Copyright and AI Training Across Jurisdictions: Fair Use, TDM Exceptions and Opt-outs

Copyright decides whether an AI developer may copy protected works into a training set without permission, and the answer differs by jurisdiction: a fair use defense in the United States, a text and data mining exception with an opt-out in the EU, wide exceptions in Japan and Singapore, and no commercial exception in the UK, Switzerland, Norway or Canada. The EU exception does not apply to content whose rightholder reserved it in a machine-readable way (Directive (EU) 2019/790 Art. 4(3)), and in the US a court decides each training use on its record.

This page compares 18 jurisdictions as of 2026-11-09 for the training sense only, and it serves both sides: rightholders deciding whether to license, and developers deciding what they may copy. The comparison sits inside AI training data licensing, where a license is the contract that settles what an exception leaves open. Seven proceedings and three rule changes are dated in the final table.

Every rule below was read in the statute, official text or court record and carries its check date; how every rule is sourced and dated is set out on the methodology page.

Does Training an AI Model on Copyrighted Works Need a License?

It needs a license wherever no exception or defense covers the copies made for training, which is the case for commercial training in the UK, Switzerland, Norway and Canada, and for content whose rightholder reserved it in the EU. Training makes reproductions: Singapore's Copyright Act 2021 s.244(3) says storing a work counts as copying it, and the other laws here treat the training copy the same way.

An exception still binds its user to conditions. The EU exception needs lawful access and no reservation (Art. 4), Japan's Art. 30-4 stops at unreasonable prejudice to the owner, and Singapore's s.244 requires lawful access. Switzerland's Art. 24d URG covers research only, so commercial training there needs a license unless another provision applies. The United States has no exception at all; 17 U.S.C. 107 leaves the question to a court, so the same training run can be fair use on one record and infringement on another.

Copyright Rules for AI Training in 18 Jurisdictions

The table sets the 18 jurisdictions side by side with the provision, the model it follows, whether a rightholder can opt out, whether a contract can exclude the exception, and its status as of 2026-11-09. "Not stated" means no primary source is recorded for this page, not that no rule exists.

Four Legal Models for Copying Protected Works Into a Training Set

Four models result: a case-by-case defense, an exception with an opt-out, a wide exception without one, and a narrow exception or none.

United States: A Case-by-Case Defense Decided in Court

US law has no training exception; 17 U.S.C. 107 lets a court find a use fair after weighing four factors, so each training copy is judged on its record. The factors are the purpose and character of the use (including its commercial nature), the nature of the work, the amount and substantiality used, and the effect on the market. In Bartz v. Anthropic (N.D. Cal. 3:24-cv-05417-AMO), the order of 2025-06-23 (Dkt. 231) held training "exceedingly transformative" and fair use, held purchased and scanned print books fair use, and did not excuse the pirated library copies. The USD 1.5 billion settlement received final approval and judgment on 2026-07-20 (Dkt. 680), covering 482,460 works. In Kadrey v. Meta (N.D. Cal. 3:23-cv-03417-VC), the order of 2025-06-25 (Dkt. 598) granted summary judgment to Meta on that record and says it "does not stand for the proposition that Meta's use ... is lawful".

The Third Circuit affirmed on 2026-09-29 (No. 25-2153) that ROSS's use of 2,243 Westlaw headnotes was "minimally transformative, at best", commercial and not fair use, and noted that ROSS's AI "was not a generative AI". In In re OpenAI Copyright Infringement Litigation (S.D.N.Y. MDL No. 3143), the United States filed a Statement of Interest on 2026-09-01, and no summary judgment ruling was located as of 2026-11-09.

Statutory damages under 17 U.S.C. 504(c), in force since 1978-01-01, run from USD 750 to 30,000 per work, up to USD 150,000 if willful and down to USD 200 if innocent. The Copyright Office's Part 3 report exists only as a pre-publication version of 2025-05-09, and it says knowing use of pirated or illegally accessed works weighs against fair use without being determinative (p. 52); the Office's work is covered in the US Copyright Office report on generative AI training.

