California Frontier AI Law (SB 53): A Compliance Guide for AI Companies

By Chanté Eliaszadeh | Updated June 2026

If you build AI and you have been told to “comply with California’s new AI law,” start with a correction that will save you a quarter of wasted legal spend: the bill most founders still name --- SB 1047 --- is not law. Governor Newsom vetoed it on September 29, 2024.¹ The frontier-AI statute that actually took effect in California on January 1, 2026 is a different bill, SB 53, the Transparency in Frontier Artificial Intelligence Act.² It is narrower than SB 1047 would have been, its heaviest duties attach only above a half-billion dollars in revenue, and most of the startups asking us about it are not directly covered by it at all --- while being squarely covered by two other California AI laws they have not heard of.

This guide sorts out which California AI law applies to your company, what SB 53 requires of the developers it does reach, and the training-data and watermarking statutes that catch far more companies than SB 53 does.

Key Takeaways

• SB 1047 is not law. It was vetoed in September 2024. Compliance roadmaps built on its 72-hour reporting clock, kill-switch mandate, and $10,000 penalties describe a bill that never took effect.¹
• SB 53 (the Transparency in Frontier AI Act) is the real law, effective January 1, 2026. It regulates “frontier developers” --- those training models above 10^26 computing operations --- and reserves its heaviest obligations for “large frontier developers” with annual gross revenue above $500 million.²
• Most startups fall below SB 53’s threshold. If you fine-tune or deploy third-party models through an API, you are almost certainly not the regulated “frontier developer.”
• You are more likely covered by AB 2013 and SB 942. California’s training-data transparency law (AB 2013) reaches any generative-AI system made available to Californians; the watermarking law (SB 942) was pushed to an August 2, 2026 effective date by AB 853.³⁴⁵
• The federal preemption picture is now actively contested, following a December 2025 executive order directing agencies to challenge state AI laws --- a fight that is unresolved and that you should not bank on.⁶

Is SB 1047 California’s AI Law? No --- It Was Vetoed

SB 1047 was vetoed, so it imposes no obligations, sets no effective date, and creates no enforcement authority. The California Legislature’s own record reads “Vetoed by the Governor” as of September 29, 2024.¹ In his veto message the Governor faulted the bill for keying its standards to model size and training cost rather than to actual deployment risk, warning that the approach could give the public “a false sense of security.”¹

The confusion is understandable. Much of what circulates online as “SB 1047 compliance” --- a 10^25 derivative-model tier, a $100 million training-cost test, a mandatory kill switch, a 72-hour incident clock, $10,000-per-violation penalties --- comes from SB 1047’s proposed text. None of it is California law. Some of those ideas reappeared, in changed form, in the bill that did pass. The figures did not.

If a vendor, template, or memo tells you to comply with SB 1047, it is describing a bill that does not exist as law. Treat that as a signal to re-check the rest of the analysis.

Does SB 53 Apply to Your Company? The Frontier-Developer Threshold

SB 53 applies to “frontier developers” --- companies that train a “frontier model,” defined as a foundation model trained using more than 10^26 integer or floating-point operations.² That is an enormous amount of compute, on the order of the largest models from the best-funded labs. Within that group, the statute draws a second line: a “large frontier developer,” defined by annual gross revenue above $500 million, carries the framework-publication and catastrophic-risk obligations that smaller frontier developers do not.²

For the typical company that asks us about this law, the practical answer is that it does not directly apply. If you fine-tune an open-weight model, build a product on top of OpenAI’s, Anthropic’s, or Google’s APIs, or train models well below the 10^26 line, you are not the regulated frontier developer --- the lab that trained the underlying model is. The law targets the handful of organizations training the largest models, not the much larger population of companies deploying them.

The threshold question is not “do we use frontier AI” --- it is “do we train a model above 10^26 operations.” For almost every startup, the honest answer is no, and the SB 53 analysis stops there. The laws in the next section are the ones that reach you.

What Does SB 53 Require of the Developers It Covers?

