Ten things to check in a generative-AI or API contract
Input data, training use, output rights, model changes, logs and liability caps.
Notes / By theme
Launching a service built on generative AI in Japan is not only a question of AI rules. The contract with the model provider, the data going in, training use, personal data, copyright, hallucination, advertising claims and sector-specific regulation all have to be looked at together.
Put the other way round: without separating those questions, there is no way to tell when the work is finished. These notes organise the subject under three headings — contracts, liability and personal data — with the primary sources in Japanese linked at each point.
What these notes cover
First
Eleven items to settle before launch. The individual questions are taken up in the sections below.
Contract
An AI product is normally built on someone else's model, which means the company in the middle is a provider to its customers and a user to its vendor at the same time. What it can promise downstream is limited by what it has secured upstream, and when something goes wrong the analysis follows information and control rather than job titles.
Input data, training use, output rights, model changes, logs and liability caps.
METI's 2026 handbook, and the difference between AI that assists and AI that replaces.
Privacy
The argument about AI tends to focus on models and outputs, while the problem that arrives first is the data going in. Japan's 2026 amendment opens a route for development that can be characterised as statistical production, and at the same time raises the consequences for high-risk handling and serious breaches.
Statistical production, under-16s, facial feature data, processors, and the surcharge.
Article 28 and cross-border transfer when the tool is hosted outside Japan.
Summary
In order: decide where your company sits in the AI value chain (developing, providing or using); confirm what the upstream contract actually gives you; look at whether the data going in contains personal data or third-party rights; and decide whether a human reviews the output or not.
Taken in that order, "complying with the AI guidelines" stops being an abstract exercise and becomes a list of clauses to change and procedures to write.
Related Japanese-language guideline indexes: AI and data, data protection, intellectual property.
Tell me how the service is put together, which model you are using and where it comes from, what data it handles and who the users are, and I can set out the questions that need answering and how to approach them. Launch work generally is described on the startup legal page.