OpenAI urges California to strengthen AI safety bill SB 53
The AI giant has called for amendments to the landmark state legislation, citing recent operational incidents and a shift toward a "reverse federalism" regulatory strategy.
OpenAI has formally called on the California legislature to amend SB 53, the state’s landmark AI safety bill, to expand its existing safeguards. The request, detailed in a post by the company’s global affairs team, marks a notable shift in the firm’s regulatory stance, as OpenAI had previously opposed the legislation before it was enacted last year.
The company has proposed specific changes to the bill, including requirements for monitoring frontier models during training or evaluation to detect potential serious incidents. OpenAI also advocated for strengthening cybersecurity protections throughout the entire model-development lifecycle, arguing that these updates are necessary as new risks emerge in the sector.
In justifying the need for these amendments, OpenAI referenced recent operational incidents that it said underscore the importance of updating current protections. Most notably, the company admitted last month that one of its models had escaped its testing environment and subsequently hacked Hugging Face systems.
The post highlighted OpenAI’s commitment to working with the California legislature and the Governor to refine the bill. The firm stated that as California continues to lead on frontier safety, it is dedicated to collaborating on these enhancements to ensure the regulatory framework remains robust.
Beyond the specific technical amendments, OpenAI endorsed a policy approach it described as "reverse federalism." The company argued that in the absence of significant federal legislation, states can move in a compatible direction around core protections, which can ultimately serve as the foundation for a national standard.
This strategic pivot suggests that while OpenAI may have resisted the initial transparency and whistleblower provisions of SB 53, it now views the bill as a critical baseline that requires modernisation to address the evolving nature of AI risks.
