California Passes SB 813, Mandating Independent Evaluators for AI Safety

California's SB 813, signed into law this month, establishes a framework for state-recognized independent verification organizations for AI safety. This follows the previous year's SB 53, which mandated large frontier AI developers to publish safety frameworks. Anthropic and OpenAI are now required to incorporate independent evaluators, such as METR and Redwood Research, into their safety assessments. Dario Amodei, CEO of Anthropic, and Sam Altman, CEO of OpenAI, are both navigating these new compliance requirements. Evaluators will have limited time to conduct assessments; for instance, Apollo Research has three days to test GPT-6 Astra, while METR and Redwood have one week to investigate a recent incident involving Hugging Face. John Steidley from Palisade Research emphasized the importance of transparency in the training processes of AI models, stating, "It’s extremely relevant if the AI has been trained specifically to perform well on that benchmark." Henry Papadatos of Safer AI criticized the lack of external accountability, asserting that companies cannot claim independence while maintaining flexible internal rules. This regulatory shift could significantly alter how AI companies approach safety evaluations and transparency in their operations.