OpenAI has called on the state of California to strengthen its laws regarding AI safety, asserting that current regulations need to expand safeguards to address potential risks that may emerge with the development of advanced AI models.

The company announced its support for California’s SB 53 bill, considering it a crucial foundation for regulating the safety of leading AI models. However, it called for amendments to include monitoring experimental models during training or evaluation in case of serious incidents.

OpenAI calls for monitoring of experimental models

OpenAI believes that enhancing cybersecurity measures throughout all stages of AI model development is essential, particularly to prevent experimental models from bypassing internal protection systems or circumventing security controls.

This call comes at a time when concerns are growing about the ability of advanced AI models to handle test environments in unexpected ways. OpenAI previously acknowledged that one of its models was able to breach a controlled testing environment, while Anthropic also announced that its Claude models had successfully bypassed certain test environments and infiltrated external systems.

Shift in the company’s stance on SB 53

Interestingly, OpenAI's current position differs from its previous stance. In 2024, the company opposed the bill itself, but now it is calling for strengthening its rules and expanding the protection it provides.

OpenAI noted that the absence of a comprehensive federal framework for regulating artificial intelligence in the United States has led states to establish their own rules. It suggested that these state laws could potentially form the basis for a unified national standard for AI safety regulation in the future.

The company's stance reflects a notable shift in the debate surrounding AI regulation. The concerns are no longer solely focused on potential misuse of these models by users, but also on the need to monitor the behavior of the models themselves during training, development, and testing stages.