What METI's checklist says to check in an AI contract

Before you sign up to an AI service or commission an AI feature, read two sets of clauses: how the vendor may use your input, training included, and who holds the rights to the output. METI's Contract Checklist for the Use and Development of AI, published on February 18, 2025, turns both into specific check points you can take into a vendor meeting.

The checklist is free, in Japanese, and written for people who are not AI lawyers. It is still 44 pages. What follows is the part a business owner or project lead needs before legal gets involved.

Who it is written for

METI names two readers. The first is in-house legal or outside counsel, who will review clause by clause. The second is business staff who understand how a contract is laid out but could not draft a clause themselves.

The second group is the person choosing the AI tool in most companies. The checklist is built so that person can spot the issues early and bring them to legal, instead of signing a click-through and finding out later.

It borrows its vocabulary from the AI Guidelines for Business: developer, provider and user. In the checklist, the side providing the service is the "vendor" and the side using it is the "user".

Three kinds of AI deal

The full text sorts AI deals into three types.

1. Using a general AI service as it is. Its example is a retailer that sends customer purchase history to a large language model service as a prompt and asks for an analysis of preferences.
2. A customized service. A vendor builds extra features, such as a chatbot tuned to your products, on top of a foundation model that a third company provides.
3. New development. You and a developer build an AI system together, such as anomaly detection trained on your own machine data.

Type 2 is where most companies end up when they hire an agency to "add AI" to a site or app, and it has a catch the checklist spells out. Your vendor is a vendor to you and a user of somebody else's model at the same time. The promises it makes you about your data can only be as good as the terms it has accepted from the model provider. Ask to see those terms, or at least the settings, for the plan your data will run through.

The input side: what you send

For input, the checklist walks through definition, provision, use, sharing with third parties, and ownership. Three points matter most.

The definition comes first. If something you send falls outside what the contract calls "input", the checklist warns that the vendor may be free to use it, unless a law says otherwise. Data that is called feedback, know-how or technical information in one clause and input in another is worth reading twice.

Use for training is the second. The checklist is blunt about why: once information has gone into a trained model, removing it is generally difficult, and so is fixing the problem afterwards. Its basic advice is to keep anything unnecessary out of the input and to build the internal rules that make that happen.

Personal data is the third. If the vendor may use what you send for its own purposes, the transfer may no longer count as entrustment under the personal information law, and becomes a provision to a third party that needs the person's consent. That is the same line the privacy regulator drew in 2023, covered in our piece on what not to paste into ChatGPT.

The output side: what you get back

On output, the main check point is whether rights in the output move to you at all, and on what conditions. If they do not, the checklist says to go in understanding that your use of the output will be limited.

This matters for anything you plan to publish or sell: copy, images, code. Whether AI output can carry copyright in the first place is a separate question, which we look at in who owns a logo or design made partly with AI.

What the checklist leaves out on purpose

The authors say they focused on data use and the split of risk. Warranties and indemnities, limits of liability, termination, and governing law and dispute resolution are not covered, and they still tell readers to review them.

They also note the checklist assumes Japanese law. Many AI services are run from abroad and choose a foreign governing law, so the same clause can work differently.

Terms that change after you sign

Section 4.6 covers something that catches a lot of teams. Services offered on standard terms of use update those terms regularly. Under the Civil Code rules on standard form contracts, a vendor can change terms without your individual consent if the change is reasonable and announced in advance with the new text and effective date.

In practice that means somebody in your company has to own the vendor's terms page and check it on a schedule. A change in training settings or data retention is easy to miss when it arrives as a notice on a website.

Logs, if something goes wrong

The security section suggests asking the vendor to keep logs beyond your own usage: authentication logs, access logs and operation logs, including those for administrator and developer privileges on the AI service and its infrastructure. If there is ever a leak, those logs decide whether anyone can tell what happened.

What the government asks of its own AI vendors

On June 12, 2026, the Digital Agency finalized DS-920, its guideline for how ministries procure and use generative AI. Ministries write requirements into their specifications from a procurement checksheet. It covers the vendor's AI governance, handling of input, output and data, output quality including false information, labeling output as AI generated, personal information and intellectual property, security, and explainability. Within reason, they are also asked to get information on the data and methods used to train the model.

The guideline binds ministries, not private companies. It is still a good list to borrow, because a vendor that can answer it for a ministry can answer it for you.

Before the first meeting

Take one page into the meeting with five lines on it: which model provider sits underneath, whether your input is used for training on the plan you will be on, where data is stored and for how long, who holds rights to the output, and who on your side watches the terms for changes. If the vendor can fill it in on the spot, the rest of the contract review gets much shorter.