If your AI-governance program doesn't build frontier AI models, SB 53 probably doesn't name you directly, but your legal or procurement team is still going to ask whether it applies to you, and "no" is only half the answer.
The Transparency in Frontier Artificial Intelligence Act (TFAIA), better known as SB 53, targets the companies that train frontier models: OpenAI, Anthropic, Google DeepMind, Meta, and a handful of others crossing the compute threshold defined below. Almost everyone else, including enterprises that license and deploy those companies' models, sits outside SB 53's direct reach. But the law reshapes what a responsible AI-governance program should be asking its model vendors, and that's the part most coverage of SB 53 skips.
Key Takeaways
- SB 53 only binds "frontier developers": entities that train or initiate the training of a foundation model using more than 10^26 floating-point operations (FLOPs). Stricter rules apply to "large frontier developers," those in that group with over $500 million in annual gross revenue.
- Large frontier developers must publish an annual frontier AI framework describing how they identify and mitigate catastrophic risk; all frontier developers, regardless of size, must publish a transparency report before or when they deploy a new or materially modified frontier model.
- Critical safety incidents go to California's Office of Emergency Services within 15 days, or 24 hours if there's imminent risk of death or serious injury. The California Attorney General can fine violators up to $1 million per violation.
- If your organization only uses or buys frontier-model-based tools, SB 53 does not create a direct compliance obligation for you, but its disclosure requirements give you a concrete, statute-backed yardstick for vendor due diligence that didn't exist before January 2026.
What SB 53 actually requires, and of whom
SB 53's scope hinges on two defined terms, and getting them right changes everything about whether the law reaches your organization.
A frontier model is a foundation model trained using a quantity of computing power greater than 10^26 integer or floating-point operations, counting the compute used in the initial training run plus any subsequent fine-tuning or material modification. A frontier developer is any entity that trained or initiated the training of one of these models. Within that group, a large frontier developer is a frontier developer whose own annual gross revenue, combined with its affiliates, exceeded $500 million in the preceding calendar year, per the enrolled bill text on the California Legislature's own site. That revenue test is why SB 53 catches Anthropic and OpenAI but leaves a well-funded startup fine-tuning an open-weight model below the FLOPs line alone entirely.
The obligations split cleanly by tier, and the split matters because it tells you which disclosures to actually expect from a given vendor:
| Obligation | Applies to | What it requires |
|---|---|---|
| Frontier AI framework | Large frontier developers only | Annual public document covering governance structure, how catastrophic risk is identified and mitigated, cybersecurity practices, and alignment with recognized safety standards |
| Transparency report | All frontier developers | Published at or before deployment of a new or substantially modified frontier model; covers release date, modalities, intended uses, restrictions, and a summary of catastrophic-risk assessment results |
| Critical incident reporting | All frontier developers | Report to California's Office of Emergency Services within 15 days of discovery, or 24 hours if the incident poses imminent risk of death or serious physical injury |
| Whistleblower protections | All frontier developers | Internal anonymous reporting channel and a ban on retaliating against employees or contractors who flag catastrophic risk concerns |
"Critical safety incident" isn't a vague catch-all. The statute defines it around a short list of concrete triggers: a model causing death or serious injury through loss of control, a materialized catastrophic risk, unauthorized access that leads a model to cause serious harm, or a model deliberately evading a developer's own safety controls. That specificity, per analysis from the Future of Privacy Forum, is what separates SB 53 from a broader "report anything concerning" mandate that would have been effectively unenforceable.
Enforcement runs through the California Attorney General, who can seek civil penalties of up to $1 million per violation, scaled to the severity of the offense. There's no private right of action; this is a regulator-enforced statute, not a basis for individual lawsuits.
Governor Newsom, in signing the bill on September 29, 2025, framed it as a hedge against federal inaction: SB 53 will "provide a blueprint for well-balanced AI policies beyond \[California's\] borders, especially in the absence of a comprehensive federal AI policy framework," while "California has proven that we can establish regulations to protect our communities while also ensuring that the growing AI industry continues to thrive," according to the Governor's official signing statement. State Senator Scott Wiener, the bill's author, called it a case of "California stepping up, once again, as a global leader on both technology innovation and safety," in the same release.
Who SB 53 doesn't touch: the deployer's actual position
Most legal-alert coverage of SB 53, from firms like Wilmer Hale, Morrison Foerster, and Goodwin, is written for the frontier developers themselves: the dozen or so labs training models above the FLOPs threshold. If your organization procures a chatbot built on GPT, Claude, or Gemini, embeds a third-party model into your product, or simply uses a frontier model through an API, you are not a frontier developer under this statute, and none of the four obligations in the table above land on you directly.
