AI Training Data Law by Jurisdiction: Copyright, Privacy, Employees and Contracts

AI training data law is the set of copyright, privacy, employment and contract rules that decide whether data may be licensed, sold or used to train AI models, and it differs by jurisdiction. No jurisdiction recorded here bans the trade outright, but the conditions differ sharply: the EU lets rightholders reserve text and data mining (Directive (EU) 2019/790, Art. 4(3)), the UK allows it only for non-commercial research (CDPA 1988 s.29A), and Canada has no text and data mining exception at all.

Every deal in AI training data licensing answers four questions: who holds the rights in the content, what legal basis covers any personal data, which staff and recording rules apply, and what the contract, trade-secret and export rules allow after delivery. This page answers them for 52 jurisdictions and 10 legal topics, each with its official citation, as of November 9, 2026, and lists the 14 dated changes of 2026.

Is It Legal to Sell or License Data for AI Training?

Selling or licensing data for AI training is legal in every jurisdiction in the table when the seller holds the rights, has a legal basis for any personal data and respects employee, recording and contract limits; no jurisdiction recorded here bans it outright. The legality sits in four layers, and a deal has to clear each one:

• Copyright and database rights: whether the seller owns the content or has the right to license it for training.
• Personal data: whether a legal basis, notice and, where needed, consent cover disclosure to a buyer.
• Employee and recording rules: works council rights, staff-data limits and recording consent.
• Contract, trade secret and export limits: customer confidentiality clauses, unfair-terms rules and export controls.

The same dataset can pass in one country and fail in another. Employee email in Germany falls under GDPR Art. 88, and the Federal Labour Court held on 2025-05-08 (BAG 8 AZR 209/21) that § 26(1) BDSG does not meet Art. 88 and stays unapplied, so an employer needs another legal basis. A recorded sales call under US federal law needs only one party's consent (18 U.S.C. 2511(2)(d)), while California requires all parties to agree (Cal. Penal Code 632).

AI Training Data Law by Topic

Ten legal topics decide most AI data deals, and each has a page that compares jurisdictions. The table lists the core instruments per topic and what each decides in a data deal, as of 2026-11-09.

The topics stack rather than replace each other. A rightsholder's license covers copyright but supplies no legal basis for personal data in the same records, and a privacy basis does not lift a works council's rights over staff emails. Export controls are covered for the United States only.

AI Training Data Law by Jurisdiction: The Complete Table

The table covers 52 jurisdictions: the EU, its 27 member states and 24 other jurisdictions, each with key instruments, the TDM rule, the privacy regime and employee and recording notes. Rows run from the EU through the member states, then the other jurisdictions, each A to Z, as of 2026-11-09:

• "Not recorded" means no instrument was confirmed in an official source, not that no law exists.
• "Not stated" means a value exists in secondary reports but is withheld until an official source confirms it.
• GDPR (Regulation (EU) 2016/679) applies in every member-state row, in addition to the national rules shown.
• Citations point to the official legislation portal, gazette, regulator or court.

How the Rules Differ Between Jurisdictions

The table shows each rule; this section shows the pattern behind it, one legal question at a time, in the order a deal meets them.

Where Training on Protected Works Needs a License

Training on protected works needs a license wherever no exception covers it, and exceptions range from US fair use to EU opt-outs to none at all. In the United States, 17 U.S.C. 107 decides fair use case by case, and statutory damages under 17 U.S.C. 504(c) run from USD 750 to 30,000 per work, up to 150,000 if wilful (amounts as of the 2024 Code). Kadrey v. Meta (N.D. Cal. 3:23-cv-03417-VC) decided the training claim for Meta on its record on 2025-06-25. Bartz v. Anthropic (N.D. Cal. 3:24-cv-05417-AMO) ended in a USD 1.5 billion non-reversionary settlement with final approval on 2026-07-20 (Dkt. 680), a settlement rather than a ruling on all training. The Copyright Office's Part 3 report exists only as a pre-publication version of 2025-05-09.

In the EU, Art. 4 of Directive (EU) 2019/790 allows commercial text and data mining unless the rightholder reserves it, machine-readably for online content. National laws word the reservation differently: Germany, the Netherlands, France, Spain and Austria name machine-readable means, while Sweden, Denmark, Finland and Slovakia require only an express reservation. AI Act Art. 53(1)(c) obliges model providers to honor these reservations, and Measure 1.3 of the GPAI Code of Practice (2025-07-10) points to robots.txt under IETF RFC 9309, which is how publishers block AI crawlers and how to block or charge them in practice.

Outside the EU, Japan's Art. 30-4 (in force 2019-01-01) permits non-enjoyment use with a proviso against uses that unreasonably prejudice the owner, Singapore's ss.243-244 cover commercial users, and the UK's s.29A stays non-commercial; the UK government's March 2026 Report on Copyright and AI calls an opt-out exception "no longer the government's preferred way forward". Four models result:

• Fair use, decided case by case: United States, with Taiwan's untested Art. 65.
• Opt-out exception: the EU and its member states.
• Closed exception: Japan (proviso), Singapore (lawful access), UK (non-commercial research only), Switzerland (research copies).
• No exception: Canada, Norway and Hong Kong, where training on protected works needs a license.

