Japan Adopts Voluntary AI Training Data Disclosure Code

Businesses that accept all or part of the code will notify the government and disclose information on their websites about model learning processes, the types of training data used, and the methods used to collect that data, according to Jiji Press.

The framework also calls for participating operators, subject to specified conditions, to disclose whether training data include material that could create copyright-infringement concerns when users or rights holders request that information. This is particularly relevant to generative systems trained on large web-scale corpora, where model developers may not publicly enumerate every underlying work.

Earlier coverage of the draft rules by Japan Today and Anadolu Agency described additional operating expectations. Operators were asked to publish an overview of automated online data-collection programs and avoid collecting from pirate sites.

Japan Today reported that the draft contemplated disclosure requests from rights holders in areas such as manga, music, and film, as well as from users who generate AI content. One cited scenario involved a rights holder preparing potential litigation after finding online AI-generated material that resembles its work.

The draft also asked operators to establish mechanisms for informing users whether training data contained copyrighted works similar to generated content, Japan Today reported. At the same time, the guidelines stated that compulsory disclosure would not be sought for trade secrets or security-related information. The draft carried no penalties for noncompliance, but companies would be asked to explain their reasons, according to Japan Today and Anadolu Agency.

The scope covering foreign providers makes the code relevant to model developers, API platforms, and application vendors that serve Japanese users, even if they are headquartered elsewhere. The reported framework is voluntary, but it creates a government-backed transparency baseline around dataset provenance and copyright-related requests.

For ML teams, comparable transparency regimes commonly increase the value of maintaining records on source categories, web-crawling controls, licensing status, dataset versions, and data-removal workflows. Those records can support externally published summaries without requiring disclosure of model weights, proprietary collection systems, or sensitive security information.

The code also illustrates a broader policy tension for generative AI: rights holders seek enough visibility to investigate possible copying, while providers seek to protect trade secrets and preserve practical access to large-scale training data. Japan's adopted principles address that tension through voluntary disclosure and conditional requests rather than direct statutory penalties.