Japan AI Promotion Act Guide: What It Means for Organisations

Japan's first AI statute sets no fines and no licensing, yet organisations that develop, provide or use AI in Japan are now expected to cooperate with government measures and follow a growing body of official guidelines that regulators, courts and customers will treat as the standard of care. The Act on Promotion of Research and Development, and Utilization of Artificial Intelligence-related Technology (Act No. 53 of 2025) is a framework law: it creates the AI Strategic Headquarters, the AI Basic Plan and a statutory basis for guidelines, rather than prohibitions. This guide explains what the Act covers, how it connects to the AI Basic Plan, the Article 13 guideline and the METI and MIC AI Guidelines for Business Version 1.2, and what a security and governance team should do about it, including for AI agents and the identities they run under. It is for compliance, security and product leads at Japanese organisations and at foreign firms serving the Japanese market, and for executives comparing Japan with the EU and Chinese regimes. English texts cited here are the government's provisional translations; the Japanese originals prevail.

Key takeaways

• It is a promotion law, not a risk law. The Act sets basic principles and responsibilities and contains no penalty provisions, unlike the EU AI Act.
• Businesses have a statutory duty to cooperate. Article 7 says utilisation business operators must cooperate with measures the national and local governments implement under Articles 4 and 5.
• The soft law now has a statutory anchor. The Article 13 guideline (19 December 2025) and the AI Guidelines for Business Version 1.2 (31 March 2026) set out what appropriate AI use looks like.
• The government can investigate and advise. Article 16 lets it collect information, analyse cases where AI misuse harms people's rights, and give guidance, advice and information to businesses.
• Version 1.2 addresses AI agents directly. It asks businesses to manage permissions for connected tools, decide where human judgment is needed and review operation logs, which are identity and access controls in practice.

Scope: who and what is covered

• AI-related technology (Art. 2): technology needed to substitute artificial means for human cognition, inference and judgment, and the information processing systems that use it to produce output from input. The definition is broad and technology-neutral; there is no risk tiering.
• National government (Art. 4): plans and implements comprehensive measures, and promotes AI use within administrative bodies.
• Local governments (Art. 5): devise measures suited to their areas.
• Research and development institutes (Art. 6): universities and research bodies conduct research, share results, develop talent and cooperate with government measures.
• Utilisation business operators (Art. 7): businesses that develop or provide AI products or services, or use AI in their business activities. They are to endeavour to make their business more efficient and advanced and create new industries through AI, and must cooperate with government measures.
• Citizens (Art. 8): deepen their understanding of AI and cooperate with measures.

The translated text does not distinguish domestic from foreign businesses and contains no express extraterritorial clause; the practical effect on a foreign provider comes through the guidelines and through Article 16 information gathering about services used in Japan. Personal data used by AI remains governed by the Act on the Protection of Personal Information. See the Japan APPI Guide.

The five basic principles (Art. 3)

1. Consistency with the Basic Act on Science, Technology and Innovation and the principles for forming a digital society.
2. Maintaining Japan's R&D capability and industrial competitiveness, recognising AI as important for national security.
3. Comprehensive and systematic promotion from basic research through to everyday and economic use.
4. Ensuring appropriateness and transparency to address risks such as crime, leaks of personal information and copyright infringement.
5. International cooperation, with Japan taking a leading role.

Key obligations

Because the Act binds the state more than businesses, the obligations an organisation actually works to come from three layers.

What the Article 13 guideline asks

• Risk management and governance: identify, assess and respond to risks across the AI lifecycle, with management-level monitoring, disclosure and training, improved continuously.
• Transparency: explain, within reasonable limits, the origin of training data and outputs, and tell users about the system's mechanisms, limitations, prohibited uses and output reliability.
• Safety and security: identify misuse risks such as cyber attacks and fraud; reduce hallucination, bias and deepfake harms; develop ways to identify AI-generated content such as digital watermarks and provenance management.
• Continuity: a business continuity plan covering normal operation and emergency response to system failure.
• Human-centric decisions: the scope and conditions of AI use should be subject to a final decision made by a human.
• Government side: each ministry appoints an AI governance officer, and local governments are asked to designate responsible officers.

What the AI Guidelines for Business add

• Roles: obligations are split between the AI developer, the AI provider and the AI business user. Version 1.2 treats retrieval-augmented generation as inference, which generally sits with the provider, while fine-tuning for alignment sits with the developer, according to PwC Japan's summary.
• Common principles: the guidelines are built around common guiding principles: human-centric, safety, fairness, privacy protection, ensuring security, transparency, accountability, education and literacy, fair competition and innovation.
• AI agents and physical AI (new in 1.2): an AI agent is defined as a system that perceives its environment and acts autonomously to reach a goal; agentic AI combines several agents; physical AI acts through actuators. Listed risks include data poisoning across more input routes, hallucination-driven actions such as unintended orders or file deletion, leakage through external integrations and misuse for malware and intrusion.
• Human judgment: businesses should decide which matters need human judgment based on importance and risk. Version 1.2 also flags automation bias and the risk that human-in-the-loop oversight becomes a formality.
• Liability signal: the Kitahama Partners summary notes the guidelines warn that failing to analyse risks, or not requiring human decisions despite knowing of serious risks, could support legal liability if an accident occurs. This is how soft law bites in Japan: through negligence claims and contracts, not regulator fines.

Timeline and current status

Both the Basic Plan and the business guidelines are on annual review cycles, so a revised plan around the end of 2026 and a further guideline version in 2027 are likely. Check the Cabinet Office AI strategy pages and METI before relying on this table.

