Japan was the first major jurisdiction to write a statutory copyright exception broad enough to cover AI training into its law, and it did so years before generative AI became a boardroom issue. Article 30-4 of the Copyright Act (著作権法第30条の4), in force since 1 January 2019, is the legal foundation that lets AI developers use copyrighted works, including works scraped from the open internet, to train machine learning models without asking the rightsholder first. Understanding what the provision actually permits, and where its limits sit, matters more than ever now that Japanese publishers and news organizations have begun taking AI companies to court and the government's Copyright Subcommittee is actively revisiting the issue.

What Article 30-4 permits

Article 30-4 allows a work to be exploited, in any way and to the extent considered necessary, whenever the purpose is not to enable a person to personally enjoy, or cause another person to enjoy, the thoughts or feelings expressed in that work. This is often called the "non-enjoyment" test. The provision lists three illustrative categories: use in testing to develop or put into practical use technology connected with recording or transmitting works, use for data analysis (statistical extraction, comparison, or classification of language, sounds, images, or other elements from a large volume of works), and other incidental exploitation that a computer processes without a human perceiving the work's expressed content. Machine learning falls squarely within the data analysis category. Because a model is trained to detect statistical patterns rather than to let a human being read, watch, or listen to a given work for its own sake, courts and the Agency for Cultural Affairs (文化庁) have treated the underlying purpose as non-enjoyment, and therefore lawful without a license, subject to the proviso discussed below.

The provision is deliberately technology-neutral. It was drafted as one of several "flexible" limitations added in the 2018 reform to cover future uses that lawmakers could not fully anticipate, generative AI among them. That flexibility is also why its exact boundaries continue to be debated rather than settled by a fixed list of permitted activities.

The "unreasonable prejudice" proviso

Article 30-4 carries a proviso: it does not apply where the exploitation would unreasonably prejudice the interests of the copyright owner, judged in light of the nature and purpose of the work and the circumstances of its use. The Agency for Cultural Affairs' 2024 policy paper, "Thoughts on AI and Copyright" (AIと著作権に関する考え方について), issued by the Copyright Subcommittee of the Copyright Division of the Council for Cultural Affairs, and the accompanying "Checklist and Guidance on AI and Copyright" published on 31 July 2024, set out how this proviso is meant to be applied. The consistent theme is market substitution: does the AI training use conflict with an existing or reasonably foreseeable market for the work, or does it displace a market the rightsholder could realistically enter.

The clearest example the Agency gives is a database of works organized and sold specifically so it can be fed into information-analysis or AI training pipelines. Reproducing that database for training purposes when a license to use it for that exact purpose is commercially available is treated as unreasonably prejudicial, because it directly substitutes for the rightsholder's own licensing market. Works protected by technological measures that block or restrict AI-oriented scraping raise a similar concern. By contrast, copying a limited number of works, or short fragments, for development testing is generally not treated as unreasonably prejudicial, because it does not meaningfully displace the market for the work itself.

A second, more contested application concerns training that is aimed at generating outputs similar in style or expression to particular existing works, for example fine-tuning a model on one artist's portfolio so it can reproduce that artist's distinctive style. Japanese commentary generally treats this as a "mixed purpose" case: if the training is conducted with the intent to enable enjoyment-equivalent outputs (a commonly cited illustration involves training an image generator specifically to output a well-known fictional character), the non-enjoyment premise underlying Article 30-4 itself may fail, independent of the proviso. This is the area legal commentators flag as most likely to produce contested litigation as commercial-scale fine-tuning on identifiable artists' bodies of work becomes more common.

Two stages, two legal regimes

The Agency for Cultural Affairs' framework is built around a distinction that is easy to lose sight of: Article 30-4 addresses the development and training stage. It says nothing about whether a specific output infringes copyright. Generation and use of AI outputs are governed by ordinary copyright rules, meaning the standard tests for reproduction, adaptation, and similarity or dependence on a prior work apply exactly as they would to a human-created work. An AI system trained entirely lawfully under Article 30-4 can still produce an infringing output if that output is substantially similar to, and was generated in reliance on, a specific protected work. This two-stage structure, training under Article 30-4 versus output under ordinary infringement analysis, is the organizing idea behind essentially all subsequent Japanese guidance on the topic.

Guidance is still evolving

The Agency for Cultural Affairs and the Ministry of Economy, Trade and Industry jointly convened a "Network of Stakeholders on AI and Copyright," bringing together creators, performers, AI developers, and service providers. On 30 May 2025 the two ministries published a summary of the network's discussions to date, cataloguing the points of continuing disagreement, including how technological opt-out measures should be recognized, how licensing markets for AI training data should be encouraged to develop, and how the unreasonable prejudice test should apply to industries such as manga, anime, and news publishing where individual creators have limited bargaining power. A new Working Team on Legal Institutions was established under the Copyright Subcommittee's Policy Working Committee, holding its first session on 11 September 2025, where its agenda included AI and copyright alongside other institutional topics such as generative AI's latest situation, designated-organization systems, compensation distribution, and record performance and transmission rights, feeding into the government's Intellectual Property Promotion Plan process for 2026. As of mid-2026, no Japanese court has yet issued a ruling that squarely interprets the unreasonable prejudice proviso as applied to AI training, which means the Agency's guidance, rather than case law, remains the operative reference point for compliance.

