California AI Laws Guide: SB 53, CCPA ADMT Rules and AI Transparency

California has no single AI act: it regulates AI through several separate laws with different targets, thresholds, regulators and start dates, and teams that track only one of them miss the rest. Frontier model developers face the Transparency in Frontier Artificial Intelligence Act (SB 53). Any business covered by the California Consumer Privacy Act that uses automated decisionmaking technology for significant decisions faces new California Privacy Protection Agency (CPPA) regulations. Generative AI developers face training data transparency (AB 2013), and large generative AI providers, platforms and device makers face content provenance duties under the California AI Transparency Act (SB 942 as amended by AB 853). This guide sets out who each law covers, what it requires, the dates confirmed as at 24 October 2026, and how security and governance teams can build one programme that meets all four. Consumer privacy law in general is covered in the US State Privacy Laws Guide; this guide covers only the AI-specific layers.

Key takeaways

• Four regimes, four audiences. SB 53 targets frontier developers, the ADMT rules target CCPA businesses, AB 2013 targets generative AI developers, and the AI Transparency Act targets large generative AI providers, platforms and capture device makers.
• ADMT compliance is due by 1 January 2027. Businesses using ADMT for significant decisions must give a Pre-use Notice, an opt-out (or a qualifying exception) and an access right by that date.
• SB 53 is about published frameworks and incident reporting. Large frontier developers publish a frontier AI framework, file transparency reports, and report critical safety incidents to the Office of Emergency Services within 15 days, or 24 hours where there is imminent danger.
• Provenance duties start in stages. The AI Transparency Act became operative on 2 August 2026; platform duties follow on 1 January 2027 and capture device duties for devices first produced from 1 January 2028.
• Agents and model weights are in scope. Unauthorised access to frontier model weights can be a critical safety incident, and AI agents that make hiring, lending or housing decisions are ADMT.

Scope: who and what is covered

Key definitions

• Automated decisionmaking technology (§ 7001(e)): technology that processes personal information and uses computation to replace, or substantially replace, human decisionmaking. Human involvement only counts if the reviewer can interpret the output, analyse it with other relevant information, and has authority to make or change the decision. Tools such as spam filters, web hosting and spreadsheets are excluded unless they replace human decisionmaking.
• Significant decision (§ 7001(ddd)): a decision resulting in the provision or denial of financial or lending services, housing, education enrolment or opportunities, employment or independent contracting opportunities or compensation, or healthcare services. Advertising is excluded.
• Catastrophic risk (SB 53): a foreseeable and material risk that a frontier model contributes to more than 50 deaths or serious injuries, or more than $1 billion in property damage, in a single incident, for example through expert-level help with chemical, biological, radiological or nuclear weapons, or cyberattacks carried out without meaningful human oversight.
• Critical safety incident (SB 53): includes unauthorised access to or exfiltration of model weights that results in death or bodily injury, harm from a materialised catastrophic risk, and loss of control of a frontier model causing death or bodily injury.

Key obligations

SB 53: frontier developers

The Governor's office describes the reporting mechanism as open to frontier AI companies and to the public, and the California Department of Technology must recommend updates to the law each year.

CCPA ADMT regulations: any covered business

• Pre-use Notice (§ 7220): before using ADMT for a significant decision, tell consumers how the ADMT works and how the decision would be made if they opt out. It may be combined with the Notice at Collection if it meets the content rules. Trade secrets and certain security, fraud prevention and safety information can be withheld.
• Opt-out (§ 7221): consumers can opt out of ADMT for significant decisions. Exceptions exist for a qualifying human appeal process, for certain admission, acceptance and hiring decisions, and for allocation or assignment of work and compensation, each with conditions. The draft security and fraud prevention exception was removed in the final text.
• Access (§ 7222): on request, explain the logic of the ADMT and the outcome of the decision for that consumer.
• Response times (§ 7021): confirm receipt of access and appeal requests within 10 business days and respond within 45 calendar days, extendable to 90.
• Risk assessments (Article 10): using ADMT for significant decisions is one of the processing activities that requires a risk assessment; certain assessments must be completed by 31 December 2027 (§ 7155(b)), updated within 45 days of a material change, and submitted to the Agency by 1 April in line with § 7157.

The same rulemaking added cybersecurity audit duties (Article 9), with the first audit reports due 1 April 2028 for businesses with over $100 million in annual gross revenue and later dates for smaller businesses up to 1 April 2030.

AB 2013: training data documentation

Developers of generative AI systems or services released on or after 1 January 2022 had to post documentation on their website about the data used to train the system, including a high-level summary of the datasets, on or before 1 January 2026, and must do so again before each later time a new or substantially modified version is made available to Californians. In practice this means a maintained dataset register: sources, licensing, whether personal information or copyrighted material is included, and collection periods.

California AI Transparency Act: provenance

• Covered providers: include latent disclosures (machine-readable provenance) in AI-generated image, video and audio content to the extent technically feasible, permanent or extraordinarily difficult to remove, and offer an AI detection tool at no cost to users.
• Large online platforms (from 1 January 2027): detect compliant provenance data, disclose it through the user interface, and not knowingly strip system provenance data or digital signatures, to the extent technically feasible.
• Generative AI hosting platforms (from 1 January 2027): may not knowingly make available a generative AI system that does not place the required disclosures.
• Capture device manufacturers (devices first produced on or after 1 January 2028): offer users latent disclosures in captured content, embedded by default, to the extent technically feasible.

Provenance matters to security teams because synthetic media now drives fraud: in the Arup deepfake fraud, a video call of deepfaked executives led to HK$200 million in transfers. See the Deepfake and AI Impersonation Guide.

