A plain-English guide to what Assembly Bill 2013 requires, who it affects, what’s at stake — and exactly how to comply before the deadline hits.

📅 Effective: January 1, 2026 🏛️ Jurisdiction: California, USA ⏱️ ~12 min read

What Is AB 2013?

Let’s start simply. Assembly Bill 2013, formally titled the Generative Artificial Intelligence Training Data Transparency Act, is a California state law that requires companies and developers who build generative AI systems to publicly disclose — in plain language — what data was used to train those systems.

Think of it as a nutritional label law, but for AI. Just like food manufacturers must tell you what’s in your cereal, AI developers must now tell the public what went into training their models.

It builds directly on the state’s existing privacy infrastructure, namely the California Consumer Privacy Act (CCPA) and its successor, the California Privacy Rights Act (CPRA).

📌 The Core Idea
AB 2013 is built on one foundational belief: if an AI system can affect people’s lives, the public deserves to know what it was trained on. Transparency isn’t a nice-to-have — it’s now the law.

Compliance deadline: January 1, 2026. And critically — it applies retroactively to models released or significantly updated since January 1, 2022.


What Does AB 2013 Cover?

The law applies specifically to generative AI systems — AI models capable of producing text, images, audio, video, code, or other synthetic outputs in response to prompts. This includes:

  • Large language models (LLMs) like chatbots, writing assistants, and search tools
  • Image generation models (e.g., text-to-image platforms)
  • Voice synthesis and audio generation tools
  • Video generation systems
  • Code generation tools used in developer environments
  • Multimodal models combining text, image, and/or audio outputs

What Must Be Disclosed?

Developers must publish a publicly accessible webpage documenting the following:

Disclosure RequirementWhat It Means in Practice
Data sourcesWeb scraping, licensed datasets, user-generated content, proprietary databases?
Data categoriesText, images, code, audio, video, or a combination?
Personal informationWas personal data included? How was consent obtained?
Copyrighted materialWas copyrighted content used? Under what license?
Data scaleGeneral size — e.g., billions of tokens, millions of images
Data collection datesWhen was the data collected or last updated?
Known limitations or gapsKnown biases, underrepresented groups, or quality issues?
⚠️ Important: Retroactive Applicability
If your generative AI product was released or significantly updated after January 1, 2022, you must retroactively document and disclose its training data by January 1, 2026. Many companies are caught off-guard by this provision.

Who Is Impacted?

Directly Impacted: AI Developers

Any company, organization, or individual that builds and releases a generative AI system for use in California — regardless of company size or revenue — falls under this law. Unlike SB 53 (which only targets companies with $500M+ revenue), AB 2013 applies to everyone.

Indirectly Impacted: Businesses Deploying Third-Party AI

If you license or integrate an AI model built by someone else, vendor due diligence just got a legal dimension. Your clients, partners, and regulators may ask you to verify the AI tools you deploy are compliant.

Who You AreYour AB 2013 ExposureUrgency
AI startup building an LLMHigh — direct compliance required🔴 Immediate
Enterprise SaaS with custom AI featuresHigh — if you trained or fine-tuned your own model🔴 Immediate
Company using OpenAI / Google APIsLow — verify vendor compliance🟡 Moderate
Healthcare AI using proprietary dataHigh — especially with personal health data🔴 Immediate
Academic / research institutionsMedium — depends on public release🟡 Moderate
Open-source AI developersMedium — public release triggers disclosure🟡 Moderate

How the Impact Happens — With Real Examples

Example 1: The HR Tech Startup

📍 Scenario

A startup launched an AI-powered resume screening tool in mid-2023, trained on scraped LinkedIn profiles, licensed job posting databases, and internal hiring data from enterprise clients.

AB 2013 Impact: The company must publish a data transparency disclosure documenting all three sources — consent on scraped profiles, licensing terms of the database, and whether the client data processing agreement permitted training use. Failure to address personal data use also triggers CCPA obligations simultaneously.

Example 2: The Marketing AI Platform

📍 Scenario

A marketing SaaS built an AI content generator trained on public blog posts, social media content, and ad copy via web scraping. They updated the model in Q3 2023 to add multilingual support.

AB 2013 Impact: Both the original model and the 2023 update fall within the retroactive window. The company must disclose the scraping methodology, copyright use, and flag any known limitations in multilingual performance.

Example 3: The Enterprise Healthcare AI

📍 Scenario

A health tech company trained a clinical documentation AI on de-identified patient notes from partner hospitals. The model was released to hospital clients in 2022.

AB 2013 Impact: Even de-identified health data triggers the disclosure requirement. The company must document how de-identification was achieved and on what legal basis. This intersects with HIPAA obligations and California’s SB 1120 healthcare AI rules.


Key Provisions to Watch Out For

🔍 Provision 1: The Retroactive Reach

Any model released or significantly modified after January 1, 2022 is subject. “Significantly modified” is broad — retraining, fine-tuning, or substantially updating capabilities likely counts. You cannot grandfather old models out of compliance.

🔍 Provision 2: Public Accessibility Requirement

The disclosure must be hosted on a publicly accessible webpage — not buried in a ToS, not behind a login wall, and not in technical jargon. Disclosures designed to technically comply while being practically unintelligible will not satisfy regulators.

🔍 Provision 3: Known Limitations Must Be Disclosed

If your training data has known demographic biases, geographic gaps, or quality issues, those must be disclosed. This isn’t just legal coverage — it’s about building AI that people can trust.

