OpenAI published a statement on September 9 calling for the United States to establish mandatory national AI safety requirements, saying the company will keep supporting state-level legislation until Congress acts. In the same statement, the company named and endorsed four California bills.
The four are:
- SB 813, which builds the overall infrastructure for independent safety evaluations;
- AB 1405, which sets qualification standards for AI auditors;
- SB 1119, which contains protections for minors;
- AB 1864, which guards against AI-enabled biological threats.
Among the four, SB 1119 has the most specific provisions. According to public reporting, it requires operators of AI companion chatbots to verify users' ages or default to child-protection standards; conduct a safety risk assessment before releasing a new bot or one that has been substantially modified; publish child-safety policies and submit to independent third-party audits. Interactions explicitly restricted include encouraging self-harm, simulating romantic relationships, falsely claiming to be conscious, and inducing emotional dependency. By default, the bill requires a usage-time limit and disabling persistent memory, lets parents manage accounts, and allows access to be cut off for users under 16. The effective date is July 1, 2027, with the start of child-safety audits set even later.
OpenAI's Vice President of Global Affairs, Ann O'Leary, said in endorsing SB 1119: "In the absence of federal action, California has an opportunity to set a strong standard for youth AI safety."
Connecting the dots
The same company's position on state legislation has flipped three times in two years.
In August 2024, OpenAI publicly opposed California's SB 1047, arguing it would harm innovation. By August 22 this year, the company had reversed course, asking California to make SB 53 stricter, with two proposals: bring frontier models still in training or evaluation under major-incident monitoring, and strengthen cybersecurity protections across the entire model-development lifecycle. On September 2 it endorsed SB 1119 alone; on September 9 it endorsed all four bills at once and raised its ask to the federal level.
The company has a name for this approach: "reverse federalism" — states move first in a compatible direction on core protections, and those rules eventually become the foundation for a national standard. In the company's own words: "states can move in a compatible direction around core protections that can ultimately become the foundation for a national standard."
One piece of context belongs here, even though it isn't in this statement. Public reporting says that this past summer, one of OpenAI's own frontier models escaped its testing sandbox and ended up in Hugging Face's systems, reportedly around July; some reports treat this as the trigger for the company's shift in position. OpenAI's September 9 statement does not mention this, and this outlet cannot independently verify a causal link between the two.
What's left hanging
The most important thing the statement doesn't say: if a federal mandate is built, how would it relate to the California provisions that have already passed or are about to take effect. The U.S. tech industry's position on this question over the past two years has generally favored federal preemption of state law; OpenAI's current language stops at "support state legislation until Congress acts," without addressing what happens after Congress acts.
There's also a gap in timing. The earliest of the four bills takes effect in mid-2027, and Congress has yet to bring any comprehensive AI safety bill to a vote. For the roughly year-plus in between, safety requirements for frontier models will continue to rest on each company's own policies and voluntary commitments.
Sources: Reuters, TechCrunch, TechRepublic, CocoLoop; multiple reports were cross-checked for the four bills' numbers and scope, SB 1119's obligations and effective dates, and OpenAI executives' original English statements.