Japan AI Policy, in depth: the AI Promotion Act, the guidelines around it, and a controls register for teams shipping AI in Japan (October 2026)

The AI Promotion Act: a law that promotes

The Act on Promotion of Research and Development, and Utilization of Artificial Intelligence-Related Technology (Act No. 53 of 2025), usually called the AI Promotion Act, passed the Diet on 28 May 2025 and was promulgated on 4 June 2025. Most of it took effect on promulgation; the chapters on the AI Basic Plan and the AI Strategic Headquarters took effect on 1 September 2025.

Its provisions that matter to engineers are short:

There are no fines and no criminal penalties in the Act. Commentators and officials have said that in serious cases the government may publicise the names of operators involved; treat that as a reputational risk rather than a statutory sanction, because the published text of Article 16 speaks only of research, analysis and guidance.

The guideline layer: Basic Plan, Article 13 and the Guidelines for Business

Three documents carry the detail the Act leaves out.

The AI Basic Plan. The Cabinet adopted the first plan on 23 December 2025 and a revised plan on 14 July 2026, about seven months later. Reporting on the revision highlights support for industry-specific ("vertical") AI and physical AI such as robot control, and stronger defences against AI-enabled cyber attacks. The plan steers funding and ministries; it does not impose duties on companies, but it predicts where future guidance will land.

The Article 13 appropriateness guideline. Decided by the AI Strategic Headquarters on 19 December 2025, it is organised in four parts: basic principles, points for developers and business users, points for national and local government, and points for citizens. It lists technical risks such as misjudgement and hallucination and social risks such as disinformation, discrimination, criminal use, over-reliance, privacy and property infringement, environmental load and cybersecurity, and asks operators to respond in proportion to their size, role and the risk involved.

The AI Guidelines for Business. Published jointly by the Ministry of Internal Affairs and Communications (MIC) and METI, these are the most operational document. Version 1.0 appeared on 19 April 2024, 1.01 on 22 November 2024, 1.1 on 28 March 2025 and 1.2 on 31 March 2026. Version 1.2 adds risk analysis for AI agents and physical AI. The guidelines assign expectations to three roles, the AI developer, the AI provider and the AI business user, and set ten common principles: human-centricity, safety, fairness, privacy protection, security, transparency, accountability, education and literacy, fair competition, and innovation. They are voluntary, but a regulator investigating an incident will ask how you applied them.

Where the binding rules are: APPI, copyright and sector law

The binding obligations for an AI product in Japan mostly come from laws that predate it.

Personal information (APPI). The Act on the Protection of Personal Information requires a specified and notified purpose of use, consent before acquiring special care-required information such as medical history or criminal records, and, for transfers to a third party abroad, consent or an equivalent-measures arrangement unless the destination is designated as equivalent, as the EU and EEA are. In June 2023 the Personal Information Protection Commission issued a caution to OpenAI and a public alert on generative AI, warning businesses not to put personal data into prompts beyond the purpose of use. For engineers this means: prompt logs, retrieval corpora, fine-tuning sets and vendor model APIs hosted outside Japan are all in scope.

Copyright. Article 30-4 of the Copyright Act permits using works for purposes that do not involve enjoying the expression, which covers much machine learning, but not where use would unreasonably prejudice the rights holder, and not when the purpose includes reproducing the expression in output. The Agency for Cultural Affairs published a general understanding of AI and copyright in 2024 explaining these limits. Infringement by outputs is judged by the ordinary tests of similarity and reliance, so a model that regurgitates a work can infringe even if training was lawful.

Sector law. Financial services, medical devices, employment and consumer protection rules apply to AI-assisted decisions in the same way they apply to any other decision. A credit model, a diagnostic aid or a hiring screen is regulated by its sector first.

Turning it into a controls register

Do not build a separate "Japan compliance" programme. Build one register of controls in which each control lists every source that motivates it, the role it applies to, and the evidence that proves it runs. Japan then becomes a set of tags on controls you mostly already have. The sketch below is enough to start.

from dataclasses import dataclass, field

@dataclass
class Control:
    cid: str
    what: str
    roles: set            # {"developer", "provider", "business_user"}
    sources: list         # citations; "binding" ones decide priority
    evidence: list        # artefacts an auditor or ministry could ask for
    owner: str = ""
    last_evidence: str = ""   # ISO date of the most recent artefact

