OpenAI Calls for Mandatory AI Safety Rules: What the California Bills Actually Change

Short answer

On September 9, OpenAI publicly called for mandatory, capability-based national AI safety requirements in the United States and announced support for four California bills: SB 813, AB 1405, SB 1119, and AB 1864. This is OpenAI’s policy position, not enacted nationwide law. California’s governor’s office separately confirmed that SB 813 and AB 1405 were signed that day; its announcement did not say that SB 1119 or AB 1864 had been signed. Builders should treat this as a regulatory-direction signal, not as a current legal obligation without checking the applicable text and date.

What happened

In a September 9 post, OpenAI Chief Global Affairs Officer Chris Lehane said the company would push for national AI safety requirements that scale with model capabilities. Until Congress acts, OpenAI said it would continue supporting state legislation. The post also called for industry standards and compatible international approaches to measuring capabilities, managing risk, preserving human control, and deciding when development should slow or stop.

OpenAI listed four California bills it supports:

• SB 813: infrastructure for independent AI safety assessments;
• AB 1405: registration, independence, transparency, and accountability requirements for AI auditors;
• SB 1119: age assurance, risk assessments, independent audits, parental controls, and safeguards for minors using companion chatbots;
• AB 1864: safeguards for gene-synthesis providers and benchtop synthesis equipment against AI-enabled biological threats.

California’s governor’s office confirmed on September 9 that SB 813 establishes a framework for independent verification organizations and AB 1405 creates registration and standards for AI auditors. This article does not turn OpenAI’s support for the other two bills into a claim that they were signed or are already in force.

What it means for AI products and agent teams

1. Treat independent assessment as a product process, not launch PR. Record the model version, tool permissions, test scenario, failure cases, remediation version, and retest result. Do not present internal testing as an independent audit.
2. Prepare incident records for high-capability features. At minimum, log prompts, model versions, tool calls, privilege changes, external-system access, human intervention, affected parties, and notification times. If unauthorized access occurs, assess notification duties against the contract, jurisdiction, and applicable rules.
3. Separate law from policy advocacy. Verify a bill’s text, signature date, effective date, covered entities, and exemptions separately. Do not promise a customer “compliance” based on a headline.
4. Use capability tiers and least privilege. For features that can access the internet, execute code, publish packages, handle minors’ data, or touch biosecurity-sensitive workflows, use approvals, isolation, egress controls, and revocable credentials.

Does this prove OpenAI is safe?

No. OpenAI also describes stronger monitoring, isolation, trajectory monitoring, and internal alignment gates, but those are the company’s statements about its own controls, not independent audit findings. A signed California bill also does not mean that every AI product has passed a safety assessment. The verifiable conclusion is narrower: policy attention is moving from voluntary company commitments toward independent assessment, auditing, incident disclosure, and controls for high-risk capabilities. The concrete obligations still depend on final legal text, implementation rules, product capabilities, and jurisdiction.

Sources and evidence boundary

• OpenAI’s official policy post: the company’s position, four California bills, and standards proposals;
• California governor’s announcement: the signing facts for SB 813 and AB 1405;
• Reuters independent coverage: cross-check of OpenAI’s call for mandatory national AI safety requirements.

This is a policy update and an operational checklist, not legal advice. Re-check the bill text, effective dates, and scope before making a product or customer decision.