Japan’s AI-training exception: Article 30-4 and the limits of creator-specific generative AI

Yuka Daimon
Atsumi & Sakai, Tokyo
yuka.daimon@aplaw.jp

Naoyuki Inui
Atsumi & Sakai, Tokyo
naoyuki.inui@aplaw.jp

Introduction

Japan occupies a distinctive position in the global debate over copyright and artificial intelligence (AI) training. Where the legality of training in other major jurisdictions turns on uncertain or conditional frameworks, fair use litigation in the US, or text and data mining exceptions elsewhere subject to research purpose limits or opt outs, Japan has a clear statutory starting point: Article 30-4 of the Copyright Act, introduced by a 2018 amendment, which means that Japan is frequently described as one of the most ‘AI-friendly’ jurisdictions for machine learning.

That description is broadly accurate, but it is incomplete and is becoming more so each year. For foreign counsel advising AI developers, content owners, platforms and licensors with Japan exposure, the useful question is no longer ‘does Article 30-4 permit AI training?’ but ‘where does the exception stop?’ Recent Japanese official and working-level materials addressing creator-specific fine tuning and so-called ‘LoRA’ overlearning suggest that the answer is more fact dependent than the ‘Japan allows AI training’ headline implies.

Article 30-4 in brief

Article 30-4 permits the exploitation of a copyrighted work, without the rightsholder’s permission, where the purpose of the use is not to ‘enjoy’, that is, not to have a person experience the expressive value of, the thoughts or sentiments expressed in that work.[1] Data analysis, including the statistical pattern extraction that underlies machine learning, is treated as a paradigm non-enjoyment purpose. Importantly, the relevant enjoyment purpose is assessed in relation to the specific work being exploited. A general intention that end users will eventually enjoy a model’s outputs does not, by itself, establish an enjoyment purpose in relation to the training works themselves.

Commercial and non-commercial AI developers alike may, therefore, in principle, use copyrighted works for machine learning without a licence, provided the use genuinely serves a non-enjoyment purpose and does not fall within the proviso discussed below. This is materially more permissive than the US fair use doctrine (case-by-case balancing, with contested outcomes still litigated), the UK’s text and data mining exception (limited to non-commercial research) or the European Union framework (a research and cultural heritage-focused exception under Article 3 of the Directive on Copyright in the Digital Single Market, otherwise known as the DSM Directive, not subject to a rightsholder opt out and a broader, purpose-neutral exception under Article 4, subject to an express rightsholder reservation). Neither the UK nor the EU regime combines Article 30-4’s two distinguishing features: a general availability to commercial AI training and the absence of a general statutory rightsholder opt-out mechanism.

The provision is not unconditional. Its proviso withdraws the exception where the use would ‘unreasonably prejudice the interests of the copyright owner’.[2] The July 2024 Checklist & Guidance provides a concrete example: unauthorised use of a database work that is offered, or is reasonably expected to be offered, as a paid AI-training dataset may fall within the proviso. Technical restrictions such as ID/password controls or robots.txt are not, in themselves, a statutory opt out, but together with past licensing or sales practices they may support an inference that the dataset is, or is expected to be, commercially offered, so that circumventing them to collect the data may engage the proviso.[3] The Agency for Cultural Affairs’ (ACA) March 2024 ‘General Understanding on AI and Copyright in Japan’, the most detailed official, though non-binding, interpretive document on the matter, was prepared by the Legal System Subcommittee of the Cultural Council’s Copyright Subdivision.[4] Two of its structural distinctions are central to where Article 30-4’s protection ends: the distinction between the development/training stage and the generation/utilisation stage and the concept of coexisting purposes.

Article 30-4 is not the only relevant provision. Where an enjoyment purpose coexists with a non-enjoyment purpose, as discussed below, Article 30-4 is unavailable. Article 30-4 is not, however, the only potentially relevant provision: in some settings, including certain uses involved in retrieval-augmented generation (RAG)-based services, Article 47-5 may require separate consideration. Article 47-5, a narrower exception enacted in the same 2018 amendment, may permit minor uses of copyrighted works that are incidental to the provision of information-search or information-analysis results. The ACA has indicated that this may extend to certain uses involved in RAG-based services, although the exception is narrow and will not apply where the use exceeds the statutory minor-use threshold, such as where a service reproduces or surfaces substantial portions of the underlying work. This article focuses on Article 30-4, but the relationship between these separate limitations is worth keeping in view, particularly for the RAG-based service discussed below.

Where the exception stops: creator-specific fine tuning and overfitting

The General Understanding states that Article 30-4 does not apply where an enjoyment purpose coexists with a data analysis purpose, providing a direct example: additional training deliberately carried out to cause a generative AI to output the whole or part of a work’s creative expression.[5] The Checklist & Guidance applies this rule to a scenario of considerable practical relevance, naming it as an illustrative case in which an enjoyment purpose may be found to coexist: additional training performed on a small quantity of works by a single creator, using techniques such as low-rank adaptation (LoRA), for the purpose of causing a model to output that creator’s creative expression.[6] Where that purpose is established on the facts, Article 30-4 does not apply and the reproductions involved may infringe copyright absent the rightsholder’s consent or another applicable statutory limitation.

