Japan's October 26 AI List Will Measure Participation, Not Compliance

Japan will begin accepting provider notifications under its new generative AI disclosure code on October 26. The government plans to publish the names of notifying companies and links to their disclosure pages. It will not certify that those disclosures satisfy the code.

That distinction limits what the first public list can prove. It can show which providers volunteered to be visible. It cannot, by itself, separate detailed implementation from a polished explanation of why little was disclosed.

Japan's Intellectual Property Strategy Headquarters published the final Principles Code on August 25, according to the National Diet Library's Current Awareness portal. The code asks generative AI developers and providers to explain their training processes, data categories, collection methods and intellectual-property safeguards. Its scope includes overseas companies whose systems or services are available in Japan.

The Cabinet Office's code page says notifications will open on October 26. The submission address and other procedural details will arrive that day. There is no filing deadline on the page. An absent company therefore cannot immediately be classified as a refusal.

The registry is an index, not a scorecard

The mechanism uses comply or explain. A provider can implement each principle, or explain why it does not. The final code expressly says it is not legally binding and does not force disclosure of trade secrets or security-sensitive information. Providers are expected to publish an acceptance statement, implementation details and reasons for any non-implementation on their own websites, then review the material annually as a general rule.

Japan's AI Act sits behind this policy but does not convert the code into a statutory disclosure duty. Article 7 of the AI Act separately requires businesses using AI-related technology to cooperate with national and local government measures. The code describes what voluntary transparency should look like. Filing under it is not proof of legal compliance, and declining to file is not proof of illegality.

The notification form makes the evidentiary limit unusually clear. The Cabinet Office will check that the filer is operating and has published acceptance material. The public list will contain the business name and its corporate website address. The office says it will not assess the substance of the notification or whether the company's publication aligns with the code. It will not answer third-party questions about the filing either.

Full and partial adherence may be distinguishable on a provider's linked page. They will not be normalized in the government list. Any comparison will require reading disclosures that companies write in different formats, at different levels of detail and with different exceptions.

A list of links becomes a market signal only when someone converts disclosure quality into a purchasing or rights-enforcement decision.

The useful evidence sits below the declaration

Principle 1 asks for more than a generic statement about responsible data use. Its disclosure fields include model identifiers and versions, architecture and design information, training methods, categories of training and validation data, whether web crawling or third-party datasets were used, crawler identifiers and collection periods. It also covers intellectual-property controls such as respecting paywalls and machine-readable restrictions, retaining training logs, avoiding pirate sites and maintaining a rightsholder contact channel.

The accompanying worked examples reveal how broad the permitted answers can be. A provider might name data types as text, images, audio and video, say that it uses licensed or public datasets, and identify a crawler. Those fields improve discoverability. They still do not amount to a dataset inventory or a license ledger.

Principles 2 and 3 are narrower response mechanisms. A rightsholder preparing legal action may ask whether specified URLs or other reference information appear in accessible training records, subject to stated conditions. A user who generated potentially similar material may make a different request by supplying the output, prompt, intended purpose and reference URL. Trade-secret, security, technical and cost constraints remain available. The code sets no fixed response deadline, although it calls for a prompt response within a reasonable period.

Those details give enterprise buyers a better test than the government list. Procurement teams can ask whether a vendor identifies each accepted principle, names the models and crawlers in scope, explains every omission, publishes a request channel and records an update date. Contract terms can then require notice when those answers change. A bare notification offers none of that assurance.

Rightsholders face a similar choice. The registry can locate a provider's policy and contact route. The value will depend on response specificity, turnaround time and whether the provider can search records at the URL or domain level. None of those outcomes exists before requests are made and measured.

October 26 establishes a baseline

The first list should be read as an adoption baseline. The relevant questions are which major domestic and foreign providers notify, how quickly they do so, and whether their linked pages expose principle-by-principle choices. Later evidence can test whether detailed filers win procurement approvals, answer rightsholder requests faster or prompt rivals to improve their disclosures.

Silence will remain ambiguous at launch. A provider may be preparing a filing, deciding whether it falls within scope, or choosing not to participate. The absence of a deadline makes early league tables especially weak.

Japan has built a low-cost discovery layer for AI transparency. It has not built an audit regime. If buyers and rightsholders compare the linked evidence and attach consequences to weak answers, the voluntary code may acquire market force. If they count names and stop there, October 26 will produce a directory.