Since 26 October 2026, Japan's Cabinet Office has been accepting notifications from generative AI businesses that sign up to the Principles-Code for Protection of Intellectual Property and Transparency for the Appropriate Use of Generative AI. The Intellectual Property Strategy Headquarters adopted the code on 25 August 2026. The code leaves the Copyright Act untouched and asks firms to disclose more about how they train. That is a better bargain than the statutory rewrites being fought over elsewhere, though its weakest point is enforcement.
The strongest case for more mandatory rules
Rights-holders have a fair complaint. A manga artist or photographer who finds an AI output resembling their work usually cannot learn whether their site was crawled. Without that information, even a legitimate claim is hard to bring. A voluntary code can be ignored by exactly the firms that most need scrutiny. Anyone who wants binding disclosure duties is responding to a real information gap.
What the code does
The Cabinet Office's English provisional translation sets out two principles. The Japanese text is authoritative. Principle 1 asks generative AI service providers to publish an outline of their models and training data on their corporate websites. That covers model name, provenance, architecture, terms of use and training methods. It also covers the types of training and validation data, whether data came from web crawling, third parties or public datasets, whether synthetic data was used, and which crawlers were used.
On intellectual property, the code asks providers to review their protection principles at least annually and to publish a summary of the review. It also asks them to honour machine-readable crawler restrictions such as robots.txt, to avoid crawling pirate sites, to keep training logs, to use provenance tools such as C2PA, and to maintain contact channels for rights-holders (Cabinet Office, Principles-Code, English provisional translation).
Principle 2 is the notable one. A party who is in, or preparing, litigation, mediation or other legal proceedings over a protected work can ask whether training or validation data contains a specific URL. The request must state sufficient reasons for the requester's status, specify the intended purpose with a pledge not to use the answer for anything else, and identify the specific grounds for the request. The code limits answers to information the provider "can easily access and verify". A footnote says the provision does not generally let creators who find similar AI output ask whether their own site was in a crawl beyond that reference information.
Why this design is defensible
Three features make the code more proportionate than a statute would be.
First, it is explicitly not binding. It says it does not mandate disclosure of trade secrets or security-sensitive information. Baker McKenzie notes that the URL-level request was "one of the most controversial elements of the draft" and that industry groups worried about exposing proprietary know-how (Baker McKenzie, April 2026). A comply-or-explain regime lets a firm that cannot disclose say why. The code says a technical impossibility, properly described, counts as an explanation.
Second, it is scoped to real risk. Systems that produce content that cannot be perceived as the original expression of existing works, or that output only statistical results, fall outside it. So do in-house systems built solely on a client's own data. The burden lands on services where infringement is a plausible concern.
Third, it reaches overseas firms. The code applies to businesses without a head office in Japan if they provide services to Japan, which matters because the largest model developers are foreign.
The design leaves Article 30-4 alone
Japan's approach to training data rests on the Copyright Act's Article 30-4, which the Agency for Cultural Affairs' non-binding "General Understanding on AI and Copyright in Japan" (May 2024) interprets for the AI context. The agency itself says that document represents the subcommittee's reading of the current act at the time of publication and is not legally binding (Agency for Cultural Affairs). The new code does not alter that legal baseline. It adds a transparency layer on top.
That contrasts with the fight over US fair use. The Electronic Frontier Foundation argued on 30 October 2026 that if plaintiffs in the training-data suits prevail, courts would have to rewrite copyright law in ways that hurt everyone except powerful media and tech companies (EFF). The Japanese approach avoids a binary court-led outcome. Developers keep legal certainty about training, and rights-holders gain a route to facts.
Where the weak points are
The code has real limits, and a pro-innovation reading should name them.
- No review of submissions. The Cabinet Office will publish a list of signatories with links to their websites. It says it will not review the content of submissions and will not respond to third-party inquiries. Quality control rests on market and reputational pressure.
- Narrow Principle 2 audience. Only parties in or preparing legal proceedings, or their lawyers or agents, qualify. Many creators will not.
- Easy-to-access limit. Because answers cover only what a provider can easily verify, a firm with poor logging can answer with little.
- Fallback to the courts. The code itself points to party interrogatories (Civil Procedure Code, Article 163) and document production motions (Article 221). Principle 2 is a cheaper first step, not a substitute for discovery.
The code also says it will keep pursuing a balance between AI advancement and IP protection by "more significant means" if needed. That is an honest admission that this is a first iteration, and the annual review cycle gives it a built-in checkpoint.
Verdict
The code asks for disclosures that are cheap for well-run developers and valuable to rights-holders. It leaves training law stable, which supports investment in Japan. The measure of success will be uptake. If major global providers sign and give substantive answers, the code will have delivered transparency without legislation. If they sign with thin explanations, Tokyo will face pressure to move from soft law to rules, and the annual review is where that decision should be tested against evidence.