GLACIS·US state AI laws·California·Updated April 2026
California AI laws, the consolidated playbook for April 2026.
Eighteen statutes and rules now govern AI in California. The most-asked questions in 2026: AB 2013 training-data transparency (live), SB 53 frontier transparency (live), CPPA ADMT (phased), the FEHA employment-AI rules (live), the SB 942 watermarking law (delayed to August 2026), and where AB 1018 actually landed (held in inactive file).
In this guide:
SB 53 (Frontier AI Transparency Act) effective January 1, 2026. Roughly 5–8 frontier developers (OpenAI, Anthropic, Google DeepMind, Meta, Microsoft) in scope. Transparency reports before deploying or substantially modifying a frontier model; annual NIST AI RMF or ISO/IEC 42001-aligned framework; 15-day critical-incident reporting (24-hour for imminent threats); whistleblower protections; up to $1M civil penalty per violation by AG.[CA1]
SB 942 (California AI Transparency Act / CAITA) operative date moved. Originally Jan 1, 2026; AB 853 (signed Oct 13, 2025) extended the operative date to August 2, 2026. Watermarking, provenance metadata, and free public AI-detection tool obligations apply to covered providers.[CA2]
AB 1018 (Automated Decisions Safety Act) – ordered to inactive file September 13, 2025 at request of Sen. Wiener. Did not become law; expect re-introduction in 2026 session.[CA3]
AB 2013 (training-data transparency) live since January 1, 2026. Public disclosure for any generative AI system available to Californians since January 1, 2022. Enforced via the Unfair Competition Law — both AG and private-action exposure.[CA4]
FEHA employment-AI rules in force since October 1, 2025. Disparate-impact framework, 4-year record retention, anti-bias testing as defense.
Executive summary
California regulates AI across training data, employment, consumer disclosure, and political advertising—more domains than any other US state. AB 2013 (effective January 2026) requires generative AI developers to publicly disclose training data information. The Civil Rights Council’s employment AI rules (effective October 2025) make employers liable for AI-driven discrimination even without intent.
SB 1001 (2019) requires bots to disclose their artificial identity, while AB 2355 (2025) mandates AI disclosure in political advertising. The ambitious SB 1047 frontier AI safety bill was vetoed by Governor Newsom in September 2024, but its concepts continue to influence national AI policy discussions.
California’s deepfake laws targeting platforms (AB 2655) and distribution (AB 2839) have faced legal challenges—blocked and struck down respectively on First Amendment and Section 230 grounds—highlighting the constitutional complexities of AI content regulation.
AB 2013: Training Data Transparency
Signed September 28, 2024 and effective January 1, 2026, AB 2013 is the first US law requiring generative AI developers to publicly disclose training data information. It applies retroactively to systems released or substantially modified since January 1, 2022.
Required Disclosures
Developers must publicly disclose:
- Description of datasets: How they further the AI system’s purpose
- Number of data points: Scale of training data
- IP content: Whether datasets include copyrighted, trademarked, or patented data
- Data acquisition: Whether datasets were purchased or licensed
- Personal information: Whether datasets contain personal or aggregate consumer information
- Processing: Any cleaning, processing, or modification to datasets
Who’s Covered
- Developers of generative AI systems
- Systems available to California residents
- Retroactive to January 1, 2022
Trade Secret Challenge
Companies must balance transparency with proprietary information protection. While trade secrets are not explicitly exempted, the disclosure requirements focus on categories and characteristics rather than specific dataset contents.
Employment AI Discrimination Rules
The California Civil Rights Council approved AI employment regulations on June 27, 2025, effective October 1, 2025. These rules apply the Fair Employment and Housing Act (FEHA) to Automated-Decision Systems (ADS) used in employment decisions.
Key Requirements
Liability Standard
- Unlawful to use ADS resulting in discrimination
- Liability even without discriminatory intent
- Disparate impact creates liability
Defense & Evidence
- Anti-bias testing can be used as defense
- Absence of testing can be evidence against
- Retain ADS records for 4 years
Record Retention (4 Years)
Selection Criteria
How ADS evaluates candidates
Outputs
Decisions and recommendations
Audit Findings
Bias testing results
Political Ads & Bot Disclosure
SB 1001
Bot Disclosure • Effective July 2019
- Platforms with 10M+ monthly US users
- Unlawful to deceive about artificial identity
- Covers commercial transactions & elections
- "Clear, conspicuous" disclosure required
AB 2355
Political AI Ads • Effective Jan 2025
- Committees with $2,000+ contributions
- Required disclaimer on AI-generated ads
- "Generated or substantially altered using AI"
- FPPC enforcement
Deepfake Laws: Legal Challenges
AB 2655 (Blocked)
Required platforms to remove deceptive election content. Blocked by federal court January 3, 2025 through June 28, 2025 on Section 230 preemption grounds.
AB 2839 (Struck Down)
Prohibited distribution of deceptive AI content near elections. Struck down October 2024 as violating First Amendment—judge ruled it "hinders humorous expression."
SB 1047: The Vetoed Frontier AI Bill
Vetoed September 29, 2024
Governor Newsom vetoed SB 1047, stating it "does not take into account whether an AI system is deployed in high-risk environments, involves critical decision-making, or uses sensitive data." Despite the veto, the bill’s concepts continue to influence AI policy discussions nationally.
What SB 1047 Would Have Required
Coverage
- • Models costing $100M+ to train
- • Models with 10²⁶+ FLOPs
- • "Frontier" AI models only
Requirements
- • Safety and security protocols
- • Shutdown capabilities
- • Third-party annual audits (from 2026)
- • 72-hour incident reporting
- • Whistleblower protections
Critical harms defined: WMD creation, cyberattacks on critical infrastructure ($500M+ damage), autonomous crimes causing mass casualties. Penalties would have been up to 10% of training computing costs.
Notable Industry Support
Despite industry opposition, SB 1047 had surprising support from within AI companies:
- xAI CEO Elon Musk publicly supported the bill
- 113+ employees of OpenAI, DeepMind, Anthropic, Meta, and xAI signed letters of support
References
- [CA1] Office of Governor Newsom, SB 53 signing statement (Sept 29, 2025) — gov.ca.gov; Future of Privacy Forum, “California’s SB 53: The First Frontier AI Law, Explained”; Brookings, “What is California’s AI safety law?”.
- [CA2] Troutman Pepper Locke, “California AI Transparency Act Amendments Signed Into Law” (Oct 2025); California legislative information for AB 853.
- [CA3] California legislative information, AB 1018 status (inactive file Sep 13, 2025); EPIC summary; Foley & Lardner (Mar 2026).
- [CA4] Crowell & Moring, “California’s AB 2013 Requires Generative AI Data Disclosure by January 1, 2026”; Goodwin, “California’s AB 2013 Takes Effect” (Jan 2026); Davis+Gilbert.
Operating AI in California?
Make the receipts. Book the Sprint.
The Glacis Agent Runtime Security & Evidence Sprint produces signed evidence receipts of training-data disclosure (AB 2013), frontier transparency reporting (SB 53), ADMT pre-use notices, and FEHA disparate-impact testing — the kind of evidence a CPPA inspector would expect. Runtime controls run inside your infrastructure with zero sensitive-data egress.
Book the Agent Runtime Security Sprint See a sample evidence pack →