AI Training Data Compliance: The Regulations, Standards and Records Data Buyers Need

Four tracks of obligation and the record each one demands

Every compliance question about acquired data falls into one of four tracks, with voluntary governance frameworks layered on top; each is answered by a filed document, not a supplier's assurance.

The provenance hub covers verifying that evidence; the Data Provenance Standards explainer covers metadata for carrying it.

Your role with the model decides which duties attach

One dataset can trigger no AI-specific duty for one buyer and several for another, because these laws attach to what you do with the model. Classify each intended use before acquisition.

One summary of the statute notes that its disclosure duty applies after a "substantial modification," but the law does not define the term [11], so fine-tuning a public model on licensed data may create your own disclosure duty (fine-tuning and provider duties). The research exclusion also ends at market placement, so capture disclosure facts while the supplier can still answer.

What Articles 10 and 53, AB 2013 and SB26-189 ask about the data

Article 10 governs high-risk training data, Article 53 sets copyright and transparency duties for general-purpose models, and two US state laws require written disclosures.

Article 10. Training, validation and testing sets for high-risk systems need data governance and management practices covering design choices, collection processes and data origin, preparation such as annotation and cleaning, assumptions, suitability, possible biases and data gaps, and must be relevant, sufficiently representative and, to the best extent possible, free of errors and complete [1]. Regulation (EU) 2026/1744, the Digital Omnibus on AI, published in the Official Journal on 24 July 2026 [12], reportedly moved the high-risk start dates shown below [13]; the Commission's AI Act Service Desk text, based on the consolidated version of 27 July 2026, marks Article 10(1) as amended [1].

Article 53. General-purpose AI (GPAI) model providers must keep a copyright policy that honors rights reservations under Article 4(3) of the Digital Single Market (DSM) Directive, and publish a training-content summary on the AI Office template [2]. The template, published 24 July 2025, applies to open-source models too [3]; a law-firm summary describes sections on general information, data sources by type and origin, and collection and processing [14]. See what buyers need from suppliers for EU summaries.

AB 2013. Documentation, due by 1 January 2026 and before each later release or substantial modification of systems released since 1 January 2022, covers dataset sources or owners, copyrighted or licensed material, personal information, collection periods and synthetic data [5] (AB 2013 supplier records).

SB26-189. Signed on 14 May 2026 to replace the consumer protections of SB 24-205, it requires developers of covered ADMT to give deployers documentation of intended uses, training data categories, known limitations and human-review guidance from 1 January 2027, enforced by the Attorney General [6], who released interim draft rules on 6 October 2026 [15] (SB 26-189 documentation).

Dates to plan around

See EU AI Act deadlines after the Digital Omnibus.

Copyright and TDM: where commercial training still needs permission

As of October 2026, no major jurisdiction gives commercial developers an unconditional right to train on accessible works; each exception carries a condition that decides the record you need. Exceptions are national, so record where each copy is made and why.

See the Copyright Office report and licensing, lawful access and pirated sources, EU opt-out checks under DSM Article 4 and whether to license or rely on fair use.

Privacy law: the de-identification standard decides which rules follow the data

Privacy compliance turns on whether the records are still personal data where you receive them, and under which standard they were de-identified; answers differ by jurisdiction.

• EU. Pseudonymised data that can be re-attributed remains personal data under the GDPR, and special categories such as health data need an Article 9(2) condition [22]; the AI Act leaves the GDPR unaffected [10]. EDPB Opinion 28/2024 says a model trained on personal data cannot be presumed anonymous and that legitimate interest must pass a three-step test [23]. As of September 2026, Digital Omnibus proposals to narrow the personal-data definition were not law [24].
• US health data. HIPAA de-identification uses Expert Determination or Safe Harbor removal of listed identifiers [25] (which method to require).
• California. CCPA "deidentified" status requires the holder to bind recipients by contract to its conditions, including no re-identification [26] (obligations a buyer inherits).
• Promises to customers and users. FTC staff warned in January 2024 that breaking promises not to train on customer data may violate FTC-enforced law, citing past orders to delete models built with unlawfully obtained data [27] (algorithmic disgorgement).

Where SourceX sources a dataset, names, emails and account numbers are removed or replaced before delivery, the method is recorded and a processed sample is checked; health records must meet HIPAA Safe Harbor or Expert Determination before a license is considered. No method is perfect, so classify the result yourself with the legal definitions compared and SourceX's guides to privacy and data laws.

Governance frameworks that turn duties into acquisition controls

Voluntary frameworks create no legal duties, but auditors and customers test against them.

