A legal checklist for launching a generative-AI service in Japan
In short
For a generative-AI service in Japan, the terms of service are only one piece. The contract with the upstream AI provider, the flow of personal data, copyright in inputs, RAG corpora and outputs, hallucination, sector-specific regulation, security and AI governance all have to be checked together. What matters most is not applying the law once the service is finished, but reflecting the legal risks in the product specification at the design stage.
Launching a service that embeds generative AI calls for the same work as any web service on terms of service and data protection, plus a set of issues specific to AI: how input data is handled, copyright in AI output, hallucination, the contract with the model provider, and AI governance.
Where you call an external model — OpenAI, Google, Anthropic — through an API, the analysis cannot stop at "our company and the user". The data flow and the allocation of responsibility have to be mapped across the whole service, including "our company and the AI provider".
Japan's Act on Promotion of Research, Development and Utilisation of AI-Related Technologies (the AI Act) came fully into force on 1 September 2025, and the AI Business Operator Guidelines were revised to version 1.2 in March 2026. A service launched now should be designed against those rules.
1 The checklist
At a minimum, work through the following before launch.
| Item | What to check |
|---|---|
| ① Your position in the AI value chain | Whether you are an AI developer, an AI provider or an AI user |
| ② Contract with the AI provider | Use of input data, training, retention, rights in output, limitation of liability |
| ③ Personal data | Whether prompts contain personal data; whether data leaves Japan |
| ④ Copyright and IP | Whether inputs, RAG data and outputs infringe third-party rights |
| ⑤ Terms of service | Whether AI-specific risks and the allocation of responsibility are reflected |
| ⑥ Hallucination | How the risk of loss caused by incorrect output is managed |
| ⑦ Third-party rights | Reputation, privacy, likeness, trade marks |
| ⑧ Sector-specific regulation | Medical, financial, legal, recruitment and other regulated fields |
| ⑨ Consumer and advertising rules | Charging, subscriptions, performance claims |
| ⑩ Security | Prompt injection, leakage of confidential information |
| ⑪ AI governance | Risk assessment, logging, monitoring, incident response |
Each is taken in turn below.
2 Identify whether you are a developer, a provider or a user
The first step is to draw the technical architecture.
A typical configuration
User → your service → external LLM API → your service → user
On that structure, even though you do not develop the foundation model, you are offering a service with generative AI embedded in it, and you may therefore have the characteristics of an AI provider.
Version 1.2 of the AI Business Operator Guidelines divides the actors into AI developers, AI providers and AI users, and sets out what is expected of each. It also states common principles: human-centricity, safety, fairness, privacy protection, security, transparency and accountability.
So begin by settling the following.
- Who supplies the foundation model
- Who writes the prompts
- Who collects data from users
- Whether the AI output is displayed as it is, or processed and checked by you
- Whether you use RAG or fine-tuning
- Whether the AI makes the final decision, or a person reviews it
The AI Act does not establish a general licensing or registration regime for AI services. It does position businesses that develop or provide AI-enabled products and services as "utilising businesses" and sets out responsibilities such as cooperating with national and local government measures.
3 Read the contract with the upstream AI provider
Where you use an external model through an API, the provider's contract has to be read. The points that matter most:
- Whether the provider uses prompts and uploaded data for training
- How long data is retained
- In which country data is stored and processed
- Whether sub-processors are involved
- What rights you have to warrant in respect of input data
- What rights you are granted in the output
- Who bears liability if third-party IP is infringed
- Whether the provider offers an indemnity
- Whether your use case falls within the prohibited uses
- What happens if the API or the model is changed or discontinued
- Whether the terms can be varied unilaterally
METI's "Checklist for contracts on the use and development of AI", published in February 2025, organises the same ground: use of inputs and outputs, provision to third parties, allocation of rights, personal data, security, log retention and amendment of terms.
The practical point
Using generative AI does not mean the only document to get right is your own terms with your users. Check that what you promise your users is consistent with the terms imposed on you by the AI provider upstream.
4 Establish whether personal data reaches the AI provider
Users type names, employers, medical histories and the substance of their problems into free-text fields. So the data flow — where the information goes, who handles it and for what purpose — has to be established.