Timeline and current status

Status as at 24 October 2026: all four regimes are enacted, and we found no 2026 amendment to AB 853 listed on the bill record. CalCompute becomes operative only upon an appropriation. On 11 December 2025 a White House executive order named California among states whose AI laws it criticised and set up a federal AI Litigation Task Force; according to the White House fact sheet the order does not itself preempt state law. Watch for litigation, and check the official texts before relying on any date.

Enforcement and penalties

SB 53 also preempts local rules adopted on or after 1 January 2025 that specifically regulate frontier developers' management of catastrophic risk.

What it means for AI agents and non-human identities

• Agents making significant decisions are ADMT: an agent that screens job applicants, sets credit limits or allocates shifts substantially replaces human decisionmaking unless a qualified human reviews with authority. AI hiring platforms already carry identity risk: the McHire default password flaw exposed applicant data through a forgotten vendor admin account. See the AI Agent Authorisation Guide.
• Model weights are crown-jewel secrets: unauthorised access to or exfiltration of model weights is part of the SB 53 critical safety incident definition. Workload identities, storage keys and CI/CD tokens with access to weights need privileged-access treatment. See the Zero Trust for AI Agents Guide.
• Loss of control is reportable: an agent acting outside its intended scope in a way that causes harm can meet the loss-of-control limb, so agent action logs and kill switches support both detection and the 15-day report.
• Access requests need traceability: to explain an ADMT decision under § 7222, you must know which model, version, agent and data produced it. Give every agent a distinct identity and log its decisions. See the Identity Governance for AI Agents Guide.
• Provenance keys are credentials: signing keys used to add digital signatures and provenance metadata must be managed like any other machine secret, with rotation and access review.

Compliance steps

1. Map applicability: for each AI system, record whether you are a frontier developer, a CCPA business using ADMT, a generative AI developer, a covered provider, a platform or a device maker. Use the AI Register and AI Inventory Template.
2. Find ADMT uses first: the 1 January 2027 deadline is the nearest for most businesses. Inventory every model, rules engine and agent that touches financial, housing, education, employment or healthcare decisions.
3. Decide opt-out or exception: where a human appeal or hiring exception fits, document how it meets the conditions; otherwise build the opt-out flow and an alternative human process. Pilot with one decision type, measure volumes, then extend.
4. Write Pre-use Notices and access responses: draft templates that explain logic and outcomes in plain language; test them with the 10-day and 45-day clocks in your request-handling system.
5. Run risk assessments: start with ADMT uses and training uses so the 31 December 2027 deadline and 1 April submissions are met.
6. Frontier developers: publish the framework, set up the incident triage that can meet 15-day and 24-hour clocks, and build the anonymous internal reporting channel.
7. Generative AI developers and providers: keep the AB 2013 dataset documentation current with each release, and test latent disclosures survive common transformations.
8. Prepare rollback: keep the previous decision process (human or older model) available so an ADMT use can be withdrawn without halting a business process.

Framework mapping and measurement

Metrics:

• Percentage of ADMT uses for significant decisions with a Pre-use Notice and opt-out or documented exception.
• Median days to answer ADMT access requests (target well under 45).
• Percentage of required risk assessments completed against the 31 December 2027 deadline.
• Percentage of identities with access to model weights reviewed in the last quarter.
• Percentage of AI-generated media outputs passing latent disclosure checks.

Common failure patterns

• Treating the "human in the loop" as automatic: a reviewer who rubber-stamps outputs without authority or analysis does not take a process out of ADMT.
• Only tracking SB 53: most businesses are not frontier developers but are CCPA businesses with ADMT duties.
• Vendor AI unaccounted for: hiring and lending platforms with built-in scoring are your ADMT, and their admin accounts are your attack surface.
• Weights guarded like ordinary data: broad storage access and long-lived tokens on model artefacts make a weights compromise, and a reportable incident, more likely.
• Provenance stripped in pipelines: image processing and CDN steps remove metadata that the AI Transparency Act requires providers and platforms to preserve.

Practitioner checklist

• Record which California AI regimes apply to each AI system.
• Inventory every ADMT use in significant decisions before 1 January 2027.
• Document opt-out flows or qualifying exceptions, including human appeal criteria.
• Publish Pre-use Notices and build ADMT access responses.
• Schedule risk assessments to meet § 7155 and § 7157 dates.
• Give each AI agent a distinct identity, owner and decision log.
• Treat model weights and provenance signing keys as privileged secrets.
• Frontier developers: test the 15-day and 24-hour incident reporting process.
• Keep AB 2013 training data documentation current per release.

Standards and references

• California Legislature (via LegiScan): SB 53 Transparency in Frontier Artificial Intelligence Act, chaptered text
• California Privacy Protection Agency: CCPA Updates, Cybersecurity Audits, Risk Assessments, ADMT and Insurance Regulations
• California Privacy Protection Agency: Final Statement of Reasons and Updated Informative Digest
• California Legislature (via LegiScan): AB 853 California AI Transparency Act, chaptered text
• CalMatters Digital Democracy: AB 2013 Generative AI Training Data Transparency
• Office of the Governor of California: Governor Newsom signs SB 53
• The White House: Fact Sheet, Ensuring a National Policy Framework for Artificial Intelligence
• OWASP: Top 10 for LLM Applications 2025

This guide summarises California's AI-specific laws and regulations as at 24 October 2026, and is not legal advice. Related NHI Mgmt Group resources: Global AI Regulation Map · US State Privacy Laws Guide · AI Governance Guide · NIST AI RMF Guide · Colorado AI Act Guide