European Union: An Exception the Rightholder Can Reserve

Directive (EU) 2019/790 allows text and data mining for any purpose by anyone with lawful access unless the rightholder reserved it, and for online content the reservation must be machine-readable (Art. 4(3)). Art. 3 lets research organizations mine without any opt-out, and member states had to transpose both articles by 2021-06-07. The detail of the exception and its reservation sits in the page on the EU text and data mining exception and its opt-out.

National wording varies. Germany requires an online reservation to be machine-readable (UrhG § 44b(3)), France uses "an appropriate manner" (CPI L122-5-3 III, in force 2021-11-26) and Sweden says "på lämpligt sätt" (15 a §). Regulation (EU) 2024/1689 Art. 53(1)(c) obliges general-purpose AI providers, from 2025-08-02, to keep a copyright policy that honors Art. 4(3) reservations, and the General-Purpose AI Code of Practice, published 2025-07-10, lists respect for robots.txt under RFC 9309 as Measure 1.3, a voluntary commitment. The training content summary required by Art. 53(1)(d) follows template C(2025) 8311 of 2025-12-05, which lists licensed datasets only as Yes or No plus modality, so it names no licensor.

Two decisions are pending. The Advocate General's opinion in C-250/25 Like Company v. Google (2026-11-03) treats the training questions as inadmissible because they are hypothetical, and says in passing (obiter) that training is as a rule permitted under the exception; the judgment has not been delivered. The Federal Court of Justice (BGH, I ZR 281/25, Kneschke v. LAION) pronounces its decision on 2026-12-17.

Japan and Singapore: Wide Exceptions With No Opt-out

Japan and Singapore permit copying for analysis or training by commercial users without any reservation mechanism, but each sets its own limit: Japan a proviso, Singapore lawful access. Japan's Art. 30-4 (Act No. 48 of 1970, replaced by Act No. 30 of 2018, in force 2019-01-01) permits use for information analysis under item (ii), except where the use would unreasonably prejudice the owner. The Agency for Cultural Affairs' General Understanding of May 2024 is non-binding and reads the proviso to reach paid analysis-ready databases and circumvention of robots.txt or ID-password measures; it says collecting from known piracy sites "should be strictly avoided" (p. 11). Japan therefore does not "allow everything".

Singapore's Copyright Act 2021 (Act 22 of 2021) s.243(b) permits using a work as an example to improve a computer program, in force 2021-11-21. Section 244 requires lawful access, and its illustrations at s.244(2)(d) cover access that depends on a paywall or database terms. Section 187(1)(c) and (3) make a contract term that excludes the exception void, whether the contract was made before, on or after 2021-11-21.

Research-Only Exceptions or None: UK, Switzerland, Norway, Canada, Hong Kong

In these five jurisdictions commercial training on protected works has no statutory exception, so it needs a license. In the UK, CDPA 1988 s.29A (inserted by S.I. 2014/1372, in force 2014-06-01) covers non-commercial research only, and s.29A(5) makes contract terms against it unenforceable. The government's Report on Copyright and AI of March 2026 says a broad exception with an opt-out is "no longer the government's preferred way forward". It counts 11,520 responses, of which 81 % preferred option 1 (licensing in all cases) and 10 % option 0, and it proposes no intervention in the licensing market at this stage.

Switzerland's Art. 24d URG (SR 231.1, in force 2020-04-01) is research-only and excludes computer programs in 24d(3). Motion 24.4596 asks to exclude AI services from the URG exceptions and to require rightholder consent for generative AI; it was transferred to the Federal Council on 2025-12-11, and no bill is published as of 2026-11-09. Norway's Act LOV-2018-06-15-40 (in force 2018-07-01) holds no TDM provision, as read on 2026-11-09, and offers extended collective licensing under s.63. Canada's Copyright Act (R.S.C. 1985, c. C-42) has no TDM exception, and s.29 fair dealing is limited to listed purposes such as research, private study, education, parody and satire. Hong Kong consulted from 2024-07-08, and no amendment bill is enacted as of 2026-11-09.