🔍 Provision 4: Interaction with Other California Laws

🚨 Common Mistake to Avoid
Don’t assume a “publicly available” or “open source” dataset means automatic compliance. You still must document it and verify whether personal or copyrighted content was embedded. “We didn’t know” is not a legal defense.

Best Practices for AB 2013 Compliance

Best Practice 1: Conduct a Full AI Training Data Audit

Map every generative AI system built or released since January 2022. For each, document data origins, personal information use, copyright licensing, known biases, and vendor contractual rights to allow training use.

Best Practice 2: Build a Living Transparency Page

Create a dedicated public-facing page at yourcompany.com/ai-transparency written in plain English, updated after every significant model change, with a visible “last updated” timestamp per model.

Best Practice 3: Renegotiate Data Vendor Contracts

Review every data supply agreement. Confirm that licenses explicitly permit AI model training use. If not — renegotiate or stop using that data. A contract amendment costs far less than a regulatory penalty.

Best Practice 4: Implement Data Provenance Tracking

Track where every piece of training data came from, when it was collected, and under what terms. The NIST AI Risk Management Framework provides solid methodology for doing this systematically.

Best Practice 5: Disclose Limitations Honestly

Document known data gaps, representation imbalances, and quality limitations. Honest disclosure builds user trust and reduces liability when models make mistakes.

Best Practice 6: Train Your Teams

Engineering, data science, product, and procurement teams all make training data decisions. Run internal training so everyone understands the disclosure obligations before they source data, not after.

Best Practice 7: Assign a Compliance Owner

Designate a Chief AI Officer, General Counsel, or dedicated AI compliance lead who is personally accountable for maintaining and updating your AB 2013 disclosures.

Best Practice 8: Monitor Regulatory Updates

California’s regulatory landscape is actively evolving. Subscribe to the CPPA’s official communications and assign someone to monitor updates — what’s sufficient today may need revision by mid-2026.


✅ AB 2013 Compliance Checklist

Use this before your January 1, 2026 deadline:

  • Identified all generative AI models released or updated after January 1, 2022
  • Completed a training data audit for each model
  • Documented all data sources, categories, and collection dates
  • Verified licensing rights for all third-party datasets used for training
  • Reviewed whether personal or copyrighted data is present in training sets
  • Created a publicly accessible AI transparency disclosure webpage
  • Written disclosures in plain, accessible language
  • Documented known data limitations, biases, or quality gaps
  • Assigned a compliance owner responsible for keeping disclosures updated
  • Established a process for updating disclosures after future model changes
  • Briefed engineering, data science, and product teams on requirements
  • Cross-checked compliance with CPPA, CCPA, and SB 942 where applicable

FAQs — Questions People Are Actually Asking

Sourced from real questions on Reddit, Quora, LinkedIn, and Google Search.

Q1. Does AB 2013 apply to my small startup if we’re not based in California?
Yes. If your AI product is accessible to or used by California residents, the law applies regardless of where your company is headquartered. If you’re deploying a generative AI product anywhere in the US, assume California compliance is required.
Q2. We fine-tuned a pre-trained model (like GPT or LLaMA). Do we still need to disclose training data?
Yes. If you fine-tuned a base model using your own data, you must disclose that fine-tuning dataset. You’re also responsible for verifying that the base model developer has published their own AB 2013-compliant disclosures. When in doubt, consult a qualified AI compliance attorney.
Q3. What counts as a “significant modification” that triggers the retroactive requirement?
If you retrained on new data, substantially altered architecture, expanded capabilities, or changed the core training dataset, that likely qualifies. Minor bug fixes or prompt updates generally do not. When in doubt, disclose — the cost of over-disclosure is zero.
Q4. Does AB 2013 require us to reveal proprietary or trade-secret training data?
No. AB 2013 requires disclosure of the nature and sources of training data, not the data itself. You disclose the category and origin (e.g., “licensed third-party text data from industry publications”) — not the raw data. Trade secrets remain protected under existing law.
Q5. What are the penalties for non-compliance with AB 2013?
Enforcement is directed through the CPPA and Attorney General’s office. Non-compliance can also trigger CCPA/CPRA violations (up to $7,500 per intentional violation) if personal data was involved in training. Early proactive compliance is far less costly than enforcement exposure.
Q6. We used publicly available web data. Is that enough to be compliant?
“Publicly available” doesn’t mean “legally cleared for AI training.” Many websites’ ToS prohibit scraping. Public data may still contain personal information or copyrighted content. You must document the scraping methodology and flag what was collected. “We got it from the internet” is not a compliance strategy.
Q7. Does AB 2013 apply to AI tools used only internally?
The law primarily targets publicly deployed systems. Strictly internal tools likely fall outside AB 2013’s immediate scope — but if it later makes decisions affecting customers or employees, CPPA ADMT rules or Civil Rights Council employment AI regulations may still apply.
Q8. Where do I publish the required disclosure? Is a PDF okay?
The law requires a publicly accessible webpage — online, indexed by search engines, no login required. A PDF could technically satisfy the requirement, but a dedicated HTML page at yourcompany.com/ai-transparency is strongly preferred for accessibility, searchability, and ease of updates.
Q9. How often do we need to update our training data disclosures?
Update whenever the model undergoes a significant change — retraining on new data, expanded dataset, or substantially altered methodology. Include a “last updated” date on your disclosure page and commit to reviewing it with every meaningful model change.
Q10. Is there an official government template for AB 2013 compliance?
As of early 2026, the CPPA has not released a standardized template specifically for AB 2013. The NIST AI Risk Management Framework provides a useful starting point. Monitor the CPPA’s official guidance page and the California Legislature’s information portal for authoritative updates.

🔗 Key Government Resources