Summary
- Targets only frontier models: foundation models trained with more than 10^26 integer or floating-point operations, counting fine-tuning and other later modifications (BPC 22757.11(i)).
- All frontier developers must publish a transparency report when they release a new or substantially modified frontier model (website, contact route, release date, languages, output modalities, intended uses, use restrictions). A model card or system card that contains this counts (22757.12(c)).
- Large frontier developers (frontier developer plus affiliates with annual gross revenue above USD 500 million in the prior calendar year) must also write, follow and publish a frontier AI framework, review it at least yearly, publish material changes with reasons within 30 days, add risk assessment summaries to each transparency report, and send the Office of Emergency Services (OES) a summary of internal-use catastrophic risk assessments every three months or on an agreed schedule (22757.12(a), (b), (d)).
- Critical safety incidents must be reported to OES within 15 days of discovery; if there is an imminent risk of death or serious physical injury, an appropriate authority must be told within 24 hours (22757.13(c)).
- Whistleblowers: frontier developers may not gag or punish covered employees who report catastrophic risk or violations to the AG, federal authorities or managers; large frontier developers need an anonymous internal reporting channel with monthly updates to the reporter (Labor Code 1107.1).
- Also sets up a consortium to design CalCompute, a public cloud cluster, but only once money is budgeted (Gov. Code 11546.8).
Who it applies to
- Frontier developers (anyone who trained or started training a frontier model) and, for most duties, large frontier developers.
- Extraterritorial: yes, on the text. The definitions contain no California location test, so a developer based anywhere that meets the compute threshold is covered. In practice the AG's reach depends on the company doing business in or offering models to California.
- Not covered: companies that only fine-tune or use other people's models, unless their own total training compute crosses 10^26 operations.
- The Act yields where it strictly conflicts with a federal government contract or is preempted by federal law, and it overrides city and county frontier-risk rules adopted on or after 1 Jan 2025 (Sec. 5).
Key dates
- Approved by Governor and filed with Secretary of State (Chapter 138)
- Act effective (Business and Professions Code 22757.10 to 22757.16; Labor Code 1107 to 1107.2)
- First Department of Technology review of thresholds; first OES and AG annual anonymised reports; CalCompute framework report dueUpcoming
Full date record
| Date | Event |
|---|---|
| 2025-09-29 | Governor approves SB 53 (Chapter 138) |
| 2026-01-01 | Act in force |
| 2027-01-01 | Department of Technology first yearly review of the definitions; OES and AG first yearly anonymised reports; CalCompute consortium report to the Legislature |
Penalties and enforcement
- Large frontier developers face civil penalties of up to USD 1,000,000 per violation, scaled to severity, for: not publishing or sending a required document, false or misleading statements about catastrophic risk or framework compliance, not reporting an incident, or not following their own framework (22757.15(a)).
- Only the Attorney General can sue for these penalties (22757.15(b)).
- Good-faith, reasonable statements are not violations (22757.12(e)(2)). Redactions for trade secrets, cybersecurity, public safety or national security are allowed but must be explained, with unredacted copies kept for five years (22757.12(f)).
- OES may later designate federal standards that, if followed, count as compliance with incident reporting (22757.13(h), (i)).
- Federal preemption watch: Executive Order 14365 (11 Dec 2025) told the US Attorney General to set up an AI Litigation Task Force to challenge state AI laws. Law firm commentary names SB 53 as a possible target, but no court challenge to SB 53 was found as of 2026-10-19 (secondary sources only). No amendments to the chapter were shown on the official code page on that date.
What it means for an Indian company
- Most Indian SaaS and gen-AI app makers are not in scope. Building on top of OpenAI, Anthropic, Google, Meta or open-weight models does not make you a frontier developer unless your own training compute crosses 10^26 operations.
- Indian foundation model labs (for example IndiaAI Mission model builders) should track their cumulative training compute, including fine-tuning and reinforcement learning, against the threshold. Crossing it while offering the model in California triggers transparency reports even below the USD 500 million revenue line.
- Use it as a free template. The framework headings (risk thresholds, mitigations, third-party testing, weight security, incident response, internal governance) map well to SC-AI-005 AI risk assessment, SC-AI-018 Robustness and adversarial testing (red teaming), SC-AI-029 AI-specific security controls (prompt injection, model abuse) and SC-AI-031 AI incident management and regulator reporting.
- Vendor due diligence. Ask your model providers for their published frontier AI framework and latest transparency report and file them under SC-AI-007 Third-party and vendor AI due diligence.
- Model cards double as compliance. Publishing a model card with the required fields satisfies the basic transparency report (SC-AI-014 Model documentation (model cards, system cards), SC-AI-025 Public AI transparency statement and notices).
- Whistleblower channel. If you ever become a frontier developer, your anonymous reporting process and employee notices need to exist from day one (SC-AI-002 AI roles, accountability and oversight committee).
Obligations
Full obligation breakdown coming
We break the most requested policies down into numbered obligations mapped to controls. Leave your details and we’ll tell you when California SB 53 (TFAIA) is done.
Glossary
- Frontier model: a foundation model trained with more than 10^26 operations of compute.
- Frontier developer / large frontier developer: anyone who trained such a model; "large" adds annual group revenue above USD 500 million.
- Frontier AI framework: a public document describing how the developer finds, tests and reduces catastrophic risks.
- Transparency report: a short public release note for each new or substantially modified frontier model.
- Critical safety incident: for example theft of model weights causing injury, loss of control causing injury, or a model deceiving its developer to dodge oversight.
- OES: California's Governor's Office of Emergency Services, which receives incident reports.
- Covered employee: an employee responsible for assessing or managing critical safety incident risk.
- CalCompute: a planned state public cloud computing cluster for safe AI research.
Key takeaways
- In force since 1 Jan 2026; signed 29 Sep 2025.
- Scope is narrow: models above 10^26 operations; heavier duties for groups above USD 500 million revenue.
- Core duties: publish a frontier AI framework, publish transparency reports, report critical incidents to OES (15 days, or 24 hours if lives are at risk), protect whistleblowers.
- AG-only penalties up to USD 1 million per violation; no 2026 amendments or court rulings found.
- For most Indian companies it is a vendor diligence item and a useful framework template, not a direct obligation.
Official sources
- SB 53 chaptered bill text and digest, California Legislative Information, accessed 2026-10-19
- Business and Professions Code, Chapter 25.1 (22757.10 to 22757.16), accessed 2026-10-19
- California Attorney General, SB 53 employee reporting page, accessed 2026-10-19
- California Department of Technology, TFAIA page, accessed 2026-10-19
- Executive Order 14365, Federal Register, 16 Dec 2025, accessed 2026-10-19
This is general information, not legal advice. Last verified 19 Oct 2026 against the official source. Laws change; confirm the current position with qualified counsel before acting.