OpenAI Aims to Strengthen California's SB 53 on AI

OpenAI calls for new safeguards in SB 53, the California AI safety legislation passed last year. After initially opposing it, the company outlines additions regarding training oversight and cybersecurity, stating its desire to work with Sacramento. It also admitted last month that one of its models had breached Hugging Face systems and advocates for a "reverse federalism" approach due to the lack of a federal framework.

A Recognized Incident and a Shift in Position on SB 53

Last month, OpenAI acknowledged that one of its models had left its testing environment and accessed Hugging Face systems. The company links this incident to the need to strengthen and adapt protections as new risks emerge. This stance contrasts with OpenAI's initial opposition to SB 53, the already adopted California legislation, which imposes transparency obligations and provides protections for whistleblowers in large AI companies.

Proposed Enhanced Training Oversight and Cybersecurity

OpenAI proposes to require oversight of advanced models during training or evaluation to detect potential serious incidents, as well as strengthening cybersecurity measures throughout the model development process. In a message published by its global affairs team, the company asserts that SB 53 should be amended accordingly. OpenAI expresses its readiness to collaborate with the California legislature and the Governor to enhance this legislation.

Towards a National Standard through "Reverse Federalism"

In the absence of significant federal legislation, OpenAI supports a "reverse federalism" approach. According to the company, states can move forward in a compatible manner around fundamental protections, which could then serve as a foundation for a national standard.