Part of the AI Regulation News hub.
Japan opens notifications under its generative AI Principle Code on IP protection and transparency, a voluntary comply-or-explain code that also reaches foreign operators
The code itself is not a legal norm. What starts on 26 October 2026 is a filing step: an operator that accepts the code publishes what it does, or why it does not, and tells the Secretariat.
Bottom line: This is a voluntary soft-law code, not legislation. The code says in terms that it is not a norm with legal binding force, and works by comply-or-explain. What is new in the window is that the Secretariat page says it will accept notifications from Monday 26 October 2026. The code's own publication date is not in the text we read, and we do not state one.
Who this affects: Developers and providers of generative AI offered to the public in Japan, including operators headquartered elsewhere, together with the IP counsel who advise them and the rights holders and litigators who may use the disclosure channels.
Issue date: Notification acceptance opened on 26 October 2026 (Cabinet Office Secretariat page). The code PDF carries no date in its text. The disclosures an accepting operator publishes are to be reviewed and updated at least annually, as a rule.
What changed: The final code is posted with a notification form and an email address. An accepting operator publishes its acceptance, what it implements and any reasons for not implementing on its corporate site, and notifies the Secretariat in the prescribed form.
Analysis: On our reading, the sharpest feature is Principle 3. A user of a generative AI service who finds a URL matching their output can ask whether that URL was in training data, but the code asks for a pledge not to use the answer in litigation or ADR filings. Principle 2 is the mirror image for people who are already in, or preparing, a dispute. The split is the code's own design; what it means in practice is our inference.
Primary sources: Secretariat page: code, English provisional translation, forms (Japanese) · Principle Code (Japanese version, PDF, 14 pages)
- Instrument (EN)
- Principle Code on IP Protection and Transparency for Appropriate Use of Generative AI (our rendering of the title)
- Instrument (native)
- 生成AIの適切な利活用等に向けた知的財産の保護及び透明性に関するプリンシプル・コード
- Authority
- Cabinet Office, Intellectual Property Strategy Headquarters Secretariat
- Jurisdiction
- Japan; also applies by its terms to operators without a head office in Japan whose generative AI is offered toward Japan
- Status
- Final code; notification acceptance opened on 26 October 2026
- Bindingness
- Not legally binding on its own wording; comply-or-explain
- Issue date / next deadline
- No publication date in the code text. Notifications from 26 October 2026. Disclosures reviewed and updated in principle every year
- Primary source
- https://www.cas.go.jp/jp/seisakukaigi/titeki2/ai_principle_code/index.html
What kind of document is this?
It describes itself as modelled on corporate-governance codes such as stewardship codes, using comply-or-explain, and as consistent with the spirit of the 2025 AI Promotion Act. Its section on method says it does not demand mandatory disclosure of an operator's trade secrets or safety and security information, and it says in Japanese that it is 法的拘束力を有する規範ではない (in our rendering, not a norm with legal binding force).
An operator may decline a principle if it explains why. The code adds that saying it has not yet built the structure for Principle 2 or 3 is not enough of an explanation: the operator should say when the structure will be finished.
Who does it reach?
It applies to generative AI developers and generative AI providers, together the operators. A developer is whoever builds a generative AI system and in substance offers it to the public; a provider is whoever offers a service built on one. Research-only actors, and those who only supply parts of a build under contract, are generally outside it.
Several carve-outs turn on who uses the system. A system specialised on one party's own data and given only to that party or its designated users is excluded, as is a system with a markedly low risk of producing infringing output. An operator that licenses a general model and adapts it for an industry can still be an operator, but the code expects it to answer for what it added, with the general model's supplier answering for the rest.
Foreign operators are in scope on the code's terms where the system or service is provided toward Japan, including, but not limited to, cases where Japanese nationals can use it.
What does Principle 1 ask an operator to publish?
Principle 1 has two halves. On transparency, an operator discloses summary information on the model (name and version, history, architecture and licensing, terms of use, training process), the training data (types of data, web-crawled and third-party datasets, public datasets, synthetic data and its purpose, and crawlers by purpose, collection period and name) and accountability (traceability of decisions, responsible persons, documentation).
On IP protection, the operator discloses its position on a list of measures. They include annual review of its IP principles, not infringing others' rights in data use, using crawlers that respect paywalls and machine-readable instructions such as robots.txt, publishing those measures per user agent and giving notice of changes, keeping training logs for a period, avoiding crawling of pirate sites, technical measures against infringing output, provenance technology such as digital watermarks or C2PA, telling users not to use output they think infringes, and a contact point for rights holders with records of how requests were handled.