That's the accurate reading of the law's actual text, and it's worth stating plainly rather than hedging around it, because the reflexive assumption inside a lot of AI-governance teams is "new AI law equals new obligation for us," and that assumption is wrong here. SB 53 regulates the small set of organizations building frontier-scale models, not the much larger population of enterprises buying and deploying what those labs ship.
Where the law does reach you is indirect, but real. Every frontier developer your organization relies on, whether that's a foundation-model API provider or an embedded generative AI feature in a SaaS product, now has to publish a transparency report and, if it clears the $500 million revenue bar, an annual frontier AI framework. That's a public paper trail you didn't have access to before 2026, and a properly run AI vendor due diligence process should be pulling those documents into its review rather than treating SB 53 as someone else's law.
SB 53 vs. the EU AI Act's GPAI obligations
For organizations operating in both California and the EU, SB 53 sits next to a more mature regime: Chapter V of the EU AI Act, which governs providers of general-purpose AI (GPAI) models. The two frameworks share a compute-threshold structure but differ in almost every other respect.
| Dimension | California SB 53 | EU AI Act Chapter V (GPAI) |
|---|---|---|
| Who is regulated | Frontier developers (compute-based); stricter tier for "large" developers over $500M revenue | All GPAI model providers, with an additional tier for "systemic risk" models |
| Compute threshold | 10^26 FLOPs defines a frontier model | 10^25 FLOPs creates a presumption of systemic risk under Article 51 |
| Baseline obligation | Transparency report at deployment (all frontier developers) | Technical documentation (Annex XI), downstream-provider documentation (Annex XII), training-data summary, and a copyright policy (Article 53) |
| Enhanced obligation | Frontier AI framework on catastrophic risk (large developers only) | Adversarial testing, systemic-risk assessment and mitigation, incident reporting, cybersecurity assurance (Article 55, systemic-risk models) |
| Incident reporting window | 15 days standard, 24 hours for imminent death/serious injury risk, to the Office of Emergency Services | Serious-incident reporting to the AI Office and national authorities under Article 55, without SB 53's fixed hour-count trigger |
| Enforcement | California Attorney General, up to $1M per violation | European Commission's AI Office, with fines up to 3% of global annual turnover or €15M for GPAI-specific violations |
| Effective since | January 1, 2026 | August 2, 2025 (models placed on market after that date; pre-existing models have until August 2, 2027) |
The practical takeaway for a multi-jurisdiction AI-governance program: the EU AI Act's GPAI regime is older, broader in scope (it reaches every GPAI provider, not just frontier-scale ones), and enforced with financial penalties tied to global revenue rather than a flat per-violation cap. SB 53 is narrower but newer, and its transparency-report requirement kicks in at deployment rather than requiring the ongoing technical documentation regime the EU built. A vendor that already produces EU AI Act Article 53 documentation is most of the way toward what SB 53 asks for, but the reverse isn't automatically true. Organizations already tracking GPAI compliance under the EU AI Act have a head start on evaluating SB 53 disclosures, since the underlying question, "can this vendor show its safety and risk-assessment work," is the same question asked in different words.
A vendor-diligence question set drawn from SB 53
Because SB 53 doesn't obligate you directly, the practical use of the law is as a checklist for evaluating the frontier-model vendors your organization already depends on. These questions map straight to the statute's own disclosure requirements, so a vendor that can't answer them is telling you something about how seriously it takes its own obligations, not just California's law.
- Is this vendor a "frontier developer" under SB 53 (does its model exceed 10^26 training FLOPs), and if so, has it published the transparency report the statute requires before deployment?
- If the vendor's revenue exceeds $500 million, has it published its annual frontier AI framework, and does that framework name the catastrophic-risk categories it actually assesses for?
- What does the vendor's transparency report say about the model's intended uses and restrictions, and does your actual use case fall inside or outside those stated boundaries?
- Does the vendor have a documented critical-incident reporting process, and can it tell you what would trigger a report to California's Office of Emergency Services versus a routine internal fix?
- Does the vendor maintain the whistleblower protections SB 53 requires, and has any public reporting suggested those channels have actually been used?
- For vendors operating in the EU as well, does the vendor's EU AI Act Article 53 technical documentation cover the same ground as its SB 53 transparency report, or are there gaps specific to one jurisdiction?
- Has the vendor disclosed any prior AG enforcement action or penalty under SB 53, and if so, what changed afterward?
