Japan's government adopted guiding principles on intellectual property protection for generative AI operators on August 26, calling on AI businesses to disclose an outline of the data and methods used to train their AI tools. Intellectual Property Strategy Minister Kimi Onoda announced the Principles Code at a news conference in Tokyo, framing the voluntary framework as a tool to boost AI transparency while protecting intellectual property rights and promoting innovation. While not legally binding, the code covers domestic AI businesses and overseas operators that provide AI systems and services in Japan.
What the Code Requires
Businesses that accept all or part of the code will notify the government and disclose, on their websites, the learning processes, types of learning data, and methods of collecting such data for their generative AI models. They will also make public whether the learning data includes information that could lead to copyright infringement if so requested by AI users and rights holders, provided certain conditions are met. The disclosure obligations apply to text, image, and other material types used as training inputs, addressing long-standing concerns that creative works are being ingested into AI models without permission.
Why It Matters
The Japanese code lands amid a wave of comparable moves by other governments. The European Union's AI Act, which entered into force in 2024 and is being phased in through 2027, requires general-purpose AI providers to disclose training data summaries under the Copyright Directive. South Korea on August 24 announced national AI ethics principles with three core values and seven guidelines, although experts noted those principles lack binding force. The United States has no federal AI training-data disclosure mandate, although individual lawsuits and FTC consent decrees have produced partial disclosures from major model providers.
Enforcement and Limits
Because the Japanese code is voluntary, its effectiveness depends on adoption. The government's strategy is reputational — businesses that refuse to participate in disclosure may face public pressure and loss of procurement eligibility — rather than punitive. Onoda said the government would "respond appropriately," including considering further measures if voluntary adoption proves insufficient. The code's coverage of overseas operators that serve the Japanese market is significant: it effectively means that any global AI provider serving Japanese customers must consider whether to comply with Japanese disclosure expectations even if its home jurisdiction does not require it.
What to Watch Through Year-End
Three checkpoints follow. The first wave of Japanese disclosures from major AI providers — particularly OpenAI, Anthropic, Google, and domestic players like SoftBank-affiliated Sakana AI and NTT — will be the first test of whether voluntary compliance produces meaningful transparency. Japan's planned Copyright Act amendments, expected to be discussed in the Diet this fall, will determine whether voluntary disclosure gives way to binding copyright safe-harbor requirements. And the bilateral AI dialogue between Japan and the EU on training-data transparency standards, which the METI and EU Commission's DG CONNECT are expected to schedule before year-end, will set the template for whether Asian and European transparency regimes can interoperate.
Top comments (1)
Interesting development, especially because the voluntary nature of the framework makes adoption almost as important as the disclosure requirements themselves.
I particularly like the distinction between transparency as a principle and transparency that is actually useful for verification. Disclosing the categories of training data and the learning process is a meaningful step, but the interesting question for me is what can ultimately be independently verified from those disclosures.
That connects with a broader problem I've been exploring in AI security: making claims about an AI system is much easier than building a mechanism that lets someone actually verify those claims.
The same tension appears at the system level with RAG, agents and AI infrastructure: knowing what a system should have access to is very different from being able to demonstrate what it actually accessed and enforced at runtime.
Really enjoyed the article. You write about these developments in a way that makes the policy implications quite concrete, rather than treating transparency as just another checkbox.
I write about the technical side of AI security, adversarial testing and verification as well, so if that angle interests you, feel free to have a look at my work here on DEV. 🔐