AI Copyright Laws by Country: 2026 Global Guide
No treaty sets a single global rule for AI and copyright. The Berne Convention harmonizes the basics across more than 180 countries, but each nation decides two fast-moving questions on its own: can AI-generated work be owned, and may AI be trained on copyrighted material? The answers diverge sharply.
Information last verified on 2026-06-25. This article presents general legal information, not legal advice. AI and copyright law is changing quickly; several regimes described here are under active reform.
Jurisdiction scope: This hub compares how major jurisdictions treat two questions, the copyrightability of AI-generated output and the legality of AI training, across the United States, the European Union, the United Kingdom, Japan, Canada, China, and Australia. It is a general comparison; see each country guide for the detail and citations. For the United States in depth, see AI copyright law in the United States.
Is there a global law on AI and copyright?
No. Copyright is territorial, meaning each country applies its own law to uses that happen on its soil, and there is no international treaty written for artificial intelligence. What exists instead is a shared floor. The Berne Convention requires its members, more than 180 countries, to protect original works automatically on creation, without registration formalities, and to recognize foreign authors. The WTO TRIPS Agreement adds that computer programs are protected as literary works under Berne. The World Intellectual Property Organization administers these instruments but has not produced an AI-specific treaty.
That floor leaves the hardest modern questions to national law, and the gaps are wide. Whether a picture or block of code generated by AI can be owned, and whether a developer may train a model on copyrighted books, images, or repositories, are answered very differently in Tokyo, Brussels, London, and Washington. The two sections that follow map those two axes.
Can AI-generated work be copyrighted? A country comparison
The dominant rule is that copyright protects human authorship, so purely AI-generated output belongs to no one. The two notable exceptions run in opposite directions: the United Kingdom protects computer-generated works by statute, and China has protected AI-assisted output through its courts.
| Country or region | Purely AI-generated work copyrightable? | Basis |
|---|---|---|
| United States | No | Thaler v. Perlmutter (D.C. Cir. 2025); Copyright Office (2025) |
| European Union | No | "author's own intellectual creation" (CJEU, Infopaq, C-5/08) |
| United Kingdom | Yes, 50-year term, but contested | CDPA 1988, s.9(3) and s.12(7) |
| Japan | No | Agency for Cultural Affairs guidance (2024) |
| Canada | No, requires human "skill and judgment" | CCH Canadian (2004 SCC 13) |
| China | AI-assisted work: yes, where human input is shown | Li v. Liu (Beijing Internet Court, 2023) |
| Australia | No | Telstra v. Phone Directories Co. (2010 FCAFC 149) |
The practical takeaway is consistent even across the split: the more genuine human creative choice goes into a work, the more likely it is protectable, and the portions a machine produced on its own are the portions most at risk of being unprotectable.
Can AI be trained on copyrighted works?
This is the most contested axis, and the spread is enormous, from Japan's broad statutory permission to Australia's lack of any exception at all.
| Country or region | Statutory training / text-and-data-mining exception? | Notes |
|---|---|---|
| Japan | Yes, broad (Copyright Act, Art. 30-4) | Most permissive; limited by an "unreasonable prejudice" proviso |
| European Union | Yes, commercial TDM with a rightsholder opt-out (DSM Art. 4) | AI Act adds training-transparency duties and opt-out compliance |
| United Kingdom | Non-commercial research only (CDPA s.29A) | No commercial exception; reform debated through 2026 |
| United States | None; fair use decided case by case | Courts split (Thomson Reuters v. Ross vs. Bartz v. Anthropic) |
| Canada | None; consultations ongoing | Fair dealing may or may not cover training; untested |
| China | None; Interim Measures require lawfully sourced data | No training safe harbor |
| Australia | None; narrow fair dealing, no general fair use | Most restrictive on training |
Europe sets the strictest rules
The European Union has the most developed framework, layering three instruments. The Copyright in the Digital Single Market Directive (2019/790) allows text-and-data-mining for any purpose, including commercial AI training, but only where rightsholders have not reserved their rights in a machine-readable way (Article 4). The AI Act (Regulation (EU) 2024/1689) then makes honoring that opt-out a binding duty for general-purpose AI providers and requires them to publish a sufficiently detailed summary of training content, backed by fines of up to 3% of worldwide annual turnover or 15 million euros, whichever is higher. On top of both sits the uniquely European "sui generis" database right (Directive 96/9/EC), which can protect a heavily invested dataset even when the individual items are not copyrightable. The United Kingdom, now outside the EU, runs a separate course covered in its own guide, including the unusual computer-generated-works provision.