Separately, and worth distinguishing from Article 30-4 entirely, a 2023 Copyright Act amendment creates a new arbitration system for works whose rightsholders cannot readily be identified or contacted, taking effect 1 April 2026. It lets a user obtain time-limited authorization from the Commissioner of the Agency for Cultural Affairs, on payment of compensation, for up to three years. This addresses a different problem, unmanaged or effectively orphaned works, but it gives AI developers and other users an additional lawful path where Article 30-4's own conditions are not met.

Enforcement is arriving through the output side, not training

The clearest sign that Japanese rightsholders are prepared to litigate arrived in 2025, but notably at the output stage rather than the training stage. Yomiuri Shimbun sued Perplexity AI in the Tokyo District Court on 7 August 2025, alleging the unauthorized reproduction and public transmission of roughly 119,000 articles through the service's search-and-answer feature, and seeking approximately 2.2 billion yen in damages. Nikkei and Asahi Shimbun followed with a joint suit in the same court weeks later, also seeking 2.2 billion yen each. Both cases center on retrieval-augmented outputs that reproduce or closely track the plaintiffs' article content, rather than a challenge to the lawfulness of any AI training conducted under Article 30-4. That distinction matters for anyone tracking the law: these suits test the output-stage, ordinary-infringement side of Japan's two-stage framework, not the training exception itself. A ruling that reads the unreasonable prejudice proviso broadly, if and when a training-stage case reaches a Japanese court, remains the development most likely to reshape Article 30-4's practical scope.

How Japan compares to the EU and the United States

Japan's approach differs structurally from both major alternatives. The European Union's Digital Single Market Directive establishes a statutory permission for text and data mining in Article 4 that applies unless the rightsholder has validly reserved its rights in an appropriate manner, including through machine-readable means under Article 4(3), giving rightsholders an opt-out from the exception rather than a presumption of consent. The EU AI Act's Article 53(1)(c) ties general-purpose AI model providers' copyright compliance obligations directly to this opt-out mechanism. The relationship between a model's training-stage use of copyrighted works and its later capacity to reproduce protected content was central to the Munich Regional Court's ruling in GEMA v. OpenAI on 11 November 2025, which addressed OpenAI's liability for protected song lyrics appearing in its models' outputs, a case worth watching for its relevance to Japan's training-versus-output distinction even though it turns on different statutory language.

The United States relies on the open-ended, four-factor fair use doctrine in section 107 of the Copyright Act rather than a dedicated statutory exception, and the case law has already split. In Bartz v. Anthropic, a California federal court held in June 2025 that training on lawfully acquired copyrighted books was "exceedingly transformative" and fair use, while separately finding that Anthropic's retention of pirated book copies was not protected. In Thomson Reuters v. Ross Intelligence, a Delaware federal court held in February 2025 that training a competing legal research tool on copyrighted headnotes was not fair use, emphasizing that the output competed directly in the same market as the plaintiff's product. Japan's Article 30-4 sits between these two models: it is a clearer, more predictable statutory rule than the case-by-case US fair use inquiry, but unlike the EU's opt-out regime it does not give rightsholders an affirmative mechanism to withdraw specific works from the exception, relying instead on the market-substitution logic built into the unreasonable prejudice proviso.

What rightsholders need to know

Article 30-4 does not require a rightsholder's permission for training-stage use as a starting point, so the practical leverage points are narrower than a licensing negotiation. Structuring a dataset or database specifically for information-analysis licensing, and pricing it accordingly, is the clearest way to bring a use within the unreasonable prejudice proviso if an AI developer bypasses that market. Applying and documenting technological protection measures, including machine-readable signals restricting AI-oriented access, strengthens a market-substitution argument. Rightsholders should also track outputs, not just training, since Japan's ordinary infringement rules, not Article 30-4, govern whether a generated work substantially reproduces a specific protected work, and this is where the Yomiuri, Nikkei, and Asahi suits are being fought. Industry-level coordination, including publisher and creator associations engaging AI developers directly, is emerging as the practical alternative to individual enforcement given the proviso's high bar.

What AI developers need to know

Training on copyrighted material collected for genuine data analysis purposes is generally defensible under Article 30-4, but that defense weakens considerably where training is deliberately aimed at reproducing a specific work's or artist's distinctive expression, where a commercial licensing market for the same data already exists, or where technological access restrictions were circumvented. The safest compliance posture treats the training stage and the output stage as separate risk surfaces requiring separate controls, since a lawful training process under Article 30-4 provides no defense if a deployed model then generates outputs that substantially reproduce a specific copyrighted work. Developers operating in Japan should also expect the Agency for Cultural Affairs' Checklist and Guidance, not the AI Promotion Act or the AI Guidelines for Business, to be the primary reference point regulators and courts look to, since copyright liability sits in binding law while Japan's AI-specific framework remains voluntary. Monitoring the Copyright Subcommittee's Working Team sessions and any ruling arising from the pending Perplexity litigation is the most direct way to track how the unreasonable prejudice standard is likely to be read going forward.

Primary sources

Related articles