Feeding these questions into a structured vendor record, rather than a one-off email thread, is what turns a disclosure law like SB 53 into an actual risk signal instead of a box to check once and forget. A vendor management module that tracks compliance documents, renewal dates, and risk scores per vendor keeps this current as frontier developers publish updated frameworks and reports year over year, rather than leaving the answers scattered across old emails from the onboarding process.
Common issues teams run into
"Our AI vendor won't confirm whether they're a large frontier developer." Revenue thresholds are self-assessed under SB 53, and vendors are not required to proactively tell customers which tier they fall into. Ask directly, and treat a non-answer as itself informative; a large frontier developer that's actually compliant has no reason to be vague about a document it's already published.
"We can't tell if our use case falls inside the vendor's stated restrictions." This is exactly what the transparency report's "intended uses and restrictions" section exists to answer. If your deployment (say, using a general-purpose model for a high-stakes hiring decision) isn't clearly covered, that's a governance gap regardless of what SB 53 requires of the vendor; pair the check with your own AI vendor due diligence process rather than relying on the vendor's disclosure alone.
"We operate in both California and the EU, and our compliance team is duplicating work." Map the two disclosure sets against each other once, using the comparison table above as a starting template, rather than running SB 53 and EU AI Act reviews as separate projects. Most of the underlying evidence, model documentation, risk assessments, incident logs, overlaps.
FAQ
Does SB 53 apply to companies that just use ChatGPT, Claude, or Gemini in their products?
No. SB 53 regulates the entities that train frontier models above the 10^26 FLOPs threshold, not the companies that license or embed those models afterward. If you're a customer of OpenAI, Anthropic, or Google rather than a company training your own frontier-scale model, SB 53's direct obligations don't apply to you.
What counts as a "critical safety incident" under SB 53?
The statute lists specific triggers: a model causing death or serious injury through loss of control, a materialized catastrophic risk, unauthorized access leading a model to cause serious harm, or a model deliberately evading its developer's safety controls. Routine bugs or minor performance issues don't meet this bar.
How much can a company be fined for violating SB 53?
The California Attorney General can seek civil penalties of up to $1 million per violation, with the amount scaled to the severity and nature of the violation. There is no private right of action; only the AG can enforce the statute.
Is SB 53 the same law as California's AI Transparency Act (SB 942)?
No, and this is a common mix-up. SB 942 governs AI-generated content labeling and detection tools for consumer-facing generative AI systems. SB 53, the Transparency in Frontier Artificial Intelligence Act, governs safety disclosures and incident reporting for the developers of frontier-scale foundation models. They address different problems and different regulated parties.
How does SB 53 compare to the EU AI Act for a company operating in both regions?
SB 53 is narrower (it only reaches frontier-scale developers) but was enforceable earlier for its incident-reporting piece, effective January 1, 2026. The EU AI Act's GPAI regime under Chapter V reaches a broader set of general-purpose AI providers and has been enforceable since August 2, 2025, with a lower compute threshold (10^25 FLOPs) for triggering enhanced "systemic risk" obligations. A vendor compliant with the EU regime is likely most of the way toward SB 53's requirements, though the reverse isn't guaranteed.
Do we need to update our AI governance program because of SB 53?
If your organization doesn't train frontier-scale models, SB 53 doesn't require you to change your own program's obligations. It does give you new, statute-backed documents to request from frontier-model vendors as part of your existing vendor risk assessment process, which is a due-diligence improvement worth making even without a direct legal mandate.
What is CalCompute, and does it affect enterprise AI users?
CalCompute is a public computing consortium SB 53 calls for, intended to support safe and ethical AI research and deployment with shared compute resources. It's aimed at expanding access to compute for research purposes rather than creating obligations for enterprises deploying commercial AI tools, so it has no direct compliance implication for a typical AI-governance program.
Who enforces SB 53, and can individuals sue over a violation?
Only the California Attorney General enforces SB 53. The statute does not create a private right of action, so affected individuals or organizations cannot bring their own lawsuit for a violation; any enforcement action has to come from the AG's office.
If your organization is building out an AI-governance program that needs to track vendor disclosures like these across dozens of frontier-model suppliers and multiple jurisdictions at once, Secure Privacy's AI Governance module classifies AI systems by risk tier, maps them to the regulations that actually apply, and keeps vendor documentation current inside the same platform used for GDPR, CCPA, and EU AI Act tracking, so a new state statute like SB 53 becomes one more mapped requirement rather than a separate spreadsheet.