Pending or Untested: Brazil, India, Taiwan, China and Open Rows

In the remaining jurisdictions the rule is a bill, an interim ruling or an untested general clause. Brazil's PL 2338/2023 passed the Senate on 2024-12-10 and sits in the Chamber as "Aguardando Parecer" as of 2026-10-27; it would require a summary of copyrighted content used, a right to prohibit use outside non-commercial research, and remuneration, and it is not law. In ANI v. OpenAI (Delhi High Court, CS(COMM) 1028/2024), the interim injunction was refused on 2026-07-24 on a prima facie reading that training is fair dealing under s.52(1)(a), and the suit is pending. Taiwan's Copyright Act (last amended 2022-06-15) has no TDM provision, so training relies on the general fair use clause in Art. 65, which is untested. China's Interim Measures for Generative AI Services (in force 2023-08-15) require training data from lawful sources without IP infringement.

South Korea, Australia and Israel stay "not stated" in the table until an official text is read for each.

What an Opt-out and a Contract Can Change, Model by Model

A reservation has legal effect only where the law gives it one: it removes content from the EU exception, but it does not change a US fair use analysis, and Japan and Singapore have no reservation mechanism at all. The table compares six rules as of 2026-11-09, reading the contract question the same way for each.

Contract override splits the models: Art. 7(1) of the Directive protects only Arts. 3, 5 and 6, so a term can switch off Art. 4, while Singapore and the UK void such terms and Finland protects research copies. The technical means of reserving content, and of charging for it, are covered in the page on AI crawlers and how to block or charge them.

Where Training Took Place and What the Model Retains

Copyright is territorial, so a claim depends on where the training copies were made and whether the trained model itself holds copies. In Getty Images v. Stability AI ([2025] EWHC 2863 (Ch), claim IL-2023-000007, 2025-11-04), Getty abandoned its Training and Development Claim because it had no evidence that training took place in the UK. The court dismissed secondary infringement, holding that the model is not an "infringing copy" under CDPA ss.22, 23 and 27(3), and found trade mark infringement in part. Getty received permission to appeal on 2025-12-16, and the Court of Appeal (CA-2026-000007) has fixed the appeal hearing for 2026-11-11 (Court of Appeal case tracker, checked 2026-11-10).

GEMA v. Suno (LG München I, 42 O 763/25, judgment 2026-07-31, not final) applied US law to the US training copies and held them not fair use, then held that memorization in Suno's v3.5 and v4 models on German servers infringes § 16 UrhG, with § 44b not applicable. Switzerland's Motion 24.4596 takes the same territorial line in point 3, applying Swiss law when content is offered in Switzerland. The two UK and German outcomes differ because one court found no training in the UK while the other found copies stored in Germany.

What Courts and Governments Say About a Market for Training Licenses

Two US courts read the licensing market differently: Kadrey found authors not entitled to a market for licensing works as AI training data, while the Third Circuit found the market for licensing headnotes to train AI "rapidly developing". The Kadrey order (Dkt. 598, 2025-06-25) rests on its record, and the Third Circuit's opinion of 2026-09-29 treats ROSS's use as non-generative and commercial. The tracker of AI copyright lawsuits over training data records each docket.

Governments lean against intervention for now. The Copyright Office's Part 3 pre-publication report (2025-05-09, p. 106) would let the licensing market develop without intervention and consider an extended collective license only where market failure is shown. The UK report of March 2026 (Section G, para. 63) proposes no intervention in the licensing market at this stage, and the Creative Content Exchange pilot was announced on 2026-01-26, with an operational pilot planned by summer 2026 (later status not verified).

Provenance runs through all of it. Bartz did not excuse pirated library copies, the Copyright Office says knowing use of pirated works weighs against fair use, and the ACA guidance says collecting from piracy sites should be avoided.

Collective routes exist in several places. Norway's s.63 provides extended collective licensing, and the Kopinor agreement with the National Library of Norway (2025-12-19) sets NOK 45,000,000 for 2026, indexed to consumer prices from 2027; the library is a public institution, so the sum is not a commercial developer's payment. STIM runs a collective AI license pilot with Songfox and Sureel for opted-in works only, and from 2026-06-01 its members approve works one by one (STIM states creators receive royalties during training, use and output use, checked 2026-11-09). The European Parliament's non-binding resolution of 2026-03-10 (P10_TA(2026)0066) calls for tools to exclude works and favors voluntary collective licensing, and the licensors involved (the Copyright Licensing Agency Limited with its Generative AI Solution, opt-in through ALCS and PLS, and GEMA with PLAI by GEMA, which publishes no remuneration figures, checked 2026-11-09) are described in collective licensing for AI training.

