Japan's law lets machines analyse works no person will ever enjoy

The process wall

Since 1 January 2019, Article 30-4 of Japan's Copyright Act, amended by Law No. 30 of 2018, has permitted using a work without the copyright owner's permission when the purpose is not for a person to perceive its expression. The Japanese text names three cases, including 情報解析 (joho kaiseki, information analysis), defined in the statute as extracting and comparing language, sound, image or other elements from a large volume of works or other information, a definition written broadly enough to cover training a machine-learning model on scraped images, years before generative image tools existed. The wall this creates for a Japanese artist is the mirror of the US and UK debates: the question is not whether a human read or viewed the work during training, but whether a machine did.

What the documents show

The statute's operative sentence, in the Agency for Cultural Affairs' own drafting, reads: 著作物は、次に掲げる場合その他の当該著作物に表現された思想又は感情を自ら享受し又は他人に享受させることを目的としない場合には、その必要と認められる限度において、いずれの方法によるかを問わず、利用することができる。 In gloss: a work may be exploited, by any method and to the extent necessary, whenever the purpose is not to let a person, oneself or another, enjoy the thoughts or feelings it expresses. Its proviso withdraws that permission where the use would, in the statute's words, 著作権者の利益を不当に害することとなる場合 -- unreasonably prejudice the copyright owner's interests, judged against the work's type, purpose and the manner of use. The Agency's General Understanding on AI and Copyright in Japan sets out how it reads that proviso for AI: a provider that trains a model specifically to make it reproduce one identifiable work's expressive features, rather than to learn general patterns, falls outside Article 30-4 and back under ordinary licensing rules.

Authorship and limitation

The mechanism is perception, not effort: an analytical process a person never experiences as the work, even where it copies the file to run that analysis, is treated differently from a use where a person ultimately reads, views or hears the copyrighted expression. The guidance's own limit follows the same logic in reverse: where training targets one artist's style specifically, or an output is held to substantially resemble and depend on a specific work, the article's protection can fail and ordinary infringement analysis, including a similarity-and-reliance test the guidance describes for outputs, takes over.

What survives

What survives for an artist working under Japanese law is the proviso, a fact-specific safety valve, plus a separate line at the output stage: training on a work is not the same legal question as generating something that resembles it. Editorially, a maker should treat Article 30-4 as a training-stage rule only, and check separately whether any specific output competes with or reproduces their protected expression.

• Read the statute's three listed cases, including the information-analysis definition, against how a specific tool describes its own training process.
• Check the 2024 guidance's own examples of when the proviso removes the training exception.
• Compare the guidance's similarity-and-reliance test for outputs against a specific generated image.

Two questions sit inside this one article: whether training itself needed a licence, and whether an output infringes; keeping them separate is what the guidance itself insists on.