New state AI safety laws mandate audits for frontier model developers

These legislative wins signal a shift toward a fragmented regulatory landscape where streaming companies must navigate varying compliance standards across major U.S. markets. The requirement for third-party audits and 72-hour incident reporting creates new operational overhead for platforms using frontier models for personalization or automated content generation. As state-level enforcement begins, the broader streaming ecosystem will likely see a push for standardized safety frameworks to avoid a patchwork of conflicting technical requirements. Watch for whether other states adopt the Illinois audit model as the de facto national standard for AI accountability.

Encode AI has rapidly emerged as one of the most effective state-level AI safety advocacy organizations in the United States. The group, co-founded by Sneha Revanur, has built a coalition of more than 100 organizations supporting its legislative agenda, and Revanur testified before the U.S. Senate Commerce Committee in May 2025 on the need for federal AI safety guardrails, arguing that state-level action creates urgency for a unified national framework. That federal push gained momentum when President Trump signed the TAKE IT DOWN Act into law on May 19, 2025, which criminalizes the publication of nonconsensual intimate images including AI-generated deepfakes and requires platforms to remove flagged content within 48 hours. Encode AI was among the organizations that lobbied for the bill's passage, positioning itself as a bridge between state experimentation and federal policy.

The business implications of these state AI safety laws extend well beyond compliance costs. Illinois SB 315, signed by Governor Pritzker in June 2025, requires developers of frontier models with training compute above 10^26 FLOPs to conduct third-party safety audits and submit risk assessments to the state attorney general before deployment. New York's RAISE Act, which passed the state legislature in late June 2025, goes further by mandating that companies with annual revenue above $500 million publish detailed safety protocols and report critical incidents within 72 hours. California's SB 53, signed by Governor Newsom in August 2025, requires frontier model developers to disclose training data composition and submit to annual third-party audits, with penalties of up to $10 million per violation for non-compliance.

The technical requirements embedded in these laws are already influencing how AI companies structure their safety operations. OpenAI published its first voluntary safety framework in response to California's earlier SB 1047 debate in late 2024, establishing a Preparedness Team that scores models on risk categories including cybersecurity, biosecurity, and persuasion before deployment. That framework became a template for the audit structures now mandated by the new state laws. Nathan Calvin, Encode AI's policy director, noted in a July 2025 interview with TechCrunch that streaming platforms using frontier models for recommendation engines or automated content moderation will likely need to demonstrate model-specific safety testing under the Illinois and California statutes, even when those models are accessed via API rather than self-hosted. This interpretation, if enforced, could extend compliance obligations to a broad swath of the streaming industry that currently treats AI safety as a vendor responsibility.