Japan's AI Promotion Act: What It Actually Requires

Japan passed an AI law in 2025 and it contains no fines, no prohibitions and no licensing. That is not an oversight or a first step to be hardened later; it is the form of statute Japan chose, and the form tells you what it is for.

What was enacted, and when

The Act on the Promotion of Research and Development and Utilisation of AI-Related Technologies (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) was passed by the National Diet on 28 May 2025 and promulgated in June 2025, with provisions taking effect on promulgation and by cabinet order thereafter. It is short—on the order of thirty articles—and most of it addresses the government rather than the governed. The authoritative text is Japanese; it is searchable through the e-Gov law search operated by Japan’s Digital Agency, and the Cabinet Office publishes the surrounding strategy documents.

Why it is a “basic act”

Japan has a recognised statutory form called a kihon-hō, a basic act, used across policy areas from disaster management to space development. A basic act declares a national objective, states principles, allocates responsibilities among the state, local government, research institutions, business operators and citizens, requires the government to produce a plan, and creates the machinery to produce it. It does not usually create individually enforceable duties, and it is not read as a failed attempt to do so.

This Act does exactly that. It states the purpose and principles in its opening articles, sets out responsibilities role by role, requires the Cabinet to adopt a Basic Plan on AI, and establishes an AI Strategy Headquarters within the Cabinet chaired by the Prime Minister with ministers as members. The Headquarters is the substantive creation: it gives AI policy a permanent home with cross-ministerial authority, which in Japanese administrative practice matters more than the absence of a penalty clause suggests.

Reading it against the EU AI Act is reading two different kinds of document. The EU AI Act is a product-safety regulation: it classifies systems by risk, attaches obligations to classes, and enforces through market surveillance and fines—see how its risk tiers are drawn. Japan’s Act classifies nothing and prohibits nothing. The comparison people reach for—“light touch versus strict”—understates the difference. They are not the same instrument dialled to different settings.

What it asks of businesses

Business operators are asked to cooperate with the state’s measures and to have regard to the principles when developing and using AI. In Japanese drafting this is a doryoku gimu, a best-efforts duty: a real statutory expression, cited in guidance and occasionally relevant to how a court reads reasonable conduct, but not a duty whose breach is itself sanctioned.

The practical content of “have regard to the principles” is supplied elsewhere, and this is the connection that makes the Act matter: the substantive expectations live in the METI and MIC AI Guidelines for Business, which the Act’s architecture leaves in place as the operating manual. A statute that says “follow the principles” plus guidance that says what the principles require is a soft-law system with a statutory anchor, and that is the design.

The enforcement mechanism is publicity

The Act empowers the state to gather information, to investigate cases where AI-related technologies have been used in ways that infringe rights or interests, to give guidance and advice, and to publish the results of its analysis. There is no fine attached and no order a company must obey. What there is, is the prospect of a government investigation naming a company in a published finding.

Whether that is weak or strong depends on assumptions about the market rather than about the law, and it is worth being honest that this is contested. Commentators sympathetic to the approach point to the reputational weight a ministerial finding carries in Japan and to the speed advantage of guidance over legislation. Critics point out that a mechanism with no order and no penalty offers an affected individual nothing, and that a foreign company with no Japanese consumer brand has little exposed to it. Both observations are correct; which dominates has not been tested, because at the time of writing no such published finding has established what the process looks like in practice.

What would change this

Three things would, and watching for them is more useful than watching for amendments to the Act itself.

• The Basic Plan. The Act requires one, and the plan is where priorities, timelines and any sectoral pressure will appear. A plan that directs a ministry to produce binding rules in a particular sector does more than the Act does.
• Sectoral regulators acting under their own powers. The Financial Services Agency, the Personal Information Protection Commission and the medical device regulator each already have binding authority and do not need this Act to use it. A binding Japanese AI rule, when one arrives, is more likely to come from one of them than from an amendment here.
• Extraterritorial pressure. Japanese exporters selling into the EU are subject to the AI Act’s reach outside the Union, which imports EU obligations into Japanese supply chains regardless of what Japanese law requires. For a large part of Japanese industry the binding AI rules are already foreign ones.

The one conclusion to avoid is that the Act makes Japan a jurisdiction without AI obligations. It makes Japan a jurisdiction whose AI obligations are located in general law and in guidance, with a statute on top that sets direction and builds machinery. That is a different compliance exercise, not an absent one.