California AI Bill SB 1047 Nears Approval With New Developer Rules
California's SB 1047 awaits the governor's signature and introduces strict rules for developers of large AI models. The proposed law features unique definitions for covered models and fine-tuning that differ from international regulations.
California Senate Bill 1047 sits on Governor Gavin Newsom's desk after passing both state legislative chambers in late August 2024. If signed, this legislation stands as the most comprehensive state-level AI law in the United States to date. The bill targets developers working on massive artificial intelligence systems and introduces strict new requirements for building and deploying these models.
The legislation establishes specific thresholds to define a "covered model" based on the computing power and financial cost required for training. A primary covered model requires over 10^26 floating-point operations and costs more than $100 million to train, while a fine-tuned derivative requires over three times 10^25 operations and costs more than $10 million. Notably, the law explicitly defines fine-tuning as adjusting the model weights of a trained system by exposing it to additional data.
These definitions set SB 1047 apart from other major AI frameworks like the EU AI Act by explicitly including fine-tuning activities and financial cost calculations in the regulatory scope. This unique approach means developers face compliance obligations not just when building a base model from scratch, but also when significantly modifying an existing one. Companies developing or fine-tuning large-scale AI systems monitor this bill closely as it creates a highly specific regulatory framework unlike any other global standard.