Generative AI Output and Japanese Copyright Law
A growing question is whether text, images, or code produced by generative AI can be used commercially in Japan. The short answer is that use is often possible, but the locus of risk shifts from "training" to "output." Japanese copyright law broadly permits AI training, yet it applies the ordinary infringement rules to the use of the generated output.
Article 30-4 (Information Analysis and Non-Enjoyment Purposes)
Article 30-4 of Japan's Copyright Act provides that a work may, in principle, be used without the copyright holder's authorization where it is used for "information analysis" or other non-enjoyment purposes. AI training is understood to fall within "information analysis," so using copyrighted works as training data is broadly lawful in itself.
- A "non-enjoyment purpose" means a use whose aim is not for a person to perceive and enjoy the thoughts or feelings expressed in the work
- Unlike the EU and the UK, Japan is notable for permitting training even for commercial purposes
Read alone, this may sound like "AI can use anything freely," but as explained below, that applies only to the training stage.
Free Tool Related to This Article
Contract Risk Checker
Try our free simulator related to this topic.
Try for free →The Proviso to Article 30-4 (Where It Unreasonably Harms Interests)
Article 30-4 contains an important proviso: it does not apply where the use "would unreasonably harm the interests of the copyright holder."
- What counts as "unreasonable harm" is judged case by case, in light of the type of work, its intended use, and the manner of use
- So it is not "always fine because it is information analysis"; depending on the manner of use, the exemption can be denied
Training (Input) and Output Are Separate Stages
This is the most commonly misunderstood point. The relaxation under Article 30-4 is limited to the training (input) stage.
- Training stage: taking in works as training data. Broadly permitted under Article 30-4
- Output / use stage: using or publishing the generated result. The user themselves directly bears the ordinary copyright-infringement risk
Even if the training was lawful, if the output relies on an existing work and is similar to it, it can infringe copyright. Here, Article 30-4—which permitted the training—does not act as a shield.
The Test at the Output Stage (Reliance Plus Similarity)
The output stage is governed by the ordinary copyright-infringement framework, which applies regardless of whether a human or an AI created the work. Infringement turns on two elements:
- Reliance (izokusei): whether the work was created by relying on (deriving from) the existing work
- Similarity (ruijisei): whether it is similar enough that the essential expressive features are shared
For generative AI, it is debated that the presence of a particular work in the training data may affect the assessment of reliance. It is not the case that "an AI generated it automatically, so it cannot infringe"; if reliance and similarity are found, the user who employs that output can be held responsible.
Where Enjoyment Purposes Coexist
Article 30-4 concerns uses for a non-enjoyment purpose. Accordingly, where an enjoyment purpose coexists with a non-enjoyment purpose—for example, where the aim includes having the model output the creative expression of an existing work as such—the use is understood not to be justified by Article 30-4 alone.
Even if something takes the outward form of "information analysis," if it is in substance aimed at reproducing expression for enjoyment, it can fall outside the scope of Article 30-4.
The Government's View (Agency for Cultural Affairs)
On these points, the government's position is set out in the "Approach to AI and Copyright" published by the Agency for Cultural Affairs (the Council for Cultural Affairs) in 2024. By discussing the training stage and the generation/use stage separately, and by organizing the scope of Article 30-4, its proviso, and the concepts of reliance and similarity, the document has become an important practical reference point.
Practical Steps
If commercial use is the premise, the basic approach is to reduce risk through management on the output side.
- Internal usage guidelines: put in writing which uses, which tools, and what manner of use are permitted
- Similarity checks: before publishing or commercializing, check that the output is not similar to an existing work
- Allocating responsibility in the vendor contract: use the contract to clarify ownership of the output, indemnification for infringement, and the terms of use
- Records: keep records of which tool and which instructions (prompts) produced the output, to support explanations of reliance and originality
Summary
In Japan, the starting point is that AI training is broadly permitted under Article 30-4, but that relaxation is limited to the training stage and is also constrained by the proviso. The output stage is governed by the ordinary infringement test of reliance plus similarity, so even lawful training does not prevent infringement where the output resembles an existing work. Commercial use is therefore possible, but the risk must be managed on the output side. Because whether a particular output infringes is a fact-sensitive assessment of reliance and similarity, we recommend consulting a professional at an early stage when the answer is unclear.