Japan’s draft AI IP Code, released April 26, 2026, by the Intellectual Property Strategy Headquarters, would impose training data transparency obligations on generative AI developers operating in Japan. According to analysis from the Center for Data Innovation, the draft requires firms to maintain records identifying whether training data came from public, private, or synthetic sources, and to document crawler activity used in data acquisition.
The mechanism at the center of the proposal is a “comply or explain” structure. Firms would publish annual compliance statements confirming they meet the disclosure requirements. Those that don’t comply must publicly explain why. That framing borrows from securities regulation governance models and places reputational accountability, rather than financial penalties, at the center of enforcement. The Basic AI Plan currently lacks monetary penalty provisions, and the IP Code’s enforcement design appears to follow that same pattern.
Why this matters for compliance teams is not just what the draft requires, but what it assumes is possible. Legal analysts at White & Case and others in the practitioner community have noted that requirements to identify training data “similar” to specific model outputs may be technically infeasible at scale for large frontier models. The draft does not appear to resolve that gap. If those feasibility concerns aren’t addressed before adoption, firms face a standard they cannot demonstrably meet, compliance statements that must either hedge heavily or risk misrepresentation.
The broader context is Japan’s regulatory pivot. This draft follows two other significant moves in roughly 90 days: Japan’s activation of an AI Strategy Council under the Basic AI Plan, ending its informal soft-law approach to AI governance, and a revision of privacy rules to allow AI training data use while adding profit-clawback penalties for intentional violations. Those earlier moves focused on governance structure and data access. The IP Code targets the output side, what’s done with the data and whether it can be traced. Together, the three instruments describe a jurisdiction that had been treated as permissive now constructing a layered statutory framework.
The primary draft documentWhat to watch: Japan’s IP Strategy Headquarters is expected to convene an Expert Investigation Team in Q3 2026 to define thresholds for “high-impact” AI models, which will determine which developers fall within the Code’s scope. That threshold decision is the single most consequential pending element of this draft. Companies below the threshold may face no obligations at all. Companies above it face the full comply-or-explain structure.
The question worth sitting with: if a firm’s compliance statement must cover training data “similar” to outputs, and that similarity can’t be technically demonstrated at scale, does a public compliance statement become a liability rather than a defense? That’s the kind of exposure question compliance counsel should be raising now, before the Expert Investigation Team’s Q3 threshold determination shapes what “high-impact” means for their systems.