California SB 53 vs New York RAISE Act: frontier AI safety laws compared

What the two laws are

California's Governor Gavin Newsom signed Senate Bill 53 on September 29, 2025. His office's announcement calls it the Transparency in Frontier Artificial Intelligence Act (TFAIA).

New York's Governor Kathy Hochul signed the RAISE Act on December 19, 2025. Her office says the signed version reflects chapter amendments, bill numbers S6953B/A6453B.

This guide uses four documents: the California announcement, the New York Senate page for S6953A, the New York governor's announcement, and a press release from the RAISE Act's Senate sponsor. The S6953A page carries an earlier amendment of the bill text, not the signed S6953B version. Where the earlier text and the signing announcements differ, this guide says so.

What SB 53 requires

The only California source supplied is the governor's announcement, which gives a summary and not the statute. According to that announcement, SB 53 does five things.

On transparency, it requires large frontier developers to publish a framework on their websites.

On incident reporting, it creates a way for developers and the public to report problems to the state.

On accountability, it protects whistleblowers and sets a penalty.

It also sets up a public computing project, which the announcement calls CalCompute.

Finally, it requires yearly update recommendations.

The announcement does not state the penalty amount, the incident-reporting deadline, or the size threshold that makes a developer "large."

What the RAISE Act requires

The signing documents describe the final law. The governor's announcement says the law requires large AI developers to publish information about their safety protocols and report incidents within 72 hours.

The Senate sponsor's release lists what the safety plans must cover. These include risk assessment, mitigation, third-party review of catastrophic risk, cybersecurity against model theft, and incident response.

The same release sets a review cycle.

On enforcement, the governor's announcement says the Attorney General can sue over missing reports or false statements.

It gives the penalty range.

The law also creates a regulator. The Senate release says the new office sits in the Department of Financial Services and is funded by fees on developers. Its duties are to enforce the law, issue rules, assess fees and publish an annual report on AI safety.

Where they differ

The documents support these differences. Some come from New York officials describing their own law against California's, so read them as those officials' statements.

• Regulator. New York creates a funded oversight office in the Department of Financial Services. The California announcement describes no such office. It gives enforcement of the civil penalty to the Attorney General and incident reports to the Office of Emergency Services.
• Reporting clock. The New York Senate release says the RAISE Act requires reporting of critical safety incidents within 72 hours, "as opposed to 15 days." The California announcement does not state a deadline, so that comparison rests on the New York release alone.
• Reporting trigger. The Senate release says the RAISE Act applies when developers reasonably believe an incident occurred.
• Penalties. The Senate release says the RAISE Act raises fines to $3 million for repeat violations. The California announcement gives no figure.
• Plan detail. The Senate release says the New York law requires developers to describe safety plans in detail. California's announcement describes a framework showing how standards were incorporated.
• California-only features. CalCompute and the Department of Technology's yearly recommendations appear in the California announcement and not in the New York documents.

Assemblymember Alex Bores, a sponsor, said in the governor's announcement that the bill goes further than California's law. Governor Hochul described the two laws as related.

What the earlier RAISE text shows

The S6953A text is not the signed law, but it shows how the bill was built. The sponsor memo lists four steps for companies that spend hundreds of millions of dollars on training. Two of them were a third-party review of the safety plan and disclosure of major incidents.

The memo also limits the scope to severe harm. The memo adds "in damage or hundreds of deaths or injuries."

The memo says the act takes effect ninety days after it becomes law. The signed version was amended, and the supplied documents do not state its final effective date.

One clause in the S6953A text matters to buyers of these models. It addresses contract terms that try to push liability onto customers. The text goes on to call such provisions void as a matter of public policy. Whether the signed version keeps this clause is not shown in the supplied documents.

What the documents do not settle

• California statute text. Only the governor's summary was supplied. Size thresholds, the reporting deadline and the penalty amounts are not in it.
• Final RAISE text. The supplied bill page shows the earlier S6953A amendment. The announcements describe the S6953B version. The earlier text includes an annual independent audit, while the Senate release mentions third-party assessment of catastrophic risk. The supplied documents do not show how the final version treats audits.
• Customer duties. None of the four documents places obligations on companies that buy or deploy frontier models.
• Federal law. The S6953A sponsor memo summarizes a section saying any part preempted by federal law does not apply. The documents give no detail on how that would work.

What security leaders can take from the requirements

The documents give no guidance written for security leaders, so this section only restates what the requirements imply.

At a developer. If your company may be a "large" developer, check the final texts for the size threshold in each state. The signing documents describe the same core duties. These are a written safety framework or protocol, cybersecurity protection against model theft, incident identification and response, whistleblower protections, and reporting to a state body. The New York clock is 72 hours from determining that an incident occurred. The Senate release also says the duty applies when you reasonably believe one has occurred.

At an enterprise customer. The documents place no duties on you. They do describe public artifacts. California requires a published framework. The New York release says developers must "write, implement, publish, and comply with" safety plans. Both states route incident reports to state bodies, and the New York release says access to those reports is expanded. Check any contract language on liability against the final New York text, given the clause described above.

For both groups, the signed New York text, the California statute text and any rules from the new Department of Financial Services office are the sources that will settle the open points.