REGISTER = [
    Control("PRIV-01", "No personal data in prompts beyond the notified purpose",
            {"provider", "business_user"},
            ["binding: APPI purpose of use", "soft: Guidelines for Business, privacy"],
            ["prompt-log PII scan report", "privacy notice version"]),
    Control("PRIV-02", "Approved mechanism before sending data to a model API abroad",
            {"provider", "business_user"},
            ["binding: APPI transfers to third parties abroad"],
            ["vendor register with hosting region", "consent or contract record"]),
    Control("COPY-01", "Output filter and takedown path for reproduced works",
            {"developer", "provider"},
            ["binding: Copyright Act, output infringement", "soft: Art. 13 guideline"],
            ["near-duplicate filter test results", "takedown tickets"]),
    Control("INC-01", "Rights-infringement incidents triaged and answerable within 5 days",
            {"developer", "provider", "business_user"},
            ["soft: AI Promotion Act Art. 7 and 16", "soft: Art. 13 guideline"],
            ["incident log", "inquiry response pack"]),
    Control("AGENT-01", "Human approval for irreversible agent actions",
            {"provider", "business_user"},
            ["soft: Guidelines for Business v1.2, AI agents"],
            ["action policy", "approval audit trail"]),
]

def gaps(register, role, today, max_age_days=90):
    from datetime import date
    out = []
    for c in register:
        if role not in c.roles:
            continue
        binding = any(s.startswith("binding") for s in c.sources)
        stale = (not c.last_evidence or
                 (date.fromisoformat(today) - date.fromisoformat(c.last_evidence)).days > max_age_days)
        if not c.owner or stale:
            out.append((0 if binding else 1, c.cid, c.what))
    return sorted(out)   # binding gaps first

Two design choices matter. Sorting binding gaps first keeps the team honest about what can actually be penalised. And recording evidence age turns the register from a document into a monitor: a control whose last artefact is a year old is a control nobody runs.

Worked example: a support assistant for Japanese banks

Consider a Singapore software company launching a generative customer-support assistant for Japanese retail banks. It hosts the model in the United States, retrieves from each bank's knowledge base, and logs every conversation.

1. Roles. The company is an AI provider; each bank is an AI business user. The model vendor behind the API is the developer. Write this down per contract, because the Guidelines' expectations differ by role.
2. Binding first. Conversation logs contain names and account details, so APPI applies. Sending them to US hosting is a transfer abroad to a non-designated country: each bank needs consent or an equivalent-measures arrangement and the provider must support it contractually (PRIV-02). Financial regulation applies to anything the assistant says about products, so answers on rates and fees come from retrieval with citations, never from free generation.
3. Soft layer next. Map transparency and accountability principles to concrete artefacts: a notice telling customers they are talking to AI, a model card, and a log that can reconstruct any answer.
4. Incident path. A customer complains that the assistant disclosed another person's balance. INC-01 triggers: preserve logs, contain the retrieval bug, notify the bank, and assess whether the APPI breach-reporting rules apply. If a ministry later asks questions under Article 16, the response pack is already assembled.
5. Agents. When the bank later asks the assistant to execute transfers, AGENT-01 applies: irreversible actions need explicit human confirmation.

Nothing here required a Japan-only control. Every item tightened a control the company should run in any market; Japan changed the tags, the transfer mechanism and the language of the notices.

Japan among other approaches

Placing Japan next to its neighbours helps with multi-market products.

Japan also exports its approach internationally. It led the G7 Hiroshima AI Process, which produced guiding principles and a code of conduct for organisations developing advanced AI in 2023, and in 2024 it set up an AI Safety Institute that publishes evaluation guidance. Expect Japanese guidelines to track those international documents, which Article 13 explicitly requires.

Failure modes

The most common mistakes teams make with Japan:

• "No fines, so nothing to do." APPI, copyright and sector law are fully enforceable, and the guidelines describe what a reasonable operator does when something goes wrong.
• Treating the guidelines as a checklist. They ask for risk-proportionate judgement. A ticked box with no evidence of reasoning is weaker than a short documented risk assessment.
• Ignoring role changes. A business user that fine-tunes a model may become a developer under the Guidelines' definitions, with heavier expectations.
• Version drift. Citing version 1.1 after 1.2 is out signals a stale programme. Version the citations in your register and review them when a new edition appears.
• English-only monitoring. Many updates appear in Japanese first. Monitor the Japanese pages of the Cabinet Office, METI, MIC and the Personal Information Protection Commission.

Trade-offs and further reading

Japan's model trades legal certainty for flexibility. Guidelines can be revised in months, as the Basic Plan was, so the rules keep pace with agents and physical AI, but you cannot point to a statutory safe harbour. Building one global register costs more up front than a per-country checklist and pays back on the second market. For keeping it current, the AI regulatory watch article shows how to detect and triage changes, and the AI regulation deep dive shows a dated obligation register across jurisdictions. Compare the binding approaches in South Korea's AI Framework Act and the principles approach in UK AI policy, and see AI governance programme structure for who owns the register.

What to do next

1. List every AI system you offer or use in Japan and record your role for each: developer, provider or business user.
2. Map personal data flows for each system, including prompt logs and vendor APIs hosted outside Japan, and confirm the APPI basis for each transfer.
3. Add Japan source tags to your existing controls register, binding sources first, and record evidence dates.
4. Write a one-page incident playbook that can answer a ministry inquiry about rights infringement within a week.
5. Add an output near-duplicate filter and a takedown path for copyrighted works.
6. Subscribe to the Japanese-language pages for the Basic Plan, the Article 13 guideline and the AI Guidelines for Business, and review the register when any of them changes.