Who is liable when generative AI causes loss in Japan?
In short
- METI's April 2026 handbook does not create a new liability regime for AI. It sets out how Japan's existing tort and product liability rules are likely to be applied when AI is involved.
- The organising idea is whether the AI was meant to assist a human decision or to replace it. The two attract different duties.
- Where AI replaces human work, requiring the user to redo every output by hand would defeat the point. The question shifts to whether the process around the AI was properly designed and supervised.
- A contract allocates risk between the parties to it. It does not bind a third party injured by the system.
- The practical test is not which label applies — developer, provider, user — but who could have known about the risk and who could have controlled it.
A generative-AI tool answers a question wrongly, a company relies on it, and loses money. An AI agent recommends something to a customer, and the customer suffers a loss. An AI inspection system passes a defective product, and that product injures someone.
Who is liable: the company that used the AI, the company that provided the service, or the company that developed the model?
On 9 April 2026, Japan's Ministry of Economy, Trade and Industry published a handbook on the interpretation and application of civil liability in the use of AI (version 1.0). It works through seven worked examples — delivery-route optimisation, legal research support, image generation, credit screening, visual inspection, autonomous mobile robots, and AI agents — and sets out how existing law may attach to each.
For in-house and external counsel, the most useful part is the distinction it draws between AI that assists a human decision and AI that is intended to replace it.
This is not a new liability regime
The handbook creates no new duties. It is an interpretive document.
What it addresses is a practical problem: AI systems are opaque, partly autonomous, and built by several companies in a chain, which makes it hard to see where responsibility sits. The handbook maps the existing rules — tort under Article 709 of the Civil Code, the Product Liability Act, and related doctrines — onto those situations.
It follows that meeting the factors it lists does not guarantee a defence, and failing to meet them does not establish liability. A court decides each case. The handbook is better read as a description of which facts are likely to matter than as a safe-harbour checklist.
Separate who was doing what
A single AI product usually involves several companies: the developer of the foundation model, the company offering it as an API, the SaaS company that embeds the API, and the customer who uses the SaaS.
| Role | What they typically do | What tends to be examined |
|---|---|---|
| Developer | Builds the model, the algorithm, the training and evaluation method | Design safety, training and evaluation, treatment of known limitations, disclosure |
| Provider | Offers the AI as a service or system | Fitness for the stated purpose, conditions of use, warnings, updates, incident response |
| User | Uses the AI in operations, decisions or customer-facing work | Appropriateness of the use, the checks carried out, human judgement, internal procedure |
In practice one company occupies several roles, and the company in the middle is a provider downstream and something close to a user upstream. So the question is not which box a company's name goes in, but, for the function that failed, what did each party design, what did each party know, and what could each party control.
"The AI got it wrong" does not decide anything
The fact of a wrong output does not establish liability. Tort still requires fault, infringement of a right or interest, loss, and causation.
What makes AI harder than ordinary software is that foreseeability and avoidability are distributed. The developer may know the model degrades under particular conditions; the user may be the only one who knows the facts of the case the model was asked about; the provider may be the only one who can see the logs.
And the expected level of human involvement changes the analysis entirely. If the service is sold as producing suggestions for a professional to verify, the user's failure to verify matters a great deal. If the service is sold precisely on the basis that it removes the need for manual checking, demanding that the user manually check everything contradicts the product.
Assistive AI and substitutive AI
| Assistive / supporting AI | Reliance / substitutive AI | |
|---|---|---|
| Concept | The AI supports a decision; a human decides | The AI is intended to take over all or a significant part of the human decision |
| Human involvement | Reviews output, investigates further where needed | Supervises the system rather than re-performing each output |
| Centre of the analysis | Did the user take appropriate care in verifying? Did the provider disclose limits the user could not see? | Was substitution reasonable? Is accuracy sufficient for it? Were process, monitoring, stop and update arrangements adequate? |
The category is not fixed by product type. The same model can sit in either depending on how it is sold, documented and actually used.
A contract-review tool used to surface issues for a lawyer to confirm is assistive. The same tool configured to approve standard contracts without human review is closer to substitutive. Writing "this output is for reference only" in the terms does not by itself make a system assistive if the marketing, the design and the deployment all say otherwise.
Assistive AI: what the user has to do
Where a human makes the final decision, the user is expected to evaluate the output rather than adopt it unconditionally.