For the frontier developers it reaches, SB 53 imposes disclosure and reporting duties rather than the prescriptive engineering mandates that sank SB 1047. There are four load-bearing obligations.

A Published Frontier AI Framework

A large frontier developer must publish a frontier AI framework describing how it governs catastrophic risk --- how it incorporates recognized national and international standards, assesses and mitigates the risk that a model materially contributes to a catastrophic event, and secures model weights against theft.² This is a public-accountability document, not a filing with a regulator: the statute makes the framework something the company posts and stands behind, with the Attorney General able to act if the company fails to publish or follows a framework that omits required elements.

A Transparency Report at Deployment

Before or when it deploys a new or substantially modified frontier model, a frontier developer must publish a transparency report covering the model’s intended uses, its capabilities and limitations, and the assessments run for catastrophic risk.² The reporting obligation is tied to the deployment event, so each materially new model is its own disclosure trigger rather than a one-time filing.

Critical-Safety-Incident Reporting to Cal OES

A frontier developer must report a “critical safety incident” to the California Office of Emergency Services within 15 days of discovery, and within 24 hours when the incident poses an imminent risk of death or serious physical injury.² This is the provision most often misstated as SB 1047’s “72 hours to the Attorney General.” Under the law that actually passed, the recipient is Cal OES and the clock is 15 days --- compressed to 24 hours only for imminent-harm scenarios. The incidents that trigger it track the statute’s “catastrophic risk” definition: a single incident causing the death of or serious injury to more than 50 people, or more than $1 billion in property damage, that a frontier model materially contributed to --- through such hazards as CBRN weapons, attacks on critical infrastructure, or a model operating beyond human control.²

Whistleblower Protections

SB 53 protects employees who report frontier-AI safety concerns, barring retaliation and requiring covered developers to maintain channels for raising them.² For any company at or near the covered tier, this changes internal compliance design: an engineer who believes a catastrophic-risk assessment was cut short has a statutory route to escalate, and the company’s response to that escalation is itself now a compliance event.

If you are a large frontier developer, the published framework and the transparency report are the two artifacts a regulator or plaintiff will ask for first. Draft them as documents you would be comfortable defending, because they are public by design.

What Are the Penalties Under SB 53?

SB 53 is enforced by the California Attorney General, with civil penalties of up to $1,000,000 per violation.² That ceiling is two orders of magnitude above the $10,000 figure that circulates with the vetoed SB 1047 text, and it reflects the statute’s focus on the largest developers rather than on startups. There is no separate criminal tier and no private right of action; enforcement runs through the Attorney General. The law also funds CalCompute, a public computing-cluster initiative meant to widen access to large-scale compute beyond the incumbent labs.²

The Other California AI Laws You Actually Need to Watch

Because SB 53 reaches so few companies directly, the more important compliance questions for most AI startups sit in two statutes that share its January-2026 moment but a far broader scope.

AB 2013 --- Generative AI Training-Data Transparency

AB 2013 requires a developer of a generative-AI system made available to Californians to post a high-level, public description of the data used to train it --- the sources, whether the data was licensed or purchased, whether it includes personal information, and the time periods it spans.³ It took effect January 1, 2026 and applies to systems released on or after January 1, 2022, so it captures models that long predate it. This is the California AI-transparency obligation most startups deploying generative features will actually owe. It is also being litigated: xAI has sued the Attorney General challenging the law on trade-secret and constitutional grounds, so the precise contours may shift.³

SB 942 --- The California AI Transparency Act (Watermarking)

The statute that genuinely bears the name “California AI Transparency Act” is SB 942, a watermarking and content-provenance law --- not SB 1047. SB 942 requires covered generative-AI providers to offer detection tooling and to attach provenance disclosures to AI-generated content.⁴ AB 853 amended SB 942 in 2025, delaying its effective date to August 2, 2026 and phasing in obligations for large online platforms and device manufacturers in 2027 and 2028.⁵ If your product generates images, audio, or video, this is the regime to calendar.

