Japan's AI Law Has No Penalties. That Makes the Interface the Compliance Layer.

Japan has an AI law. It came into full force on 1 September 2025. It carries no penalties.

Not lighter penalties than Europe's. None. There is no fine, no prohibited practices list, no conformity assessment and no regulator that can stop you shipping. The obligations on companies are duties of effort.

Every summary I have read treats this as a story about regulation. I think it is a story about design, and I want to explain why in a way that does not require you to care about legislation.

What the law actually is

The Act on the Promotion of Research, Development and Utilisation of AI Related Technologies is Act No. 53 of 2025. It was promulgated on 4 June 2025 and came fully into force on 1 September 2025, when the provisions establishing the AI Strategy Headquarters took effect. It is Japan's first basic law on AI.

The word in the title that matters is promotion. This is a law designed to encourage a technology, not to fence it. It sets out national strategy, establishes a headquarters inside the Cabinet, and assigns responsibilities to the state, to research institutions, to local government and to businesses.

The provision that reaches an ordinary company is Article 7, which covers the responsibilities of businesses that use AI. It applies to you if you merely use generative AI in your operations, which after last year's adoption numbers is essentially every Japanese company of any size. What it asks for is cooperation with national measures and, in substance, that you actively use AI to improve and advance your own business.

Read that again. The legal duty runs towards using the technology, not towards restraining it.

Compare the European AI Act, which classifies systems by risk, prohibits some outright, imposes conformity obligations on high risk uses and attaches administrative fines that scale with global turnover. Japan looked at that and chose the opposite instrument.

I am not going to argue about which is correct. I will point out the consequence, which is not much discussed.

With no enforcement, the interface is the enforcement

When a regulator can fine you, compliance becomes an organisational function. Somebody owns it, it has a budget, and it produces documents that prove things.

When there is no penalty, none of that machinery gets built, and I have never seen a company build it voluntarily. So what actually stands between a model's output and a customer?

The screen. That is the whole list.

Whether a person sees the model's answer before the customer does. Whether the confirmation step is a real decision or a button someone clicks four hundred times a day without reading. Whether the interface shows what the answer was based on. Whether there is an undo, and how long the window is. Whether the uncertain answers look different from the confident ones. Whether the person reviewing has any way to tell.

Those are all interaction design decisions. Every one of them gets made by somebody, usually implicitly, usually in an afternoon, usually by whoever is building the feature. In a jurisdiction with penalties, a lawyer eventually audits those decisions. In Japan, nobody does.

Which means the design decision is the governance. Not a support for governance. The thing itself.

The guidelines say this almost outright

Japan's substantive AI guidance lives outside the statute, in the AI Business Operator Guidelines, jointly issued by METI and the Ministry of Internal Affairs and Communications. Version 1.2, published in March 2026, clarified the human in the loop expectation for AI agents, meaning systems that take actions rather than only producing text.

Human in the loop is discussed as though it were a policy. It is not. It is a description of a screen.

You cannot implement human in the loop by writing it in a document. You implement it by deciding where the confirmation sits, how much context the reviewing human is given, whether they can see the alternative the model rejected, how long they have, and what the default is if they do nothing. Get those wrong and you have a human in the loop who approves everything, which satisfies the guideline and protects nobody.

I have built the failing version of this. On an early build of one of my own apps I put a confirm step in front of an automated action because it felt responsible. Within a week of using it myself I was clicking through the confirmation without reading it, because it was identical every time and correct almost every time. The step was theatre. What fixed it was making the confirmation look different when the model was uncertain, so the unusual cases interrupted the habit and the ordinary ones did not ask for attention they did not need.

That is a design change. No policy document would have caught it, because on paper both versions have a human in the loop.

The Cabinet has diagnosed the real problem, and it is a design problem

On 23 December 2025 the Cabinet approved Japan's first AI Basic Plan, subtitled "Japan's revival through trustworthy AI". Its goal is for Japan to become the country where AI is easiest to develop and use in the world.

Chapter one contains two sentences more candid than anything a consultant would put in a deck. The first: in Japan, AI has not become something actively used in daily life or at work. The second: for a technology in which basic research and social implementation sit close together, the lack of implementation has become a major obstacle to Japan's AI development.

The plan is built on four pillars. Use AI, build AI, lead on AI governance, and continuously transform towards an AI society. Under the fourth, it names as a task to be worked out the division of roles between people and AI.

Deciding what a person does, what a machine does, and where the handover happens is the definition of interaction design. It has been the definition since long before anyone used the word AI. That sentence in a Cabinet decision is a national design brief, and I do not think the design industry here has noticed it is addressed to them.

Why trustworthy is the interesting bet

The plan's whole strategic argument is that Japan lost the investment race and should compete on trustworthiness instead. It says so plainly: the emphasis is on reproducing the value of reliability that Japan has built up in the physical world.

My instinct was that this was a consolation prize dressed up as a strategy. I have changed my mind, for one reason.

Trust in a product is not produced by the model. It is produced by the interface. Users cannot inspect weights or evaluate benchmarks. What they can do is notice whether the system admits when it does not know, whether it is consistent, whether the undo works, whether it explains itself at the moment they need it rather than in a help article, and whether it has ever quietly lost their work.

Every one of those is a design property. So if Japan's national AI strategy is trustworthiness, then Japan's national AI strategy is a design strategy, and the country's competitive advantage depends on a discipline that 0.8% of its manufacturers once credited for their own success.

That is either the best joke in Japanese industrial policy or the largest available opportunity here. I think it is both.

What to do with this if you ship software in Japan

The absence of penalties is genuinely useful and I would not pretend otherwise. You can build here without a conformity assessment and without a legal review gate on every model change. That is a real speed advantage and companies should use it.

What I would not do is confuse no penalty with no consequence. The consequence just arrives through a different channel, which is a customer who stops trusting the product and cannot articulate why.

Four things I would hold yourself to, none of which any Japanese regulator will ask you for.

1. Write down, for each place a model output reaches a user, who is accountable when it is wrong. If the answer is nobody, that is your highest risk surface and it took ten minutes to find.
2. Make your confirmation steps differ by confidence. An identical confirmation on every action trains the habit that defeats it.
3. Show the basis, not the reasoning. Users do not want a chain of thought. They want to know which document, which record, which field this came from, so they can check the one thing they actually doubt.
4. Make undo the primary safety mechanism rather than confirmation. Confirmation asks people to predict a problem. Undo lets them fix one they can see. Predicting is much harder and we are much worse at it.

Japan has decided not to regulate this at the level of the model. That decision pushes the entire question of whether AI behaves acceptably down into the layer where the product meets a person, and hands it to whoever is designing that layer.

If you build products here, the law just made you the compliance function. Nobody sent a memo.

Sources

• 人工知能関連技術の研究開発及び活用の推進に関する法律 (Act No. 53 of 2025), Cabinet Office. Promulgated 4 June 2025, fully in force 1 September 2025. Article 7 covers the responsibilities of businesses using AI.
• Cabinet Office, 人工知能基本計画, Cabinet decision of 23 December 2025. The implementation passage is in chapter 1. The four pillars and the division of roles between people and AI are in chapter 2.
• METI and the Ministry of Internal Affairs and Communications, AI事業者ガイドライン, version 1.2, March 2026.
• Regulation (EU) 2024/1689, the European AI Act, for the comparison.