Japan: governing without penalties
Lesson 3 of 5 in Asia-Pacific: Japan, South Korea, Singapore, India, and Australia.
Japan has spent a decade proving you can take AI seriously without threatening anyone. The lineage runs from Society 5.0 (the 2016 vision of an AI-integrated society) through the Social Principles of Human-Centric AI (2019) to the consolidated AI Guidelines for Business (METI/MIC, April 2024) — all voluntary, all iterated with industry, under a philosophy Japanese policymakers named agile governance: steer continuously with guidance, reputation, and dialogue rather than fixed statutory commands, because the technology outruns any parliament.
In 2025 Japan finally legislated — and stayed perfectly in character. The AI Promotion Act (Act on the Promotion of Research, Development and Utilization of AI-Related Technologies) passed the Diet in May 2025 and was promulgated in June 2025. It is a basic law: it declares AI development a national priority, assigns responsibilities, builds coordination machinery — and contains no penalties, no fines, and no prohibited practices. Businesses owe a duty to cooperate with government AI measures. That is the entire coercive apparatus.
The machinery matters more than the mandates. The Act creates a Cabinet-level AI Strategy Headquarters — chaired by the Prime Minister, with every minister a member — obliged to produce a national AI Basic Plan. And it hands government a quietly sharp tool: the power to investigate harmful AI incidents and publish what it finds, naming problematic actors — guidance backed by reputation in a market where reputation is the currency. Japanese lawyers call it the name-without-shame-being-optional mechanism: no fine, but no anonymity either.
Around the basic law sits a distinctive stack: Article 30-4 of the Copyright Act — since 2019 among the world’s most permissive text-and-data-mining exceptions, allowing copyrighted works to be used for machine learning without permission unless the use unreasonably prejudices the owner — made Japan a favoured jurisdiction for model training. The Japan AI Safety Institute (February 2024, under the IPA) was among the world’s first. And Japan authored the G7 Hiroshima Process, exporting its guidance-first instincts into the international code of conduct you met in the soft-law module.
Myth: “A law without penalties is just a press release”
The Act allocates budget authority, creates the PM-chaired Headquarters, obliges a national Basic Plan, and legitimises investigation-and-publication of harmful cases. In a coordinated economy, ministry guidance plus public naming moves corporate behaviour measurably — the same mechanism that makes “voluntary” administrative guidance effective across Japanese regulation generally. Binding is a spectrum, not a switch.
Myth: “Japan has no AI rules at all”
Sectoral and adjacent law binds plenty: APPI (data protection) governs training data and profiling; the pharmaceutical and medical-device regime covers medical AI; financial-services guidance covers algorithmic trading. What Japan lacks — deliberately — is horizontal AI-specific obligations. The Guidelines for Business supply the horizontal layer as soft law.
Fact: Japan’s copyright regime is an AI industrial policy
Article 30-4 permits ingestion of copyrighted works for information analysis — including model training — without licence, subject to the unreasonable-prejudice carve-out. It predates the GenAI boom (2019) and has survived intense creator-industry pressure since, with the Cultural Affairs Agency issuing interpretive guidance (2024) rather than amendment. Compare the EU’s opt-out-based TDM exception, which forces model trainers to honour rights reservations.
Fact: the Act still has teeth in procurement and planning
The Basic Plan cascades into ministry budgets and public procurement expectations. A vendor ignoring the Guidelines for Business will find that “voluntary” compliance is a de facto condition of selling AI to the Japanese state and to the large firms that align with it — the soft-law-hardened-by-hooks pattern you saw with TC260 in the China module, in a politer register.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.