and decertify an evaluator.
From law to implementation
California passed SB 813 and AB 1405 this summer, establishing a framework for independent verification organizations. A subsequent executive order accelerated the deadline for finalizing designation criteria from January 2028 to May 2027 and convened experts to consider where independent verification could play a broader role.
Fathom said the report reflects its recommendation to phase in licensing requirements as the evaluator market grows rather than wait for it to mature before imposing requirements on developers. The organization argued that a mandate would help attract capital and talent to the sector.
Fathom CEO and co-founder Andrew Freedman welcomed the report while acknowledging that implementation would bring questions about legislation, rulemaking, timing and costs. He urged state and federal governments to require transparency and independent risk assessments.