Japan's Article 30-4: the Text-and-Data-Mining Copyright Exception

Japan is routinely described as the jurisdiction where training AI on copyrighted works is simply legal. The provision behind that reputation is Article 30-4 of the Copyright Act, introduced by the 2018 amendment and in force since 1 January 2019. It is broader than the European equivalent in one dimension and narrower in another, and it contains a proviso that does most of the work people assume it lacks.

What the provision says

Article 30-4 permits exploitation of a work, by any means and to the extent considered necessary, where it is not for the purpose of enjoying or causing another person to enjoy the thoughts or sentiments expressed in it. It gives examples, of which the relevant one is use for data analysis — extracting and comparing information from a large number of works — and use in the course of computer information processing that does not involve perceiving the expression.

Two features distinguish it from the EU regime described in the Article 4 CDSM opt-out page. There is no rightsholder opt-out: a reservation of rights in terms of use does not disapply Article 30-4, so the machine-readable-reservation debate consuming Europe has no counterpart here. And there is no express lawful-access condition of the European kind, though obtaining copies by circumventing technical measures or in breach of other law raises its own problems. Official English material on the Act and on the Agency for Cultural Affairs’ interpretation is published by the Agency for Cultural Affairs.

The non-enjoyment test

The organising concept is enjoyment, which in this context means appreciating the expression — reading the novel, viewing the picture, listening to the song — as opposed to processing it as data. Training a model to learn statistical structure is not enjoyment; the machine does not appreciate the expression, and the copies exist to extract information rather than to be experienced.

The Agency for Cultural Affairs published a general understanding on AI and copyright in March 2024, approved by the Subdivision on Copyright of the Council for Cultural Affairs, which works through how the test applies to modern systems. Its most consequential point is about mixed purposes: where a use has an enjoyment purpose alongside the data-analysis purpose, Article 30-4 does not apply. It is not a question of which purpose predominates. If enjoyment of the expression is among the purposes, the exception is unavailable.

That has real consequences for common practices. Fine-tuning on a specific creator’s works with the object of generating outputs in that creator’s distinctive expression is the paradigm case the guidance identifies as falling outside the exception, because the point of the exercise is the expression itself. Retrieval systems that hold copies in order to present the expressive content back to a user are not doing non-enjoyment processing either; different provisions, including the minor-use provisions for computerised information services, are the ones in play there.

The proviso that limits it

Article 30-4 ends with a proviso: the exception does not apply where the exploitation would unreasonably prejudice the interests of the copyright owner in light of the nature or purpose of the work or the circumstances of its exploitation. This is the part left out of the summaries, and it is the part that makes the provision a standard rather than a rule.

The clearest agreed example concerns databases sold for the purpose of information analysis. Where a rightsholder markets a dataset specifically for machine analysis, copying it without a licence for exactly that purpose destroys the market the rightsholder has established, and the guidance treats that as unreasonable prejudice. The reasoning generalises to a principle: the exception is not a licence to displace an existing licensing market for the analytical use itself.

Beyond that example, the boundaries are unsettled. Whether large-scale scraping of a sector’s output unreasonably prejudices the interests of the rightsholders in that sector is arguable and has not been resolved by a Japanese court in a way that settles it for AI training. Anyone asserting a confident answer about where the proviso bites is stating a position, not reporting the law.

Outputs are a separate question

Article 30-4 addresses the training stage. It says nothing about generation. When a model produces something similar to an existing work, ordinary infringement analysis applies under Japanese law: whether the output is similar to the protected expression, and whether it relies on the earlier work — the two elements of similarity and reliance.

The March 2024 guidance discusses how reliance is assessed where a generative system is involved, including the position where a user deliberately prompts towards a known work and the position where a work was in the training data. A developer whose training was covered by Article 30-4 therefore has no answer to an infringement claim about a specific output; those are different acts with different analyses, and liability may attach to the user, the service provider, or both depending on the facts.

The common misreading

The claim “training AI is legal in Japan” is a compressed version of something narrower. What Article 30-4 supports is: exploitation for a purpose that does not involve enjoying the expression, to the extent necessary, unless doing so would unreasonably prejudice the copyright owner’s interests. That is a permission with two conditions and a general limit, and each has been the subject of official interpretation precisely because it is not self-applying.

Two further points for anyone relying on the provision commercially. Japanese copyright law governs acts in Japan; training conducted or exploited elsewhere is subject to that jurisdiction’s law, so Article 30-4 does not travel with a model. And Japan’s separate legislation promoting AI research and use, enacted in 2025, is a promotion and coordination statute rather than a copyright reform — it does not widen Article 30-4, and it is covered in the AI Promotion Act page. The policy debate in Japan has if anything moved towards clarifying the limits of Article 30-4 rather than broadening it.