The Personal Information Protection Commission has stated that a business entering personal data into a generative-AI service must confirm that doing so falls within its stated purpose of use. It has also cautioned that entering personal data without the data subject's consent, where that information is used for a purpose other than generating the response — training the model, for example — may breach the Act on the Protection of Personal Information (APPI).
Where the provider is outside Japan, the following also need to be considered.
- Whether the transfer is a third-party provision
- Whether it can be characterised as entrustment of handling
- Whether Article 28 of the APPI on provision to a third party in a foreign country has to be addressed
- Whether you rely on the recipient's conforming framework
- Whether a data processing addendum secures appropriate handling
Characterising the arrangement as "entrustment" does not automatically take it outside the cross-border rules. The Commission's Q&A indicates that entrustment to a party in a foreign country requires compliance with Article 28(1) other than in defined cases.
Privacy by design
Design choices matter here: not sending unnecessary personal data to the model at all, masking it, or selecting an API configuration in which input is not used for training.
5 Sort out copyright in inputs, RAG data and outputs
For generative AI, copyright divides into two stages: ① input and training, and ② use of the generated content.
Input and training
Where third-party text, images or databases are entered into a model or used as RAG data, whether that use is permitted under the Copyright Act has to be checked.
Article 30-4 provides a limitation on rights for information analysis and similar uses, but it does not make every use of a work in connection with generative AI free. The purpose and the manner of use are assessed individually.
The Agency for Cultural Affairs published its "Views on AI and copyright" in March 2024 and a "Checklist and guidance on AI and copyright" in July of the same year. In particular, when you build your own RAG corpus, the source of the data and the purpose of its use both have to be checked; "it is generative AI, so it is allowed" is not an analysis.
Use of the output
Where AI output resembles an existing work, infringement can arise. Nor does copyright necessarily subsist in AI-generated content: that depends on the extent of creative human involvement.
The terms of service should therefore address:
- That the user holds the rights necessary for the data they input
- A prohibition on inputs that infringe third-party rights
- That no warranty is given that output does not infringe third-party rights
- Ownership where rights do subsist in the output
- Takedown handling for unlawful or infringing content
6 Reflect AI-specific risks in the terms of service
Beyond the provisions used for any web service, the terms for a generative-AI service need to deal with:
- That output may contain errors
- That accuracy, completeness and currency are not warranted
- That output should not be relied on directly for medical, legal or investment decisions
- Prohibited prompts and uses
- That the user holds the rights necessary for their input data
- That input may be processed for the purpose of providing the service
- Rights in the output
- Handling of output that infringes third-party rights
- That output may change as models and specifications change
- What happens if the service or the API is suspended
- Liability and its limitation
A caution
Writing "this is AI, so we accept no liability" does not resolve the problem. The allocation of responsibility in the terms and the reduction of risk in the design of the service are both required. The requirements for incorporating terms as standard terms, and the limits on exclusion clauses, are covered in a separate article.
In April 2026 METI published a "Guide to the interpretation and application of civil liability in the use of AI", setting out how tort liability and related doctrines apply under current law where an AI service causes loss. The standard of care expected of a business varies with how the AI is used, the degree of human supervision, and the response to foreseeable risks.
7 Plan for loss caused by hallucination
Generative AI produces plausible statements that are not true. For a casual conversational product the consequences are limited; for services used in medicine, law, finance and investment, credit, recruitment, education or significant operational decisions, an incorrect answer carries weight.
Depending on the risk profile, consider:
- Restricting the reference material through RAG
- Displaying sources
- Displaying a confidence level
- Guardrails that decline to answer in defined cases
- Human review in the loop
- Not allowing AI alone to complete important decisions
- Retaining logs
- Keeping the system in a state where an incident can be investigated
The AI Business Operator Guidelines likewise treat safety, transparency and accountability as core common principles.
8 Check reputation, privacy, likeness and other third-party rights
Copyright is not the only exposure. If the model answers a question such as "has the president of Company X committed a crime?" by generating a conviction that does not exist, defamation and related claims can follow.
For image and voice generation, the issues extend to likeness rights, publicity rights, privacy, trade marks, the Unfair Competition Prevention Act and reputation. Where a service can generate images or voices of real people, countermeasures against deepfakes and impersonation belong in the design.