The permissive end: Japan and the United States
Japan and the United States sit at the permissive end, but for different reasons and on different axes. Japan is permissive on training: Article 30-4 of its Copyright Act, in force since 2019, lets anyone use works for "information analysis," which the government has confirmed includes building AI training datasets, subject to a proviso protecting rightsholders from unreasonable harm. The United States has no such statute and instead leaves training to fair use, which courts are applying inconsistently. On the ownership axis, both deny copyright to purely AI-generated output. So a model trained freely in Japan may still produce output that neither Japan nor the United States will let anyone own.
China's different path on AI output
China stands apart on the ownership question. In Li v. Liu (Beijing Internet Court, 2023), a court held that an image generated with Stable Diffusion was protectable because the user's selection and arrangement of prompts, parameter choices, and iterative refinement reflected enough personal intellectual investment to count as an original work, with authorship vested in the human user rather than the model. An earlier case involving Tencent's "Dreamwriter" reached a similar result for an AI-assisted article, while a separate ruling denied protection to purely software-generated output. The throughline is human contribution, but Chinese courts have been more willing than United States authorities to find that prompting and curation supply it.
Why this matters if you build or ship across borders
Because copyright is territorial, a product sold in several countries is exposed to each country's law at once. Code that is unprotectable in the United States might be protected in the United Kingdom; training that is lawful in Japan might breach an opt-out in the European Union. Companies that operate internationally often default to complying with the strictest market they touch, since one jurisdiction's infringement or training violation can disrupt a global launch. The country guides below set out each regime in detail.
This is general legal information, not legal advice. It compares the law of several jurisdictions as of 2026-06-25 and does not address your specific facts. These regimes are evolving, and several are under active legislative reform. Consult a lawyer qualified in the relevant country before acting.
Related articles
- AI copyright law in the United States
- Is AI-generated code copyright infringement?
- European Union AI copyright laws
- United Kingdom AI copyright laws
- Japan AI copyright laws
- Canada AI copyright laws
- China AI copyright laws
- Australia AI copyright laws
Last updated: 2026-06-25. Statutes, directives, and cases cited reflect their status as of 2026-06-25; several regimes are under active reform.
Frequently Asked Questions
Is there an international AI copyright law?
No. There is no treaty written for artificial intelligence. The Berne Convention and the WTO TRIPS Agreement set shared minimum standards across more than 180 countries, but AI-specific rules on ownership and training are decided nationally and vary widely.
Which country is most permissive about training AI on copyrighted works?
Japan. Article 30-4 of its Copyright Act broadly permits using copyrighted works for information analysis, including AI training, subject to a proviso protecting rightsholders from unreasonable harm. It is widely regarded as the most permissive regime.
Can you copyright AI-generated art anywhere?
In some places. The United Kingdom protects computer-generated works by statute (CDPA s.9(3)), and Chinese courts have protected AI-assisted images where the user showed meaningful creative input (Li v. Liu, 2023). Most other countries require human authorship and deny protection to purely AI-generated output.
Does the EU AI Act apply to companies based outside the EU?
Yes. The AI Act applies to providers that place general-purpose AI models on the EU market regardless of where the provider is established, so non-EU developers serving EU users can fall within its training-transparency and copyright-compliance duties.
Is software copyrighted internationally?
Yes. Under the Berne Convention and TRIPS, computer programs are protected as literary works in member countries. Protection covers the expressive code, not the underlying ideas, algorithms, or programming languages.
Which country's law applies to my cross-border AI product?
Copyright is territorial, so each country's law applies to uses that occur there. A product sold in multiple countries is subject to each one's rules at once, which is why many builders comply with the strictest market they operate in.
Sources and References
- Berne Convention for the Protection of Literary and Artistic Works (WIPO)(wipo.int)
- WTO Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPS), Art. 10(wto.org)
- Thaler v. Perlmutter, No. 23-5233 (D.C. Cir. Mar. 18, 2025)(media.cadc.uscourts.gov).gov
- Regulation (EU) 2024/1689 (EU AI Act), Art. 53(artificialintelligenceact.eu)
- Directive (EU) 2019/790 (DSM Copyright Directive), Art. 4(eur-lex.europa.eu)
- Copyright, Designs and Patents Act 1988 (UK), s.9(3)(legislation.gov.uk).gov
- Copyright Act of Japan, Art. 30-4 (Agency for Cultural Affairs)(bunka.go.jp)
- CCH Canadian Ltd v Law Society of Upper Canada, 2004 SCC 13(canlii.org)
- Beijing Internet Court, Li v. Liu (2023)(chinaiplawupdate.com)
- Telstra Corporation Ltd v Phone Directories Company Pty Ltd [2010] FCAFC 149(austlii.edu.au)