Generative AI Training Data Transparency Act obligations mapped to SWT3 witness procedures. Dataset disclosure requirements, data provenance tracking, and copyright attestation for GenAI developers.
Who this is for: GenAI developers required to disclose training data, compliance officers managing AB 2013 documentation, legal counsel advising on training data transparency and trade secret risk, and data governance teams tracking dataset provenance.
Enforcement: January 1, 2026 (in effect). AB 2013 applies to all generative AI systems or services released on or after January 1, 2022, and made available to Californians. Developers must post high-level training data summaries on their websites. xAI has filed a federal lawsuit challenging the statute on First and Fifth Amendment grounds. Enforcement continues during litigation.
California Assembly Bill 2013, officially the Generative AI Training Data Transparency Act (TDTA), creates the first US state-level mandate for developers to publicly disclose what data their generative AI models were trained on.
The law requires developers to publish a "high-level summary" of the datasets used to develop or train their generative AI systems or services. Key provisions:
AB 2013 does not require disclosure of trade secrets or proprietary information. Developers may describe datasets at a summary level without revealing specific data sources that constitute trade secrets.
AB 2013 requires developers to address each of the following 12 categories in their public disclosure:
| # | Disclosure Category |
|---|---|
| 1 | Sources or owners of the datasets |
| 2 | Description of how the datasets further the intended purpose of the AI system |
| 3 | Number of data points in each dataset |
| 4 | Whether the data is protected by copyright, trademark, or patent |
| 5 | Whether the data is entirely in the public domain |
| 6 | Whether the developer purchased or licensed the data |
| 7 | Whether the datasets include personal information |
| 8 | Whether the datasets include aggregate consumer information |
| 9 | Whether the generative AI system uses synthetic data in its training |
| 10 | Date ranges during which the data was collected |
| 11 | Whether the data was collected from the internet |
| 12 | Whether the developer obtained consent for the data use |
Each AB 2013 disclosure obligation maps to one or more SWT3 witness procedures. Anchors generated by these procedures serve as cryptographic evidence that the required documentation exists, has been verified, and is being maintained.
| AB 2013 Category | SWT3 Procedure | What It Witnesses |
|---|---|---|
| Data sources (1) | AI-DATA.1 |
Data provenance documentation status |
| Dataset purpose (2) | AI-SBOM.1 |
AI bill of materials with component descriptions |
| Copyright/IP status (4-5) | AI-LIC.1 |
License and IP status tracking |
| Purchased/licensed (6) | AI-DATA.2 |
Data licensing verification |
| Personal information (7-8) | AI-DATA.3 |
Personal data presence attestation |
| Synthetic data (9) | AI-MDL.5 |
Model weights and synthetic data method attestation |
| Public disclosure (all) | AI-TRANS.1 |
Transparency disclosure evidence |
| Bias from training data (implicit) | AI-FAIR.1 |
Bias measurement from data composition |
Categories 3 (data point counts), 10 (date ranges), 11 (internet collection), and 12 (consent) are captured as factor values within the AI-DATA.1 and AI-DATA.3 procedures. The SWT3 Witness Anchor records the attestation timestamp and factor hash, providing immutable proof that these details were documented at a specific point in time.
AB 2013 requires: Disclose sources or owners of datasets and describe how they further the system's intended purpose.
SWT3 witness behavior: Witnesses data provenance documentation at every training pipeline run. Factor A records the number of documented sources. Factor B records the number of sources with verified provenance. The verdict is PASS when all sources are documented and verified.
Each anchor includes a SHA-256 fingerprint that is locked to the tenant, procedure, and factor values at the moment of attestation. This means any change to the provenance documentation produces a different fingerprint, creating an auditable trail of documentation updates.
Assessor Tip
AI-DATA.1 anchors with continuous PASS verdicts demonstrate ongoing provenance tracking, not just a one-time disclosure filing. Check the anchor timeline to verify the developer is maintaining provenance records as datasets change.
AB 2013 requires: Disclose whether data is protected by copyright, trademark, or patent, whether it is in the public domain, and whether it was purchased or licensed.
SWT3 witness behavior: Witnesses license status verification for each dataset component. Factor A records the total number of components. Factor B records the number of components with a verified license status. The verdict is PASS when all components have documented license determinations.
This procedure directly addresses the three IP-related disclosure categories (4, 5, and 6) by requiring that each dataset component has a recorded license classification: copyrighted, public domain, purchased, licensed, or other.
Assessor Tip
Compare AI-LIC.1 attestation counts against the dataset component count in the public disclosure. Gaps between Factor A and Factor B indicate undocumented components that may represent incomplete disclosure.
AB 2013 requires: All 12 categories of information must be publicly posted on the developer's website.
SWT3 witness behavior: Witnesses the existence and completeness of public disclosure documentation. Anchors prove when the disclosure was first published and when it was last verified as live and accessible.
Because AB 2013 requires ongoing public availability (not just a one-time filing), periodic AI-TRANS.1 attestations create a verifiable record that the disclosure remained accessible over time.
Assessor Tip
The anchor timestamp on AI-TRANS.1 proves the disclosure was live on or before the compliance deadline. Compare the earliest anchor date against the January 1, 2026 effective date to confirm timely compliance.
| Examiner Question | Where to Look |
|---|---|
| Are training data sources documented? | AI-DATA.1 anchors --- provenance status per pipeline run |
| Is copyright/IP status tracked? | AI-LIC.1 --- license verification count vs component count |
| Is personal data presence disclosed? | AI-DATA.3 --- personal data attestation |
| Is the public disclosure live? | AI-TRANS.1 --- disclosure timestamp and verification |
| Is synthetic data usage documented? | AI-MDL.5 --- synthetic data method attestation |
| Are bias risks from data assessed? | AI-FAIR.1 --- disparity measurement from data composition |
Install the SDK and start generating witness anchors for your training data pipeline:
# Install the SDKpip install swt3-ai# Witness training data provenancefrom swt3_ai import Witnesswitness = Witness(api_key="axm_live_xxx")# Attest data provenance for a training runwitness.witness_rag_context( corpus_id="training-v3", chunk_count=1_200_000, corpus_present=True)
For full SDK documentation, see the SDK Docs. To create a free account and get an API key, visit Sign Up.