Japan’s Draft AI IP Code Misses the Mark, Undermining US Alignment

Japan has spent years building a well-earned reputation as one of the world’s most thoughtfully calibrated AI regulatory regimes. Its AI Promotion Act, enacted in 2025, is explicitly pro-innovation, and the Hiroshima AI Process, launched under Japan’s G7 presidency, helped set the tone for voluntary, interoperable governance at the international level. That record has put Japan and the United States on the same side of many core AI policy questions—alignment that matters because both President Trump and Prime Minister Takaichi are seeking to deliver a “new golden age” in U.S.-Japan relations built on deeper economic and technological integration, where AI will be central. Unfortunately, Japan is close to finalizing a draft code on intellectual property (IP) and AI that would impose requirements that are technically impossible to meet and abandon the light-touch, pro-innovation approach to AI governance that Tokyo and Washington have been championing together on the global stage.

The draft code, issued by Japan’s Intellectual Property Strategy Headquarters, seeks to create a system in which rightsholders can trace how AI models use their content in training. It would require AI model developers to disclose an outline of how their systems are built and trained—including model architecture and training methods—as well as the types and sources of training data, whether drawn from public datasets, private datasets obtained by web crawling, third-party sources, or synthetic data, alongside details of their crawler activity. AI firms would also be expected to retain records of the training process, data use, and system activity over time and maintain documentation that allows tracing key development choices, data inputs, and system behavior. The requirements apply to any developer or provider offering generative AI systems or services in Japan, regardless of where they are headquartered, and companies must publish annual compliance statements and submit notifications to the Cabinet Office, which maintains a public registry of firms and their responses. Those who do not comply must publicly explain their non-compliance.

The draft code has several significant technical and design flaws. First, the code’s disclosure requirements rest on technical assumptions that don’t hold for large-scale AI systems. Confirming whether a domain appeared in training data (as per principle 2) is difficult enough, but principle 3’s requirement to determine whether training data contains content “identical or similar” to a specific generated output is not technically feasible for frontier models that don’t retain granular records of individual training inputs. Modern AI models learn statistical relationships across large datasets, not discrete, traceable contributions from individual sources. The draft code itself hedges on this point, qualifying several requirements with language such as “to the extent technically possible and reasonable.” But this hedging means firms will likely invoke technical infeasibility as grounds for non-compliance, producing a public registry populated with explanations rather than disclosures, which would serve neither regulators nor rightsholders.

This requirement, therefore, creates a compliance obligation that cannot be met in practice. A more effective approach already exists within the draft itself: the code’s own Principle 1 includes honoring robots.txt as a machine-readable opt-out and advancing watermarking and provenance tools. Tokyo should build from those workable foundations rather than layering technically infeasible traceability requirements on top of them. It should align disclosure expectations through multilateral efforts such as the Hiroshima AI Process to ensure they are technically grounded and interoperable across jurisdictions.

Second, the code introduces a recordkeeping obligation without defining its limits. It requires firms to retain documentation on system development, data use, and model behavior, but does not specify a clear retention period. Instead, it defers to separate guidance that defines the timeline only as “within a reasonable range.” That ambiguity leaves firms without a clear compliance endpoint. Without knowing how long records must be kept, companies cannot determine when they have satisfied their obligations—or when they are exposed to public non-compliance findings under the code’s own registry mechanism.

Third, the code mandates disclosures at a level of detail that risks exposing sensitive technical and commercial information without a clear accountability benefit. Ironically, a set of rules meant to enhance IP rights disregards the importance of trade secrets. Requiring granular insight into system design and data practices could reveal proprietary methods and make it easier for third parties to identify vulnerabilities, manipulate data inputs, or reverse-engineer system design. During the public comment process, both domestic and international firms warned that these requirements go beyond what is necessary for transparency and could undermine legitimate competitive advantages. As written, the provision risks discouraging lawful and widely accepted data practices in machine learning, while producing disclosures that are less informative than they appear. Knowing that a URL was included in a crawl target says nothing about whether the associated content survived data cleaning. The draft code’s flaws are not just technical; they put Japan on the wrong side of a global debate where it has, until now, reinforced a more balanced approach. The question over how governments govern AI is far from settled, and other countries are watching to see which approaches take hold. A Japan that shifts course weakens the broader case that like-minded democracies can govern AI without stifling it.

The United States is still firmly on the side of light-touch, pro-innovation AI governance. The Trump administration has laid out a clear federal direction on AI and IP through its new national AI policy framework, one that explicitly rejects the kind of heavy disclosure and traceability burdens that Japan’s draft code would impose. Even the U.S. Copyright Office, whose May 2025 report on AI training and copyright seemed to draw the ire of the administration for being too restrictive of AI innovation, stopped well short of calling for anything like Japan’s approach—pointing instead toward voluntary licensing and market-based solutions rather than mandatory disclosure regimes. Japan, which has spent years standing shoulder to shoulder with the United States in pushing back against unnecessarily heavy-handed AI regulation, would, with this draft code, find itself on the other side of that argument.

Japan’s Intellectual Property Strategy Headquarters should revise the code before finalizing it. The underlying concern—how to address IP in the context of generative AI—is legitimate. But requirements that cannot be met, combined with disclosure obligations that go beyond what is necessary, will not solve that problem. They risk narrowing the set of firms willing to operate in Japan without improving protection for rights holders. A more effective approach would focus on enforceable standards, technically feasible transparency, and alignment with international efforts already moving in that direction.

Image credit for social media preview: Bryan Jones/Flickr