Japan’s Intellectual Property Strategy Headquarters released a draft AI IP Code on April 26, 2026, proposing a training data transparency regime that would represent a significant departure from the country’s previous voluntary approach to AI governance. According to analysis of the draft, the code would require generative AI developers to maintain granular records of training data types and sources, distinguishing between public, private, and synthetic data, and to document crawler activity used to gather that data.
The draft’s most operationally significant mechanism is its “comply or explain” structure. Under the proposed framework, companies would publish annual compliance statements. Firms unable to meet the disclosure requirements wouldn’t face financial penalties under the current draft, they’d instead be required to publicly explain their non-compliance status. That distinction matters: the enforcement model is reputational and transparency-based, not monetary, which analysts describe as consistent with Japan’s governance tradition while also reflecting the limits of the current regulatory toolkit.
This isn’t Japan’s first AI governance move in recent months. The country activated its AI Strategy Council under the Basic AI Plan, signaling the end of its soft-law era, and separately revised its privacy rules for AI training data. The IP Code, if adopted, would be the third distinct instrument in roughly 90 days, each one adding a new layer of statutory obligation where voluntary guidelines previously stood.
The draft also creates a definitional challenge that legal analysts have flagged. Requirements to identify training data “similar” to specific model outputs would be technically difficult, and potentially infeasible, for large frontier models at current capability levels. Developers who’ve trained on billions of documents can’t readily reconstruct which specific inputs correspond to any given output. According to legal analysis from White & Case, this feasibility gap is one of the draft’s central unresolved tensions.
What comes next matters. Japan’s Intellectual Property Strategy Headquarters is expected to convene an Expert Investigation Team in Q3 2026 to define thresholds for “high-impact” AI models, the designation that would trigger the code’s full obligations. Until those thresholds are set, developers face uncertainty about whether their systems fall within scope.
Global AI developers who’ve treated Japan as a light-touch jurisdiction should reassess. The IP Code draft isn’t final, and its requirements may change before adoption. But the direction is clear: Japan is moving from encouraging good behavior to requiring verifiable records of it. For compliance teams managing multi-jurisdictional AI governance programs, the IP Code adds a third major transparency regime, alongside the EU AI Act’s general-purpose AI provisions and emerging US disclosure expectations, that needs to be tracked and planned for.
The draft status is critical context here. Every requirement described in this brief reflects the proposed text as characterized by secondary sources. None of these obligations are currently in force. The primary source, the official draft publication from Japan’s IP Strategy Headquarters, should be reviewed directly when available.