9 Check the sector-specific regulation that applies to your function
There is no single industry called "AI services". The statutes that apply change with what the AI actually does. Each of the following fields has its own regulation, and in several of them professional licensing rules as well.
- Medical and healthcare
- Finance and investment
- Insurance
- Credit assessment
- Recruitment and staffing
- Legal advice
- Real estate
- Education
- Services for children
Why this belongs at the design stage
Replacing a human process with AI does not remove the regulation that applied to it. The question to ask of the business model is: what happens when an act that requires a licence or a professional qualification when performed by a person is performed by AI instead?
Discovering this after development is well advanced often means redesigning the service. Check sector regulation and licensing as early in the design as you can.
10 Check advertising claims and charging
Because generative AI is new, marketing tends to make strong claims about capability. If you intend to say "100% accurate", "equivalent to a professional", "never wrong", "fully automated" or "the highest accuracy in the industry", the substantiation has to exist.
A representation that leads consumers to believe the quality is significantly better than it is may amount to a misleading representation of superiority under the Premiums and Representations Act. A representation made in error, without intent, can still be caught.
Where paid plans or subscriptions are sold online to consumers, the distance selling rules of the Act on Specified Commercial Transactions also apply, with requirements on the display of price, payment timing and cancellation terms, and on the final confirmation screen.
11 Design security and logging
Generative AI carries security risks that differ from those of a conventional web service:
- Prompt injection
- Leakage of the system prompt
- Leakage of confidential information held in the RAG corpus
- Output of another user's data
- Leakage of API keys
- Unintended actions by AI agents that call external tools
So establish, at the design stage, who can reach which data through which model.
To make an incident investigable, consider retaining logs, to the extent necessary, of:
- Prompts
- AI responses
- The model used
- The model version
- The material referenced through RAG
- Actions executed by AI agents
METI's contract checklist likewise treats security levels, audit rights and log retention as discrete items.
12 Do not leave AI governance until after launch
A single legal review is not enough. Models are updated, API terms change, training and retention policies change, features are added, new AI rules and guidelines are published, and new risks and attack techniques emerge.
A continuing cycle
Design → risk assessment → release → monitoring → incident response → improvement
Put a mechanism of this kind in place, sized to the scale and risk of the service. Version 1.2 of the AI Business Operator Guidelines, published on 31 March 2026, provides materials and checklists on establishing and monitoring AI governance.
In conclusion
Launching a generative-AI service in Japan is not a matter of checking compliance with the AI Act alone. In practice it means working across the AI Act and AI governance, the APPI, copyright and IP, contracts, civil liability, consumer law, and whatever sector-specific regulation applies to the function.
The most important point is not to write terms of service once the product is built, but to check the legal risks during design and, where necessary, change the product specification.
Measures such as "do not send user personal data to the external model", "do not let AI alone complete an important decision" and "only put data we may use into the RAG corpus" cannot be achieved by terms of service.
Conclusion
Legal work on a generative-AI service is not only applying law to a finished product. It is a question about the design of the service itself: what arrangement achieves the commercial objective while keeping the legal risk manageable.
Principal sources
- Cabinet Office, Act on Promotion of Research, Development and Utilisation of AI-Related Technologies
- e-Gov, the text of the same Act
- MIC and METI, AI Business Operator Guidelines (version 1.2)
- METI, Checklist for contracts on the use and development of AI
- METI, Guide to the interpretation and application of civil liability in the use of AI
- Personal Information Protection Commission, alert on the use of generative AI services
- Personal Information Protection Commission, Q&A on the APPI guidelines
- Agency for Cultural Affairs, Views on AI and copyright and Checklist and guidance on AI and copyright
- Consumer Affairs Agency, on misleading representations of superiority
- Consumer Affairs Agency, guide to the Act on Specified Commercial Transactions
These documents are published in Japanese. They are indexed, with 343 others, in the guideline index filterable by industry and function.
This article organises the general legal issues as at September 2026 and is not legal advice on any specific matter. Any particular service requires individual analysis against its business model, technical architecture and contracts. For advice from the design stage of a generative-AI service for the Japanese market, please use the contact form.