• NIST AI RMF 1.0 (AI 100-1, January 2023) structures risk work under Govern, Map, Measure and Manage [7]; NIST says it is being revised as part of the White House AI Action Plan, with no revision published as of early October 2026 [28].
• NIST AI 600-1, the Generative AI Profile (July 2024), lists 12 risks, including data privacy, intellectual property, and value chain and component integration, with coded actions such as GV-1.1-001 [29]; acquisition controls fit the last two (NIST controls for acquired data).
• ISO/IEC 42001:2023 sets requirements for an AI management system, with controls in Annex A; certification is optional [8] (Annex A data and supplier controls).
• ISO/IEC 5259 covers data quality for machine learning: Parts 1 to 4 appeared in 2024 [30] and Part 5, on governance, in 2025 [31]; measures belong to the training data quality hub.

Start internal rules from the governance policy template, pre-acquisition risk scoring and the data governance definition.

AI training data compliance checklist per dataset

Run these nine checks before a dataset enters any pipeline; together they produce every record the tracks above ask for.

1. Classify use and role. Intended uses, and which roles above each one creates.
2. Name the source. System of record, legal owner, prior holders, collection period.
3. Prove the acquisition route. License, purchase or access terms; dated opt-out checks for web-derived content.
4. Fix permitted uses. Allowed uses, term, deletion duties and regulator access.
5. Classify personal data. Status per jurisdiction, de-identification standard and method, special categories.
6. Capture disclosure facts. IP status, synthetic share, data categories and volume, in publishable wording.
7. Log preparation. Annotation, cleaning and filtering, bias examination and known gaps.
8. Record location and retention. Where copies are made, and retention periods for data and records.
9. Record sign-off. Legal, privacy and security internal approvals against the dataset ID.

Illustrative example: invented to show structure; it does not describe an available dataset.

dataset_id: ds-0142
description: "Field service work orders, commercial HVAC, 2019-2024"
supplier: "Supplier B (US facilities services company)"
role_assessment:
  eu_gpai_provider: true
  eu_high_risk_use: false
  ab2013_developer: true            # model offered to the public in California
  co_sb26_189_developer: false
acquisition:
  route: direct_license
  license_ref: LIC-2026-031
  permitted_uses: [pre-training, fine-tuning, evaluation]
  prohibited_uses: [resale, retrieval_display]
  regulator_disclosure_allowed: true
  training_location: US
disclosure_facts:                   # feeds AB 2013, the EU summary and SB26-189 documentation
  source_owner: "Supplier B"
  ip_status: "owned by supplier; customer attachments removed"
  personal_information: "names and phone numbers replaced with surrogates"
  collection_period: "2019-01 to 2024-12"
  synthetic_share: 0.0
  data_categories: ["work orders", "technician notes", "parts usage"]
privacy:
  jurisdictions: [US-CA]
  standard: "Cal. Civ. Code 1798.140(m) deidentified"
  method: "NER plus rules, surrogate replacement; post-processing sample checked"
  contract_terms: ["no re-identification", "keep in deidentified form"]
art10_governance: not_applicable    # complete if the data feeds a high-risk system
retention:
  license_end: 2029-06-30
  deletion_duty: "delete data copies within 30 days of license end"
approvals: {legal: 2026-09-14, privacy: 2026-09-16, security: 2026-09-16}

Keep entries in a training data use register, package them for audit readiness, and ask suppliers through a due diligence questionnaire or due diligence checklist.

If you source operational data through SourceX, every dataset goes through rights review, which checks that the business owns or may share the records and that required consents are in place, and diligence materials on source, rights, preparation and allowed use are prepared per dataset for your review (legal framework, data governance). You can list the compliance records your review requires when you describe the data.

Start here: compliance guides by question

Each guide below takes one question from this map further; for sourcing and licensing beyond compliance, start from the AI data buyer's guide.

• How do we document data for the EU? The training-content summary for licensed datasets, the GPAI Code copyright chapter, Annex IV technical documentation and models trained outside the EU.
• What do US states require? US state AI laws that reach training data, model cards that reference licensed data and CCPA risk assessments for AI training.
• Does an exception cover our training? Japan Article 30-4, UK commercial training, Singapore's exception and China's training data rules.
• Sector rules: bank model risk, insurers' external data, FDA devices, export controls.
• Who signs off? In-house counsel review of AI data licenses.

Mistakes that create compliance debt

• Treating a security review as a compliance review. Encryption protects files; it does not show rights, legal basis or disclosure facts.
• Assuming an exception travels. The UK exception is non-commercial only [18]; Japan's yields where analysis licenses are sold [19].
• Signing a license that blocks disclosure. If it bars public description or regulator access, you may be unable to meet these duties.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Bring your compliance requirements to a data request

Describe the operational data you need and the records your compliance review requires. SourceX looks for US businesses that hold that data, checks the data and the supplier's licensing permissions, and manages a license that defines which records are included, what they can be used for and how delivery happens. Nothing is contracted until a supplier agrees, and a request does not guarantee a matching dataset. Submit your data and compliance requirements.