OpenAI Reverses Course on California SB 53, Now Urges Stronger AI Safety Rules

OpenAI is now calling on California to strengthen SB 53, a state AI safety bill the company previously opposed — a notable reversal that reflects how quickly the regulatory calculus around frontier AI is shifting.

Key takeaways

• OpenAI has reversed its earlier opposition to California's SB 53 AI safety bill and is now urging the state to strengthen its provisions.
• SB 53 is California legislation that would impose safety requirements on developers of the most powerful AI models.
• The reversal comes after OpenAI quietly disbanded its internal preparedness team — the group previously responsible for assessing catastrophic model risks.
• Stricter state-level safety rules could influence what models are publicly available, how they are trained, and what outputs they are permitted to generate.
• No final text of the strengthened bill has been published; the scope of any new requirements remains unconfirmed.

From opponent to advocate — what changed at OpenAI

OpenAI's prior position on SB 53 was straightforward opposition. The company, along with several other major AI developers, argued that California-specific rules would fragment the regulatory environment and create compliance burdens that disadvantaged U.S. companies relative to international competitors. That argument has not disappeared — it remains a common refrain in industry lobbying — but OpenAI is no longer making it about this bill.

According to TechCrunch, OpenAI is now actively asking California legislators to go further than the current draft. The company has not published a detailed public statement explaining the shift, so the precise reasoning is still unclear. What is observable is the timing: the reversal follows months of public criticism over OpenAI's internal safety practices, including the disbanding of its preparedness team and a high-profile sandbox breach.

For AI creators, the practical question is what stronger safety requirements would actually govern. SB 53, as drafted, targets developers of the most capable frontier models — the very systems that underpin image generators, video tools, and the large language models used for prompt assistance. Tighter rules on those developers could affect which model capabilities are released publicly, which fine-tuning approaches are permitted, and how quickly new generations of generative models reach the tools creators use daily.

What SB 53 actually covers — and what is still open

SB 53 is a California state bill that would impose safety obligations on developers of the most powerful AI models, generally defined by compute thresholds — the amount of processing power used to train them. The bill as it stood required developers to implement basic safeguards and document model risks before deployment.

OpenAI's call to strengthen those provisions likely means raising those thresholds, adding audit requirements, or expanding the categories of risk that must be assessed. None of that additional language is finalized. Until a revised draft is published, the concrete effect on any specific model or tool remains speculative.

What is already clear is that California has outsized influence here. Because most major AI developers are headquartered in the state, California law effectively sets a national floor. A bill that passes with meaningful teeth would apply to the companies building the models that power nearly every AI image and video tool on the market.

The trust problem shaping this shift

The broader context is a trust deficit that has been building for months. Anthropic CEO Dario Amodei described the mounting resistance to AI as "fundamentally a crisis of trust" — and OpenAI's pivot on SB 53 can be read through the same lens. Voluntarily endorsing external oversight is one of the more credible signals a company can send when internal governance has come under scrutiny.

For creators who depend on frontier models for their work, the outcome of SB 53 is worth tracking. If California passes a strengthened version, it will shape the training, deployment, and capability boundaries of the models that sit behind the image generators and style catalogs they use every day. The legislative session is ongoing, and the final text — including whatever OpenAI has specifically requested — has not yet been made public.