ACA working team material of September 2025 records, among consultation matters raised through the ACA’s legal advice helpline for creative sector professionals, a creator’s report that a LoRA model had apparently been produced from their works and released under their name without authorisation. The material accompanies that report with the general explanation that intentionally causing overfitting of this kind may involve a coexisting enjoyment purpose and may, therefore, result in infringement, consistent with the 2024 documents, without determining that an infringement occurred in regard to the reported instance.[7] This signals that the 2024 framework is being used to respond to real-world concerns, although the material is illustrative working-level reporting rather than a new determination by the Subcommittee.

Two nuances matter. First, Japanese copyright law protects concrete creative expression, not abstract style; imitation of a shared technique or sensibility is not, without more, infringement. The concern is narrower: training deliberately targeted at reproducing a particular creator’s protected expression through creator-specific data curation. Second, the enjoyment purpose inquiry is objective and fact specific, turning on the size and specificity of the dataset, the purpose for which the additional training was designed and the capabilities and use of the resulting model, not the developer’s private intent alone.

Training-stage permission does not resolve output-stage risk

Even where training is lawful under Article 30-4, outputs generated with the resulting model are assessed under ordinary infringement principles, which require both similarity and dependence. Notably, where the relevant work is shown to have been included in the training data, the General Understanding indicates that dependence will ordinarily be inferred even if the user was unaware of that work, although the ultimate determination remains fact specific and a matter for the courts.[8]

A company may, therefore, lawfully train an AI model on a broad corpus and yet face exposure where an output is substantially similar to a specific training work. Further, while an infringing act of training does not, without more, entitle a rightsholder to destruction of the trained model, where the model can be shown to generate, with high probability, outputs reproducing protected expression from specific training works, the model itself may potentially be characterised as a reproduction of those works and destruction may be available; Japanese courts have not yet determined the circumstances in which such relief would be granted,[9] a powerful incentive to prevent memorisation, and arguably the framework’s most important practical takeaway.

This is not theoretical. Since August 2025, Japan’s major newspaper publishers have brought copyright infringement actions against Perplexity AI Inc before the Tokyo District Court: the Yomiuri Shimbun Group filed first,[10] seeking approximately JPY 2.17bn, followed by a joint action by the Asahi Shimbun Company and Nikkei Inc,[11] each seeking JPY 2.2bn and additionally invoking the Unfair Competition Prevention Act. No judgment on the merits has yet been issued in either action: the first oral hearings were held in May 2026. Perplexity has sought dismissal. In the Yomiuri action it reportedly raised threshold objections, inadequate specification of the claims and lack of international jurisdiction, while stating publicly that its service operates lawfully in Japan.[12] The proceedings nonetheless illustrate that a RAG service involves several distinct acts, collection, storage, retrieval-database creation and generation and communication of outputs, each assessed independently under Japan’s copyright rules.

The wider 2026 context: managed flexibility, not simple permissiveness

The debate sits within an increasingly layered AI governance landscape. The AI Guidelines for Business, a non-binding instrument first issued jointly by two ministries (the Ministry of Internal Affairs and Communications (MIC) and the Ministry of Economy, Trade and Industry (METI)) in April 2024, were updated to Version 1.2 in March 2026.[13] Separately, the Japan Fair Trade Commission published its Report Regarding Generative AI Version 2.0 in April 2026, examining these markets from a competition law perspective.[14] A more targeted, sector-specific measure is also under discussion: in December 2025, the secretariat of the Intellectual Property (IP) Strategy Headquarters published for public comment a draft ‘Principle Code’ on IP protection and transparency for generative AI, which would ask AI developers and providers to disclose an overview of their IP safeguard measures on a comply-or-explain basis and, on request from a rightsholder pursuing or preparing litigation, to state whether specific works were used during training. The comment period closed on 26 January 2026, and the IP Promotion Plan 2026, adopted in June 2026, identifies the adoption of the Principle Code as a forthcoming measure.[15] None of these developments amends Article 30-4. Their relevance is structural: Japan is managing AI and IP not through a single comprehensive statute in the manner of the EU AI Act, but through an unamended core exception, non-binding, yet increasingly detailed interpretive guidance, sector-specific soft law and adjacent regulatory pressure. Managed flexibility rather than simple permissiveness or an approaching clampdown.

Practical implications for foreign counsel

Advising on Japan-facing AI training or deployment now calls for more than confirming that Article 30-4 applies. In practice, counsel should: separate training-stage and output-stage analysis, mapping each stage of any RAG pipeline; treat creator-specific or narrowly curated fine tuning as materially higher risk; build output-side safeguards against memorisation and near-reproduction; allocate training- and output-stage infringement risk expressly in licences, warranties and indemnities; and monitor guideline updates, competition law developments and the pending Perplexity litigation. In regard to the burden of proof, the position depends on the claim and requirement at issue, and injunctive relief, unlike damages, requires neither fault nor proof of loss. A party relying on Article 30-4 or 47-5 should be ready to substantiate the facts supporting the limitation, including the purpose, technical structure and scope of the use. A rightsholder may, in turn, point to licensing markets, technical restrictions and the manner of exploitation to argue that the Article 30-4 proviso applies. Detailed records of data provenance and system design may, therefore, matter in any dispute.

Conclusion

Article 30-4 remains, on its face, one of the broadest general statutory exceptions for AI training among the major jurisdictions, and it has not been narrowed by amendment. What has changed is the sophistication with which its boundaries are articulated. Regulatory attention to creator-specific fine tuning, litigation over RAG-based services and the enduring gap between training-stage permission and output-stage liability together show that the real work for foreign counsel lies not in confirming that Article 30-4 exists, but in mapping precisely where, on the facts, its protection runs out.