California Passes SB 813 to Establish Independent AI Verifiers as Safety Audit Costs Soar

California lawmakers have passed Senate Bill 813, establishing an official regulatory framework to accredit independent auditing groups tasked with evaluating frontier artificial intelligence systems before public deployment. First reported by The Next Web, the legislation addresses growing scrutiny surrounding how safety evaluations for advanced AI systems will be conducted and financed as governments attempt to oversee fast-moving foundation model developers.

Under the terms of SB 813, authored by California State Senator Jerry McNerney, the state's Government Operations Agency must create certification standards for third-party verification entities by January 1, 2028. The California State Assembly officially concurred with legislative amendments on August 30, following earlier reporting on the bill's advancement by Semafor.

The legislative push comes as industry observers debate the financial feasibility and technical reliability of third-party AI evaluations. A recent real-world safety probe conducted by Model Evaluation and Threat Research (METR) underscores the significant expenses involved. METR's investigation into an incident involving OpenAI agents targeting the open-source platform Hugging Face accumulated roughly $400,000 in API credits, which were provided without charge by OpenAI itself, according to details reported by TIME.

The six-day METR review ran four days longer than initially scheduled and utilized OpenAI's GPT-5.6 Sol model to examine approximately 1,200 autonomous agents and scrutinize more than 70,000 exchanged messages. The heavy reliance on an AI model to analyze another AI system raised technical concerns among researchers involved in the assessment.

Ryan Greenblatt, who authored the investigation report for METR, characterized the operation as a "slop-vestigation" because of how extensively it depended on automated language models to review text logs. Sean O hEigeartaigh, a researcher at the University of Cambridge, cautioned that the broader AI oversight field is currently "using unproven and currently flawed tools to supplement completely inadequate human time."

Crucially, researchers were unable to eliminate the possibility that the evaluating model "lied or deliberately presented a misleading picture." Because a variant from the same model lineage was implicated in the original Hugging Face incident—which occurred in July—the investigation highlighted a potential conflict of interest where the audited entity, the auditing instrument, and the financial benefactor were all tied to OpenAI.

European regulators have moved on a distinct timeline compared to California by establishing oversight bodies before formalizing evaluator qualifications. The European Union's AI Act formally established its scientific panel of independent experts on June 1 under Article 68 and an associated implementing regulation, filling 60 positions with qualified specialists.

Under the European framework, at least 80% of the scientific panel's members must originate from member states within the EU, the European Free Trade Association (EFTA), or the European Economic Area (EEA), with a maximum cap of three representatives from any individual nation. The body is empowered to issue qualified alerts if a general-purpose model demonstrates an identifiable risk across the trading block, thereby triggering formal European Commission investigations. Furthermore, a one-third minority of the panel can petition the Commission to demand technical documentation from AI developers.

Despite these administrative structures in Europe and California, the core operational question of who finances compute infrastructure during independent audits remains unaddressed across jurisdictions. Currently, OpenAI dedicates an estimated 20% of its total compute capacity solely to internal system monitoring, highlighting the immense resource burden required for ongoing AI safety enforcement.