California requires frontier AI makers to publish safety plans

In short: California has a new law that tells makers of the most powerful AI systems to publicly explain how they plan to keep those systems safe.

What happened

Governor Gavin Newsom signed SB 53, also called the Transparency in Frontier Artificial Intelligence Act, on September 29, 2025. The law applies to developers of large “frontier” AI models that are available in California.

Starting January 1, 2026, covered companies must publish a safety and transparency framework on their websites. Think of it like a public safety manual, where a company explains what risks it has identified and what steps it says it will take to reduce them.

The law focuses on very capable, general-purpose AI models that could potentially be used to cause large-scale harm, such as major property damage, injury, or death. It also says these public frameworks should be based on national or international standards and common industry practices.

SB 53 also creates CalCompute, a public computing consortium run through the state’s Government Operations Agency. In plain terms, this is a state-backed way to share access to computing power, which is the expensive hardware needed to train and test AI systems.

The bill does not set direct safety requirements for how companies must build their models, and it does not create new legal liability for developers. It follows Newsom’s 2024 veto of a stricter AI bill, SB 1047, which he said could slow innovation.

Why it matters

For the public, the main change is more visibility. Companies that offer the most powerful AI in California will need to show their risk plans in writing, which can make it easier for customers, researchers, and regulators to compare claims and spot gaps.

Source: NYTimes