Licensing Works for Training: Benefits, Risks, Downsides and Alternatives, Including Not Licensing

A rightholder can license works for training in every jurisdiction here, but what a license adds depends on the local model, and not licensing remains a valid choice. The lists apply to the rightholder or data holder; the developer's view is the mirror image.

Benefits

• A license is the lawful route to training use in the UK, Switzerland, Norway and Canada, where no commercial exception exists.
• In the EU, a machine-readable reservation combined with a license works because Art. 7(1) lets a contract exclude Art. 4.
• A license documents provenance, which matters where courts and agencies treat pirated sources as weighing against the user.

Risks

• In Japan and Singapore an exception makes a license optional for the developer, which can limit what a rightholder obtains.
• US fair use findings on a record (Bartz, Kadrey) show that a court may allow training without a license.
• A seller that does not own all rights in the data cannot license what it does not hold, as the rights audit explains.

Downsides

• Auditing ownership and operating a reservation cost time and money before any license is signed.
• The AI Act training summary names no licensor, so a license does not produce public credit; the template C(2025) 8311 asks only Yes or No and a modality for licensed datasets.

Alternatives

• Reserve the content and do not license it.
• Join a collective license such as an extended collective scheme or STIM's pilot.
• License only works the holder owns outright.
• Do not license at all, which the page on should you sell or license your data to AI companies treats as a valid outcome.

Decisions Still Pending as of November 2026

Seven open proceedings and three pending rule changes can move these rules; each row gives the next dated step and its source. The table lists them by next dated step as of 2026-11-09, with the ROSS row re-read on 2026-11-10.

The next fixed dates are the BGH pronouncement on 2026-12-17 and the ROSS rehearing petition filed on 2026-11-03; the other rows carry no scheduled date beyond the Kadrey related trial in 2027.

Beyond This Comparison: Jurisdiction Pages, Court Records and the Seller Process

Each jurisdiction, case and seller step has its own page.

• AI training data law by jurisdiction: privacy, employee and contract rules beside copyright.
• EU rules for AI training data: the other EU instruments.
• EU AI Act Article 53: the provider's copyright policy and training summary.
• Rights audit before licensing: checking that the seller owns what it licenses.
• Sell or license company data to AI companies: the full seller process.
• United States AI training data law: federal and state rules.
• Bartz v. Anthropic and the settlement: the settlement and its claims process.
• Machine-readable TDM reservations: reservation formats.
• Pirated books in AI training: shadow libraries and provenance.
• Is web scraping for AI training legal: scraping beyond copyright.

What This Page Does Not Cover

This page covers copies made to train AI models, not whether AI outputs can be protected by copyright or who owns them. Removal tools and the reuse of generated media are outside this reference too.

How This Page Is Sourced and Dated

Every rule here cites the statute, official text or court record, with the date it was checked. The rules for sources and dates are set out in how we source, date and check every figure. This page is general information about copyright law, not legal advice for any specific deal.

Frequently Asked Questions

Are there AI copyright laws?

Yes, as exceptions in the EU, Japan, Singapore and the UK (s.29A), but not as an AI-specific statute in the US. The US decides training under fair use (17 U.S.C. 107), and the EU rule is Directive (EU) 2019/790 Arts. 3 and 4.

Is AI breaking copyright laws?

It depends on the jurisdiction and the record. Bartz and Kadrey found fair use on their records, the Third Circuit found ROSS's non-generative use not fair use on 2026-09-29, and the Munich judgment of 2026-07-31 is not final.

Which countries allow AI training without a license?

Japan within its proviso, Singapore with lawful access, and the EU unless the rightholder reserved the content. Each of these exceptions has conditions, so none permits training on everything.

Does blocking AI crawlers in robots.txt count as a legal opt-out?

In the EU a machine-readable reservation counts as an opt-out, and robots.txt is the format named in the voluntary Code of Practice Measure 1.3. In Japan the ACA reads circumvention of robots.txt into the proviso, and in the US a block has no effect on a fair use analysis.

Can a rightholder stop training that already happened?

No rule compared here gives a retroactive right to remove a work from a trained model. A reservation acts on future mining, and courts award damages or injunctions case by case.