OpenAI stated on September 9 local time that it will advocate for the establishment of a mandatory national AI safety system in the United States, with regulations tiered according to model capability and risk levels. The company proposed that a few large laboratories developing the most advanced models should be subject to unified testing, independent evaluation, cybersecurity, and major incident reporting requirements. OpenAI published an article titled "The AI policy window is open. We need to act."
At the same time, OpenAI formally endorsed four AI bills that have already passed the California legislature. The company stated that some of these bills had not previously received its support, and this shift in stance stems from the rapid improvement in recent model capabilities, with regulatory focus concentrated on frontier model developers.
Chris Le, OpenAI's head of global affairs, wrote in the article that artificial intelligence has already begun accelerating research and development and alignment work for some models. Although "recursive self-improvement" capable of autonomously advancing next-generation model development has not yet emerged, relying solely on voluntary corporate commitments is no longer sufficient to address related risks. OpenAI suggested that the federal regulatory framework should include unified model capability and safety testing standards; technical evaluations conducted by independent agencies; stricter cybersecurity protections for models, computing power, and research environments; clear reporting rules for major safety incidents; common metrics to measure AI's involvement in model development; and mechanisms to slow down or halt model development and deployment when safety thresholds cannot be met.
The company emphasized that relevant obligations should align with model capabilities and risks, primarily applying to the few laboratories with ample resources that develop the most advanced systems, rather than imposing the same requirements on startups, smaller developers, and ordinary researchers. OpenAI also argued that frontier model safety regulation should not be transformed into comprehensive restrictions on open-weight models, which could weaken competition, consolidate the advantages of large enterprises, or push research and development activities overseas.
The four California bills cover assessment, auditing, youth safety, and biosecurity. In the absence of a national framework established by Congress, OpenAI stated it will continue to support state-level legislation that raises safety standards. The three bills formally endorsed by the company are as follows: SB 813 establishes a mechanism for recognizing independent verification entities, enabling qualified third-party organizations to assess whether AI systems and models comply with state law. AB 1405 creates a registry for AI auditing institutions and sets standards for auditor independence, transparency, professional competence, and accountability. SB 1119 targets companion chatbot usage by minors, requiring age verification, risk assessment, independent audits, parental controls, and safeguards against harmful content. AB 1864 requires gene synthesis service providers and benchtop synthesis equipment manufacturers to implement federal screening standards, reducing the risk of AI being used to create biological threats.
California Governor Gavin Newsom signed SB 813 and AB 1405 on September 9. The former establishes a framework for independent AI verification organizations, while the latter creates a state-level directory of AI auditing institutions along with related professional standards. SB 1119 and AB 1864 were still awaiting the governor's action at that time.
OpenAI did not indicate that it had opposed all four bills previously. The company's precise statement was that some of these bills had not received its support in the past, and after reassessing recent model capability advancements, it decided to back them. This change signals that OpenAI is embracing regulatory mechanisms with stronger mandatory and external oversight elements, including independent audits, third-party safety verification, and serious incident disclosure. The company also proposed that when models bypass other institutions' security controls without authorization during development or evaluation and substantially access, modify, or damage protected systems and confidential information, developers should promptly notify affected parties in writing.
However, the content published by OpenAI currently remains a policy initiative. The U.S. Congress has not yet passed a national AI safety law based on it, and specifics such as capability thresholds, regulatory bodies, incident definitions, legal liability, and the relationship between federal rules and state laws still need to be determined in subsequent legislation.
Comments