Legal Basis for Personal Data in a Training Set, Country by Country

Personal data in a training set needs a legal basis where the seller is established before it is disclosed, and the models range from GDPR's six bases to China's separate consent. GDPR (Regulation (EU) 2016/679, applying from 2018-05-25) adds stricter conditions for special categories in Art. 9, and the EDPB's Opinion 28/2024 of 2024-12-17 holds that models trained on personal data are not anonymous in all cases. The detail for the EU sits on the page on GDPR and AI training data.

• United Kingdom: recognised legitimate interests, Art. 6(1)(ea) UK GDPR, in force 2026-02-05 under the Data (Use and Access) Act 2025.
• China: notice plus separate consent under PIPL (in force 2021-11-01), with fines up to CNY 50 million or 5% of turnover.
• Canada: PIPEDA s.7.2(4) excludes transactions whose primary purpose is the sale of personal information from the business-transaction exception.
• Mexico: the LFPDPPP, in force since 2025-03-21, allows transfer under the privacy notice or consent and exempts dissociated data.
• United States: sectoral rules, such as HIPAA Safe Harbor's 18 identifiers (45 CFR 164.514), plus state laws; under the CCPA, licensing personal information for money is a sale.

The FTC staff warned on 2024-02-13 that retroactive, surreptitious changes to privacy terms to allow AI training may be unfair or deceptive. No privacy instrument is recorded for Australia, Brazil, Chile, India, Israel, New Zealand, Norway or Vietnam, and those cells stay empty.

Staff Communications, Works Councils and Recorded Calls

Employee emails, chats and call recordings carry rules on top of privacy law: works council rights in several EU countries and recording-consent rules everywhere. In Germany, technical monitoring falls under co-determination (BetrVG § 87(1) Nr. 6), on top of the BAG ruling described above. Switzerland limits employee data to job suitability or the contract (OR Art. 328b), and Spain requires the works council to be informed on algorithms (RDLeg 2/2015 art. 64.4.d).

Recording rules travel with the audio. Germany punishes unauthorized recording with up to 3 years (§ 201 StGB), Austria with up to 1 year or 720 daily rates (§ 120 StGB), and Switzerland makes recording without all participants' consent punishable (StGB Art. 179bis, 179ter); none of the three is a one-party-consent country. The US federal rule is one-party consent (18 U.S.C. 2511(2)(d)), but 2511(1)(c)-(d) makes disclosing unlawfully intercepted contents an offence in itself, which reaches a later sale. A staff archive therefore also needs employee notice and consent for internal communications.

• Consent: Netherlands (WOR art. 27(1)(k), (l)); Austria (ArbVG § 96a, replaceable by the Schlichtungsstelle).
• Consultation: France (Code du travail L2312-8, information-consultation, not consent).
• Negotiation: Sweden (MBL 11 §, duty to negotiate before important changes).

Contract Terms, Confidential Datasets and Export Rules

Contract, trade-secret and export rules decide what a buyer may do with licensed data after delivery. Customer contracts and confidentiality duties can also bar the license itself.

• Unfair terms: the EU Data Act (Regulation (EU) 2023/2854, applying from 2025-09-12) voids unfair terms unilaterally imposed on another enterprise under Art. 13, but excludes the main subject matter and price (Art. 13(8)); the Commission's model terms, Recommendation (EU) 2026/2394 of 2026-10-29, are non-binding.
• Trade secrets: the DTSA (18 U.S.C. 1836, 1839) sets a 3-year limitation, Directive (EU) 2016/943 allows up to 6 years (Art. 8(2)), the UK applies SI 2018/597, and Japan's UCPA Art. 2(7) protects shared data with limited access.
• Export and foreign buyers: the DOJ bulk data rule (28 CFR Part 202, effective 2025-04-08) restricts transfers to countries of concern above volume thresholds, for example health data on more than 10,000 US persons; the EAR treats release to foreign staff as a deemed export (15 CFR 734.13), and ITAR covers defense technical data (22 CFR 120.33, 120.50).
• Voice and likeness: Tennessee's ELVIS Act has applied since 2024-07-01, while the federal NO FAKES Act (S. 4591, introduced 2026-05-20) is still a bill.

Disclosure Duties That Make a Licensed Dataset Visible

A licensed dataset can become visible after the deal, because some laws make AI developers disclose what they trained on. Under AI Act Art. 53(1)(d), providers publish a training content summary on the Commission's template, C(2025) 8311 final of 2025-12-05, which shows licensed datasets only as Yes or No plus modality, never the licensor's name. The AI Office can enforce it from 2026-08-02, with fines up to 3% of worldwide turnover or EUR 15,000,000, whichever is higher (Art. 101(1)). California's AB 2013 (Cal. Civ. Code 3110-3111) has applied since 2026-01-01; its challenge is on appeal in the Ninth Circuit (No. 26-1591), and an amicus brief there lists whether datasets were purchased or licensed among its disclosure categories. Connecticut's P.A. 25-113 requires privacy notices from 2026-07-01 to say whether personal data is used to train large language models.

