Eleven months after the Diet enacted the Act on Promotion of Research, Development and Utilization of AI-Related Technologies — Japan's so-called AI Promotion Act — the implementation phase is reopening a fight the country thought it had settled in 2018. At the center sits Article 30-4 of the Copyright Act, the provision that lets developers ingest copyrighted works for machine learning without rightsholder consent, so long as the use is not aimed at ‘enjoying’ the expression itself. Japanese manga artists, book publishers, and the Content Overseas Distribution Association (CODA) have intensified calls in 2026 to amend the carve-out, while the Agency for Cultural Affairs (Bunka-chō) continues to defend it as a foundation of Japan's competitiveness in AI.
From a pro-innovation standpoint, the answer is not to gut Article 30-4. It is to keep the rule and tighten the edges where genuine harm appears — outputs that substitute for licensed works, and scraping that bypasses paid access. Japan got the architecture right; the implementation can be improved without surrendering its comparative advantage.
What Article 30-4 actually does
Introduced through the 2018 amendment to the Copyright Act and effective from January 2019, Article 30-4 permits the use of copyrighted works ‘in cases where it is not intended to enjoy or cause others to enjoy the thoughts or sentiments expressed in the work.’ In practice, that covers text and data mining (TDM) and, on Bunka-chō's reading, the training of generative models. The provision applies regardless of the developer's commercial purpose and irrespective of how the work was lawfully obtained — features that make it one of the broadest TDM exceptions in the world, broader than the EU's two-tier rule under Articles 3 and 4 of the 2019 DSM Directive and well beyond the UK's stalled commercial TDM proposal.
Crucially, Article 30-4 contains a built-in safety valve: the exception does not apply where use would ‘unreasonably prejudice the interests of the copyright holder.’ In a March 2024 General Understanding on AI and copyright, the Cultural Affairs Council's Copyright Subcommittee clarified that this proviso bars, among other things, training that reproduces a database compiled specifically for AI training and licensed for a fee, and use of works obtained by circumventing access restrictions. Outputs that reproduce or are substantially similar to training works can still infringe under ordinary copyright rules.
Why the carve-out is under pressure
Three pressures have converged in 2026. First, the AI Promotion Act — Japan's first standalone AI statute — passed in 2025 with a soft-law, guidance-first architecture and an explicit aim of making Japan a hub for foundation model development. That framing has, paradoxically, energised creators who fear the state is doubling down on a regime that gives them no consent right and no statutory remuneration. Second, manga and anime piracy losses — long estimated by CODA in the multi-billion-dollar range annually — are now being conflated with AI training, even though the two harms differ in kind. Third, several US lawsuits against AI labs and the EU AI Act's transparency duties for general-purpose models have given Japanese creators a comparative argument: why should they be the only major creative economy without a meaningful opt-out?
The pressure is real, but the diagnosis is wrong. Article 30-4 does not authorise the harms creators are most worried about. Output-level infringement, scraping behind paywalls, and the wholesale reproduction of an artist's distinctive style in a fine-tuned model targeted at that artist are all already actionable — either as standard infringement or under the ‘unreasonable prejudice’ proviso. The work to be done is enforcement and clarification, not repeal.
The case for keeping the rule
Japan is one of the few jurisdictions where domestic AI labs — Sakana AI, Preferred Networks, NTT's tsuzumi, NEC's cotomi — can plausibly train competitive Japanese-language models without negotiating with every publisher in the country. That matters for cultural sovereignty: a Japan that can only license foreign models trained on foreign data will end up with worse Japanese-language AI than a Japan that can train its own. Article 30-4 is the legal foundation that makes domestic training viable for non-incumbents. Replacing it with an opt-out or a compulsory licence would, in practice, transfer that capacity to whichever players can afford to license catalogues at scale — which is to say, the largest US firms.
What a proportionate fix looks like
A measured update could preserve Article 30-4 while addressing the legitimate complaints driving the backlash:
- Statutory clarity on the ‘unreasonable prejudice’ proviso. Bunka-chō should codify the March 2024 General Understanding into binding guidance, including the bar on circumventing technical access controls and on training datasets that substitute for licensed AI-training corpora.
- Output-side liability that bites. Generative outputs that are substantially similar to identifiable works should attract liability on the model deployer, not just the end user, with a safe harbour where reasonable provenance and filtering measures are in place.
- Transparency, not consent. Borrow from the EU AI Act's training-data summary obligation for general-purpose models: developers operating in Japan should publish a sufficiently detailed summary of training data sources, enabling rightsholders to assert claims where the proviso is breached.
- A licensing market the state actually helps build. Public domain corpora, opt-in licensing platforms, and Bunka-chō-brokered collective deals for manga and literary archives would give creators a route to be paid without dismantling the underlying exception.
The choice Tokyo faces is not ‘protect creators’ versus ‘protect AI labs.’ It is whether to use a scalpel — clearer rules on outputs, scraping, and transparency — or a hammer that would knock out one of the few structural advantages Japan has in the global AI race. The scalpel is the right tool. Article 30-4 should stay.