That does not mean redoing everything. How much verification is appropriate depends on the size of the potential loss, the accuracy of the tool, the expertise of the user, the cost of checking, and what the provider disclosed. Summarising an internal document and deciding on someone's credit application are not the same exercise.
Nor does human review discharge the provider. If the developer or provider knows about a weakness that the user cannot observe from the interface — degraded accuracy under particular conditions, categories of input the model was not built for — that is a design and disclosure question regardless of who reviews the output.
Worth confirming for assistive AI
- Is it decided who makes the final check on AI output?
- Is it decided when to escalate or investigate further?
- Are the conditions the model handles badly communicated in a form users can act on?
- Are logs kept, so that the basis of a decision can be reconstructed afterwards?
- Is any external decision being made on the sole ground that the AI said so?
Substitutive AI: not re-checking is not automatically negligent
Where AI is intended to replace human work, the handbook treats two things as conditions of that substitution being reasonable: that there is a genuine need or benefit in reducing human involvement, and that accuracy and safety are sufficient for the task being handed over.
Given that, requiring the user to re-perform by hand every task the AI was bought to perform would remove the point of the system. So the user's duty shifts from catching each individual error to designing and managing the process around the AI: deciding where it may be relied on, detecting the conditions where it should not be, monitoring for anomalies, being able to stop and roll back, and re-evaluating performance after a model update.
Correspondingly, more is expected of the provider and developer. If the product is sold on the basis that customers will rely on it, accuracy and safety appropriate to that purpose, risk-reduction measures, clear statements of limits, and disciplined updating all carry more weight than they would for an assistive tool.
What the seven examples have in common
| Example | Issue | Practical reading |
|---|---|---|
| Delivery route optimisation | A driver follows a suggested route and an accident occurs | Separate what the driver could check from hazard information only the routing system held |
| Legal research support | AI assists research or drafting and an error results | For professional tools, the professional's own verification matters — but limits only the provider knows are a separate question |
| Image generation | Output infringes a third party's rights | Separate the user's prompt and review from the model's own risk-reduction measures and disclosure |
| Credit screening | Unsuitable outcome in a screening decision | The weight of the decision and the reality of human review change what verification is required |
| Visual inspection | AI that replaced manual inspection misses a defect | If substitution assumed performance at least equal to humans, the focus is process safety, not re-inspection |
| Autonomous mobile robot | Autonomous movement causes a physical accident | Design, safety mechanisms, operating environment, maintenance and update responsibility |
| AI agent | AI proposes, purchases or executes on a person's behalf | Scope of authority, confirmation points, integration with external services, stop and recovery, logs |
Across all of them the method is the same: look past what happened inside the model, at the business process the AI sat in, the role each party had, the information each held, and who was in a position to prevent the loss.
Does a contract settle it?
Between the parties, largely yes — and that is worth doing properly. Warranty scope, liability caps, allocation of third-party IP claims and the division of investigation costs can all be agreed in advance, and the ten points to check in an AI contract covers how.
But a contract binds the parties to it. If a person who is not a party is injured by the system, a limitation of liability in the terms of service does not automatically bind them.
So an AI contract is better designed as a statement of who controls which risk and who bears the cost of investigating and responding to an incident, rather than as an attempt to make liability disappear.
What the chain means in practice
The longer the AI value chain, the more that information and control sit in different companies.
The developer knows the model's weaknesses. The provider holds the logs. The user knows the facts of the particular case. None of them can see the whole picture, and the loss happens at the join.
That is why the contractual work and the operational work are the same work: logs, disclosure documents, escalation rules and change notifications exist so that, when something goes wrong, what each party knew and did can actually be established.
Worth confirming now
- Classify each AI use in the business as assistive or substitutive, by task rather than by product
- Write the boundary between what the AI decides and what a human confirms into internal rules and specifications
- For uses that could cause serious loss, record how accuracy and safety were assessed
- Check the upstream vendor's terms: liability cap, log provision, incident response, model change
- Check that what you promise customers does not exceed what the upstream contract gives you
- Check that inputs, outputs, model version, settings and the human decision can be traced after an incident
- Where third parties could be harmed, handle the risk through design, operation and insurance as well as contract
In short
When AI causes loss, the fact that AI was involved does not locate the responsibility.
For assistive AI, the question is whether the user properly evaluated the output and made the decision. For substitutive AI, the question is less whether a human repeated the work and more whether the accuracy justified relying on the system and whether the surrounding process was sound.
Liability is not distributed by label. It follows information and control — which is why the time to organise logs, disclosure, change management and indemnities is before the incident, not after.