What Changed
On August 18, 2026, an expert panel convened by Japan's Cabinet Office broadly approved a draft "principle code" that asks generative AI developers and service providers — foreign operators serving the Japanese market included — to publicly disclose which models they use, what data trained them, and how that data was collected. Rights holders in manga, music, and film would gain a formal channel to ask, under specified conditions, whether a particular webpage was scraped into a training set (Nippon.com/Jiji Press). The code sits on top of the "Act on the Promotion of Research and Development and Utilization of AI-Related Technologies," enacted in May 2025 — a framework statute the Cabinet Office itself describes as designed to advance AI innovation while managing risk, not to punish it (Cabinet Office, AI Act overview).
Crucially, the code carries no penalty provisions. Japan is using a "comply or explain" model: operators can either follow the disclosure norms or publicly state why they won't. The Cabinet Office plans to finalize the code in the near term.
The Case For Disclosure
The strongest argument for the code isn't sentimental. Japan's creative industries — manga, anime, music, film — are a genuine economic asset, and their publishers have spent years watching generative models trained on unlicensed scrapes compete with the works those scrapes were built from. Japan's existing copyright framework, anchored in the Agency for Cultural Affairs' 2024 "Approach to AI and Copyright," already permits AI training under the flexible Article 30-4 exception for non-expressive use (Bunka-chō, AI and Copyright) — but permission to train isn't the same as knowing what was trained on. A rights holder who cannot find out whether their work was scraped cannot even evaluate whether a dispute exists, let alone litigate one. Disclosure obligations are the minimum infrastructure any enforcement regime — voluntary or otherwise — needs to function. That's a defensible floor, not regulatory overreach.
Why the Design Choices Matter More Than the Headline
Where this code earns credit is in what it deliberately avoids. It does not create a new liability standard, a licensing mandate, or a private right of action. It does not touch Article 30-4, the provision that actually determines whether AI training is lawful under Japanese copyright law. It asks for transparency, not permission — which means a firm that discloses accurately faces no new legal exposure simply for having trained on the wider open web. That's the correct sequencing: information first, adjudication (if any) later, through existing copyright remedies rather than a bespoke AI tribunal.
The extraterritorial reach — the code explicitly applies to "overseas operators offering services for Japan" — will draw the most industry pushback, and reasonably so. A mid-sized foreign lab serving Japanese users now faces a disclosure obligation calibrated to Japanese administrative expectations, layered on top of the EU AI Act's own training-data summary requirement and whatever the US eventually settles on. None of these regimes are identical in scope or format. Firms serving multiple jurisdictions will either standardize on the most demanding disclosure format globally — a quiet Brussels-and-Tokyo effect — or maintain jurisdiction-specific compliance packages, a real cost smaller developers will feel more than incumbents.
The Comply-or-Explain Hedge Is the Whole Point
What keeps this from tipping into the more punitive model some rights-holder groups wanted is the absence of penalties. A firm that decides the disclosure regime doesn't fit its architecture — say, a foundation model with training data drawn from thousands of licensed and scraped sources where webpage-level attribution isn't technically tractable — can say so publicly and continue operating. That's not a loophole; it's the feature that keeps a transparency mandate from becoming a de facto licensing requirement enforced through paperwork burden alone. Contrast this with an EU-style regime where non-compliance triggers fines calculated as a percentage of global turnover: Japan's approach lets the market, and reputational pressure from rights holders and users, do the enforcing instead of a regulator.
The guidelines will also apply to overseas operators offering services for Japan — but they carry no monetary penalty for opting out publicly.
The risk worth watching is not the code itself but what replaces "comply or explain" if voluntary compliance is thin. Japan's Cabinet Office has signaled this is meant to balance rights protection with innovation, not to be a first step toward a licensing regime. If large labs treat public explanation as a costless dodge — filing a boilerplate non-compliance notice rather than genuine disclosure — political pressure from the manga and music industries, influential domestic constituencies, could push the next revision toward binding rules. The code's authors have effectively given industry a probationary period to prove that soft law works before Japan reaches for harder tools.
The Regional Signal
Japan's approach is notable mainly for what it isn't: neither the EU's binding training-data-summary requirement under the AI Act nor the more permissive, litigation-driven US posture. It's a third path — disclosure-first, penalty-free, extraterritorial in reach but enforced through reputation rather than fines. For a market this consequential to global AI firms, and for creative industries this economically significant domestically, that's a reasonable bet on transparency over prohibition. Whether it holds depends entirely on whether "explain" ever means something.