Run the analysis in this order: Are you a frontier developer under SB 53 (almost certainly not)? Do you offer a generative-AI system to Californians under AB 2013 (quite possibly)? Do you generate synthetic media under SB 942 (if so, August 2, 2026 is your date)? Most companies owe nothing under the first and real obligations under the second or third.

Will Federal Law Preempt California’s AI Statutes?

The honest answer is that it is being fought over right now, and you should not build a compliance plan around federal rescue. Through 2025 the federal posture turned deregulatory: a proposed multi-year moratorium on state AI laws moved through Congress, a national AI Action Plan issued in July 2025, and a December 11, 2025 executive order directed federal agencies to identify and challenge state AI laws the administration views as obstructive.⁶

What none of that has done is repeal California’s statutes. An executive order cannot by itself displace a duly enacted state law; preemption generally requires an act of Congress or a conflict a court is willing to enforce, and the moratorium proposals have drawn bipartisan resistance. The practical reading for a company operating today is that SB 53, AB 2013, and SB 942 are operative law on their stated effective dates, and the preemption question is a live risk to track --- not a reason to defer compliance. The federal fight is real, but it is a reason to watch, not a reason to wait.

A Compliance Triage for California AI Obligations

The table below maps each statute to the company it reaches and the first action it requires. Use it to decide where, if anywhere, you have work to do.

For most AI startups the entire California AI-law obligation reduces to one or two public disclosures --- a training-data summary under AB 2013, and provenance tooling under SB 942 if you generate media. SB 53 is a law to understand, not usually one to comply with.

Need Help Sorting Out Which California AI Law Applies to You?

Astraea Counsel advises AI companies on California and federal AI compliance --- threshold analysis under SB 53, training-data disclosures under AB 2013, provenance obligations under SB 942, and the federal preemption landscape. Explore our AI & Emerging Tech legal services.

Related Resources

• Federal AI Regulation Landscape --- Federal legislation, agency rulemaking, and the preemption fight
• AI Model Training Data Rights --- Copyright and disclosure compliance for training datasets
• Caremark Duty for AI Agent Deployment --- Board-level oversight obligations for AI systems
• AI & Emerging Technology Services --- Comprehensive AI legal counsel
• Contact Us --- Discuss your compliance needs

---

Footnotes

1. California Senate Bill 1047, Safe and Secure Innovation for Frontier Artificial Intelligence Models Act (2023—2024 Reg. Sess.), legislative history: “09/29/24 Vetoed by the Governor,” available at https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202320240SB1047; Governor Gavin Newsom, Veto Message, SB 1047 (Sept. 29, 2024). ↩ ↩² ↩³ ↩⁴

2. California Senate Bill 53, Transparency in Frontier Artificial Intelligence Act (2025—2026 Reg. Sess.), signed Sept. 29, 2025, effective Jan. 1, 2026, codified at Cal. Gov. Code §§ 22757.11—22757.15 (frontier-model and large-frontier-developer definitions, § 22757.11(i)—(j); catastrophic-risk definition, § 22757.11(c)(1); critical-safety-incident reporting, § 22757.13(c); civil penalty, § 22757.15), available at https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB53. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹

3. California Assembly Bill 2013, Generative Artificial Intelligence: Training Data Transparency (2023—2024 Reg. Sess.), effective Jan. 1, 2026, available at https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202320240AB2013. ↩ ↩² ↩³

4. California Senate Bill 942, California AI Transparency Act (2023—2024 Reg. Sess.), available at https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202320240SB942. ↩ ↩²

5. California Assembly Bill 853 (2025—2026 Reg. Sess.), signed Oct. 13, 2025, amending SB 942 (delaying its effective date to Aug. 2, 2026 and phasing in generative-AI hosting-platform, large-online-platform, and capture-device-manufacturer obligations), available at https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260AB853. ↩ ↩²

6. Executive Order, “Ensuring a National Policy Framework for Artificial Intelligence” (Dec. 11, 2025), directing federal agencies to identify and challenge state AI laws, available at https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/; see also the proposed state-AI moratorium in 2025 federal budget legislation and the federal AI Action Plan (July 23, 2025). ↩ ↩²