Japan's Copyright Act Article 30-4 — Training AI Without Permission

When the question is AI and copyright under Japanese law, the first provision to read is Article 30-4 of the Copyright Act. The short answer: in Japan, collecting and copying works in order to train AI can, as a rule, be done without the copyright holder's permission, because the article permits uses not aimed at "enjoying" the thoughts or sentiments expressed in a work, to the extent considered necessary and by any method. Two things take a use outside it: a purpose of enjoyment that coexists with the analytical one, and a use that would unreasonably prejudice the copyright holder's interests. Publishing or selling images and text made with AI is a separate stage, judged not under this article but by the ordinary infringement test of similarity and reliance.

Article 30-4 was created by the 2018 amendment of the Copyright Act (Act No. 30 of 2018). According to the Agency for Cultural Affairs, the amending act passed the House of Councillors on 18 May 2018, was promulgated on 25 May and took effect on 1 January 2019.

The text

The main clause, as published in Japan's official e-Gov law database, says (our translation): a work may be used, by any method and to the extent considered necessary, in the cases listed and in any other case where the purpose is not to enjoy, or have others enjoy, the thoughts or sentiments expressed in it — except where, in light of the nature and purpose of the work and the manner of use, this would unreasonably prejudice the interests of the copyright holder.

Three examples follow: testing in technology development (item 1), information analysis (item 2) — extracting, comparing and classifying information on the elements of a large volume of works or other information — and computer processing that involves no human perception of the work's expression (item 3).

Where it came from

The Agency's explanatory guide of 24 October 2019 sets out the history.

Two points stand out. Old Article 47-7 limited information analysis to "statistical" analysis; the new provision dropped that word, which the guide says brings in recording works as training data for AI developed by deep learning. And the guide states that an AI reading works to learn is not "enjoyment", a human act, so both collecting training data and supplying it to others for AI development are covered.

"A machine-learning paradise"

In January 2019, just after the article took effect, attorney Taichi Kakinuma of STORIA Law Office showed AI developers a slide titled "Japan = machine-learning paradise." According to @IT, he explained that similar European and UK provisions were limited to non-commercial purposes, whereas in Japan models could be trained for profit too. The copyright scholar Tatsuhiro Ueno used the same label in the title of a 2021 journal article, and it has since recurred in legal commentary. A December 2025 Nishimura & Asahi newsletter notes that Ueno, in a 2024 book, called Japan's provision stronger than other countries' because it places no limit on the works covered or the acts permitted.

That comparison, however, dates from early 2019. In April of the same year the EU adopted the Digital Single Market Directive, whose Article 4 permits text and data mining, commercial uses included, unless rights holders have expressly reserved their rights in an appropriate manner, such as by machine-readable means. The difference shifted from "commercial or not" to whether rights holders have a way to opt out.

The 2024 General Understanding

The spread of generative AI made the reach of Article 30-4 contested. After hearings and a public consultation, the Legal System Subcommittee of the Council for Cultural Affairs' Copyright Subdivision compiled its General Understanding on AI and Copyright on 15 March 2024. The document itself says it has no legal force and makes no definitive legal assessment of any particular generative AI.

On style, it says sharing a style that is merely an idea does not infringe, but additional training aimed at reproducing the shared creative expression of a small body of one creator's works may fall outside Article 30-4. On 31 July 2024 the Agency followed with a Checklist and Guidance on AI and Copyright, with one part for AI developers, providers and users and another for rights holders.

Japan, the EU and the United States

In the United States the dispute is moving through judgments and settlements. The first ruling to address training squarely was Bartz v. Anthropic (2025); other leading cases are The New York Times v. OpenAI and, decided in the UK, Getty Images v. Stability AI. Their stories are told on those pages. On protection for AI output itself, US law also requires a human author (Thaler v. Perlmutter), broadly in line with the creative-contribution approach in Japan's General Understanding.

What it changed

Article 30-4 protects training, not output or distribution. That distinction has shaped Japan's disputes in the years since it took effect.

For developers, the Agency's guide gave a clear expectation that building training datasets in Japan, and supplying them for AI development, needs no individual permission. According to the Nishimura & Asahi newsletter, attorney Matsushita argued in 2025 that the real problem arises only when AI outputs expression similar to existing works, so liability at the use stage suffices; the same newsletter summarises the prevailing view as favouring infringement findings at the training stage, with injunctions, for AI that can reproduce particular works in bulk.

Rights holders' concerns have moved beyond training. On 7 August 2025 the Yomiuri Shimbun sued the AI search service Perplexity in the Tokyo District Court, alleging infringement of its reproduction and public-transmission rights and seeking about 2.168 billion yen. On 26 August the Asahi Shimbun and Nikkei jointly sued the same company for a total of 4.4 billion yen. According to Impress Watch, they say Perplexity ignored their robots.txt refusals and copied and stored articles without permission, repeatedly displaying them in answers. This is a dispute over RAG-style use, not training itself.

On 27 October 2025 the Content Overseas Distribution Association (CODA) asked OpenAI not to train its video model Sora 2 on members' content without permission. OpenAI accepts opt-out requests from rights holders, INTERNET Watch reported, but Japanese law has no statutory opt-out, so the two approaches now meet head-on. In manga, the Japan Cartoonists Association issued a joint statement on 31 October 2025 calling on AI companies to obtain permission, make training data transparent and pay fair compensation to rights holders who consent.

For training, the text of Article 30-4 has not changed since 2019. What has changed is that the dispute has spread from training itself to output, distribution and payment for training data.