A 2018 revision to the Japanese Copyright Act (effective 2019) added Article 30-4, which explicitly permits the use of copyrighted works for the purposes of 'data analysis,' including the training of machine-learning models, without requiring authorisation from the rights holder. The provision predated the recent wave of generative-image models but has become one of the principal legal references in international debates about AI training data.
The text of Article 30-4
Article 30-4 of the Japanese Copyright Act (added 2018, effective January 2019) permits the use of copyrighted works 'without seeking enjoyment of the thoughts or sentiments expressed' in the work, including for purposes of analysis. The provision was drafted with text-and-data mining and machine-learning training in mind, but its language is broad: the exception applies regardless of the type of work used, as long as the purpose is not the work's enjoyment.
Practical scope
Under Article 30-4 as currently interpreted by the Japan Patent Office and the Agency for Cultural Affairs, training a machine-learning model on copyrighted manga, anime, or illustrations does not require authorisation from the rights holder. The exception applies to both commercial and non-commercial training. The provision is among the most permissive in any major copyright system regarding AI training data and is one reason several international AI companies have prioritised Japan in their operational structure.
Limits and ongoing debate
Article 30-4 has limits. The exception does not cover model outputs that would themselves infringe copyright if produced by a human — a generated image substantially similar to an existing illustration remains infringing under separate provisions. The exception is also subject to a clause permitting authorities to act if a use 'unreasonably prejudices the interests of the copyright owner.' Japanese illustrators and manga artists have argued through 2023 and 2024 that current generative-image use of their work crosses this threshold; the Agency for Cultural Affairs has held consultations and issued non-binding guidance, but the underlying provision remains in place.
International context
Article 30-4's permissive position contrasts with stricter approaches in the European Union (the 2019 EU Copyright Directive's text-and-data mining provisions permit training but allow rights holders to opt out) and with the unresolved US position, where ongoing litigation (the New York Times v. OpenAI case, Andersen v. Stability AI, and others) will eventually produce a clearer rule. The differing positions across jurisdictions create regulatory arbitrage opportunities for AI companies.