Where an item cannot be disclosed, the operator gives the reason. The Secretariat is to publish further examples of how detailed each disclosure should be; we did not open that companion document.
How do Principles 2 and 3 work?
Principle 2 is for someone who is bringing, or preparing, litigation, mediation or ADR over their own rights in a work, or their lawyer or authorised representative. They may ask whether a URL they supply appears in the data used for training and validation. The request is to give reasons showing they fall in that category, state the purpose of the answer with a pledge not to use it for another purpose, and identify the URL and why it is relevant. If a provider cannot answer, it is to name the developer of the underlying model.
Principle 3 is for someone who generated content with the operator's service. They supply the output and the prompt, state the purpose and pledge not to use the answer for any other purpose or for litigation or ADR filings, and identify a URL with matching or similar content. The operator then answers whether that URL is in the training data, limited to what it can readily access and check.
In both, the code expects the operator to make a reasoned effort at a detailed and clear answer, to consider seriously any trade-secret objection, and to say what its policy is. It allows anti-abuse measures such as fees or limits on request numbers, provided they do not discourage genuine requests. It sets no response deadline and asks for a reasonably prompt answer. It also notes that civil procedure tools such as party inquiries and document production orders remain available.
How does notification work?
The Secretariat page says notification under paragraph 1(4) of the code is made by filling in the form and sending it as a PDF to the stated email address, and that acceptance started on Monday 26 October 2026 (令和8年10月26日（月）より受付を開始します). An operator that cannot disclose on its own site can send material for the Cabinet Office site. The page says a list of notifying operators will be published, with some delay possible.
The code adds that the Secretariat does not review the content of notifications and does not answer third-party inquiries about them. In plain terms, the filing is a public statement by the operator, not a certificate.
What we did not verify
What we opened: the 14-page Japanese-language code PDF, read as extracted Japanese text, and the Japanese Secretariat notification page. Our English renderings are ours and are labelled where used. Japanese quotations are reproduced in the original characters from those documents.
What we did not open: the English provisional translation, the companion document of examples of summary disclosures, the notification form itself, the list of notifying operators, and any operator's published disclosure. We did not check whether any operator has yet notified.
What we refuse to claim: a publication date for the code. The PDF text does not state one, and secondary sources mention 25 August 2026, which we could not confirm from the primary. The PDF file metadata shows a creation timestamp of 24 August 2026 US Pacific time, which is a file property and not the code's date. We also make no claim about how courts, rights holders or the Secretariat will treat an explanation or a non-notifying operator, or that any operator is covered.
Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.
For an operator offering generative AI toward Japan, the live decision from 26 October is whether to accept the code and publish. The practical groundwork is the same either way: a map of what training data and crawlers sit behind each model, a robots.txt and paywall position, a rights-holder contact point, and a response policy for Principle 2 and 3 requests that does not rely on saying the process is not ready.
Source File
https://www.cas.go.jp/jp/seisakukaigi/titeki2/ai_principle_code/index.html
Open the Secretariat page and read the notification section for the 26 October 2026 start. Then open the Japanese code PDF and read section 1(3) on method and section 1(4) on notification, then Principles 1 to 3. Compare each paragraph against our description.
法的拘束力を有する規範ではない (our rendering: not a norm with legal binding force) · Cabinet Office IP Strategy Headquarters Secretariat, Principle Code, section 1(3)
FAQ
Is Japan's generative AI Principle Code legally binding?
On its own wording, no. It says it is not a norm with legal binding force and uses comply-or-explain, where an operator either implements a principle or publishes its reasons.
When can operators notify the Secretariat?
The Secretariat page says acceptance started on Monday 26 October 2026, by emailing the prescribed form as a PDF.
Does the code apply to foreign AI companies?
By its terms, yes where the generative AI system or service is provided toward Japan, including, but not limited to, cases where Japanese nationals can use it. We did not verify any enforcement consequence.
Who can ask an operator whether a URL was in its training data?
Under Principle 2, a person bringing or preparing litigation, mediation or ADR, or their lawyer. Under Principle 3, a person who generated content with the service, who is to supply the output and prompt and pledge not to use the answer for litigation or ADR filings.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.