What Changed in 2026 and What Is Still Pending

Fourteen dated events in 2026 changed or may change the rules, and pending items are marked as pending. The table lists them newest first, with status as of 2026-11-09.

Pending and final stay apart in this table. Bartz v. Anthropic is final, while Kneschke v LAION, C-250/25 and Getty Images' appeal against the High Court judgment of 2025-11-04 ([2025] EWHC 2863 (Ch)) are pending; Spain's AI bill lapsed, and the EU's Digital Omnibus proposal on data and GDPR (COM(2025) 837 final) is still a proposal. Rulings and settlements are tracked in the case tracker of AI copyright lawsuits over training data.

Benefits, Risks, Downsides and Alternatives of Licensing Under These Rules, Including Not Selling

Licensing under these rules can turn rights the law already reserves into income, but each layer adds cost and some breaches are criminal, so not licensing stays a valid outcome.

Benefits

• License income where the law requires a license: in the UK (s.29A non-commercial only), Canada and Norway, and for EU content with a valid Art. 4(3) reservation.
• Documented consent and works-council agreements make the dataset easier for a buyer to accept.
• A recorded reservation keeps the seller's bargaining position against crawlers that honor it.

Risks

• Recording offences: passing on unlawfully recorded calls can be a crime, for example under § 201 StGB in Germany.
• Regulator action: the Garante's decision n. 755 of 2024-11-02 fined OpenAI EUR 15 million, and the Tribunale di Roma upheld OpenAI's opposition in judgment n. 4153/2026 (published 2026-03-18); the Irish DPC announced an inquiry into Grok training on 2025-04-11.
• Foreign buyers: the DOJ bulk data rule can bar a sale of US sensitive data above its thresholds.

Downsides

• Cost: a rights audit, privacy notices, de-identification and works-council procedures come before any income.
• Visibility: the AI Act summary and laws such as California's AB 2013 can reveal that a dataset was licensed.

Alternatives

• License only de-identified or non-personal data, which removes the personal-data layer.
• Keep the data and reserve text and data mining rights, without licensing.
• Join a collective license: the Kopinor agreement with the National Library of Norway of 2025-12-19 sets a fee of NOK 45,000,000 for 2026, paid by a public library rather than a commercial AI developer.
• Do not sell or license at all; the options are compared on the page on alternatives to selling data to AI labs.

Where the Law Meets the Seller Process, the Buyers and the Courts

The law tables answer what is allowed; the pages below apply it to a sale, a buyer and a contract.

• How to sell or license company data to AI companies.
• Who buys AI training data, every buyer and program.
• Data assets in insolvency, sold under court supervision.
• The rights audit before licensing, which checks who owns each dataset.
• The EU AI Act Article 53 training data summary and the copyright policy behind it.

How This Law Reference Is Sourced and Dated

Every rule on this page cites the official text or a court or regulator record and the date it was checked. Pending bills are never written as law. This page is general information about the law, not legal advice for any specific deal. The rules for sources and dates are set out in how we source, date and check every figure.

Frequently Asked Questions

Is it illegal to sell data for AI training?

No jurisdiction recorded here bans selling data for AI training outright. Specific data is barred instead: Maryland's MODPA has banned the sale of sensitive data since 2025-10-01, the DOJ bulk data rule limits transfers to certain foreign buyers, and unlawful recordings cannot be passed on.

Is it legal to sell data to AI companies in California?

It is legal in California when the CCPA's rules for a sale of personal information are met, because licensing personal information for money counts as a sale. Developers also disclose training data under AB 2013 from 2026-01-01. The detail is on the page on California CCPA and AB 2013.

Can AI companies train on publicly available data without a license?

Public availability alone does not give a license. In the EU, Art. 4 of the DSM Directive allows commercial mining of online content unless it carries a machine-readable reservation; in the US, Van Buren v. United States (2021-06-03) limits computer-misuse claims to gates that are closed, while copyright still applies.

Do AI companies have to name the data they licensed?

No, not under the EU template. The AI Act training content summary (C(2025) 8311 final of 2025-12-05) discloses licensed datasets only as Yes or No plus modality, so the licensor stays unnamed unless the contract or another law requires disclosure.

Can a contract switch off a text and data mining exception?

In the EU it can, because DSM Art. 7(1) protects only Arts. 3, 5 and 6 against contract override, not Art. 4. Singapore voids such terms (s.187(1)(c)) and the UK makes them unenforceable (s.29A(5)). The EU text and data mining exception page covers the opt-out.

Does Japan allow all AI training?

No: Art. 30-4 of the Copyright Act permits non-enjoyment use but excludes uses that unreasonably prejudice the copyright owner. The Cultural Council's General Understanding on AI and Copyright (English overview, May 2024) interprets the proviso and is non-binding.