Japan Proposes IP Safeguards for AI Training

Japan’s Cabinet Office has proposed a revised Principle-Code for generative AI businesses that outlines new IP protection principles. The draft, discussed on August 18, introduces a ‘comply or explain’ model requiring companies to either follow the principles or publicly explain why they do not.

Scope and Approach

The code, titled Principle-Code for Protection of Intellectual Property and Transparency for the Appropriate Use of Generative AI, applies to developers and providers of generative AI, including foreign firms whose systems are offered in Japan or to Japanese nationals. It aims to balance AI development with the safeguarding of intellectual-property rights and to increase transparency for rights holders and users. The Cabinet Office’s meeting materials note that the draft incorporates changes made after a public consultation that ran from December 26 2025 to January 26 2026.

The framework is non-binding and follows a ‘comply or explain’ approach. Companies must establish IP-protection principles, clarify responsibility for implementation, review them annually, and publish their substance. They are also required to disclose their crawler measures for each user agent and to notify when those measures change.

Data Collection and Transparency

The draft addresses how businesses acquire training material, not just what their models produce. It calls for respect of access restrictions, including paywalls, and for the use of crawlers that obey machine-readable instructions such as robots.txt. Businesses should avoid crawling so-called pirate sites and disclose crawler measures. The code also proposes that training-related logs be retained for a certain period.

Rights holders can seek information about the use of their works in AI development. While the draft does not mandate the public release of every training item, it does allow for the disclosure of specified information about models, training data, and collection methods. A rights holder pursuing a legal remedy may request confirmation of whether a particular URL or identifier was used in training or validation, limited to what the business can readily access and confirm. AI users have an equivalent mechanism for their outputs. The draft recognises limits where information is proprietary, including trade secrets.

Training Safeguards and Rights-Holder Interaction

The proposal also includes measures to prevent infringing outputs. Where possible, businesses should employ technical safeguards such as digital watermarking and the Common-Crawl-Project-Agency (C2PA) standard to verify content origin and provenance. They must establish contact points for rights holders, clarify the requirements for an approach, and keep records of responses. Companies are also expected to advise users not to employ outputs that appear to infringe.

Japan’s broader IP policy highlights transparency around training data and the relationship between AI development and copyrighted works as areas requiring further attention. The draft code does not amend existing copyright rules; instead, it adds governance and disclosure practices to the current framework. This distinction is reflected in the ‘comply or explain’ model, which makes the code non-binding rather than a statutory obligation backed by penalties.

The code sits within Japan’s wider AI governance effort under the Act on Promotion of Research and Development, and Utilization of Artificial Intelligence-related Technology, promulgated on June 4 2025 and fully in force on September 1 2025. The AI Basic Plan was adopted by Cabinet decision on December 23 2025. The draft code, however, remains a proposal and is not yet law or regulation.

The next step will involve assessing how these principles will function in practice, balancing rights-holder requests with trade secrets, security concerns, and the technical difficulty of tracing individual works through large training datasets.

---