The AI Basic Plan in brief

The plan, titled "Japan Rebooted" through "Trustworthy AI", has four pillars: Adopt AI, Create AI, Enhance AI Trustworthiness and Collaborate with AI, with three principles: innovation alongside risk mitigation, agile PDCA style response and integrated domestic and international policy. On safety it plans to roughly double the staff of Japan's AI Safety Institute, benchmarked against the UK AI Security Institute's roughly 200 personnel, widen its evaluation scope, and build evaluation testbeds. It names deepfakes, disinformation (supporting watermarking and authenticity verification), cyber attacks and AI-enabled fraud as risks, and discusses AI agents and an "AI economic sphere" in which agents transact with each other.

Enforcement and penalties

• No sanction does not mean no consequence: sector law still applies in full. APPI for personal data, copyright, consumer protection and financial regulation govern AI harms, and the Act's own principles name crime, personal information leaks and copyright infringement.
• Information gathering: under Article 16 the government studies cases where AI misuse harms citizens' rights and may give guidance or advice to the businesses concerned. Expect requests for information after high-profile incidents.

See the EU AI Act Compliance Guide and the China Generative AI Measures and AI Labelling Rules Guide for the binding regimes.

What it means for AI agents and non-human identities

Version 1.2 of the AI Guidelines for Business is one of the first government guidelines to give specific attention to AI agents. Its three agent controls map directly onto identity practice.

The guideline's example of hallucination-driven file deletion is not hypothetical: in the PocketOS database deletion, a coding agent found an account-scoped token in the codebase and deleted a production database and its backups. Scoped credentials and a human gate on destructive actions are exactly what Version 1.2 asks for. See the AI Agent Authorisation Guide and the Identity Governance for AI Agents Guide.

Framework mapping

Compliance steps

1. Map your roles. For each AI system, record whether you act as AI developer, AI provider or AI business user in the guidelines' terms. One organisation is often all three.
2. Gap-assess against Version 1.2 and the Article 13 guideline. Use the role-specific appendices as a checklist; small organisations can start with the business user checklist only.
3. Set up governance. Name an accountable executive, a cross-functional AI committee and a policy that references the guidelines. Large groups can align this with an ISO/IEC 42001 management system; smaller firms can run a lighter register and quarterly review.
4. Inventory AI agents and their identities. List each agent, its tools, credentials, owners and the actions it can take. Remove shared and long-lived keys first.
5. Define human judgment points. Classify agent actions by impact and require approval for irreversible or external actions. Pilot on one workflow in observe-only mode, compare agent proposals with human decisions, then enable automation; keep a switch that returns the agent to propose-only mode as a rollback.
6. Log and review. Capture tool calls and decisions, and schedule a log review that someone actually performs.
7. Prepare for information requests. Keep documentation on training data origin, limitations and incidents ready, because Article 16 and Article 25(2) requests are the most likely government touchpoint.
8. Track revisions. Re-run the gap assessment when the Basic Plan or the guidelines are revised.

Detection and measurement

• Coverage: share of AI systems with a recorded role, owner and risk assessment.
• Agent credential hygiene: share of agents with a dedicated, scoped, short-lived credential (target 100%).
• Human gate coverage: share of high-impact agent actions that pass through an approval step.
• Approval quality: rate at which reviewers reject or edit agent proposals; a rate near zero can signal the formal, rubber-stamp oversight Version 1.2 warns about.
• Log review cadence: days since the last documented review of agent operation logs.
• Guideline currency: days between a new guideline version and completion of the gap reassessment.

Common failure patterns

• Reading "no penalties" as "no obligations": sector law, contracts and civil liability still apply, and the guidelines shape what a court may see as reasonable care.
• Working from Version 1.0 or 1.1: older internal policies miss the AI agent and physical AI content added on 31 March 2026.
• Formal human-in-the-loop: approvals clicked through without review, which Version 1.2 specifically calls out.
• Agents on shared or over-scoped keys: the PocketOS case shows how one broad token turns a reasoning error into data loss. See the API Key Management Guide.
• No owner for government requests: nobody can answer quickly when asked about training data origin or an incident.
• Treating Japan as an EU AI Act subset: Japan has no risk classes or conformity assessment, so EU documentation helps but does not map one to one.

Practitioner checklist

• Record your role (developer, provider, business user) for every AI system used in Japan.
• Confirm you are working from AI Guidelines for Business Version 1.2 and the 19 December 2025 Article 13 guideline.
• Name an accountable executive and an owner for government information requests.
• Inventory AI agents, their tools and the credentials they use.
• Replace shared or long-lived agent credentials with scoped, short-lived ones.
• Define and enforce human approval for high-impact agent actions.
• Log agent tool calls with agent and user identity, and review the logs on a schedule.
• Document training data origin, limitations and prohibited uses for each AI service.
• Check APPI and other sector obligations alongside the AI guidelines.
• Diarise annual reviews of the Basic Plan and guidelines.

Standards and references

• Japanese Law Translation: Act on Promotion of Research and Development, and Utilization of Artificial Intelligence-related Technology (Act No. 53 of 2025)
• Cabinet Office: Artificial Intelligence Basic Plan (provisional translation)
• Cabinet Office: Guideline for Ensuring the Appropriateness of R&D and Utilization of AI-Related Technology (provisional translation)
• METI: AI Guidelines for Business Ver 1.0 Compiled
• PwC Japan: AI Guidelines for Business Version 1.2 changes (Japanese)
• Kitahama Partners: AI, Blockchain and IoT Newsletter on Version 1.2
• White & Case: Japan's first AI legislation becomes law
• OWASP: Top 10 for LLM Applications 2025

This guide summarises Japan's AI Promotion Act and related government guidelines as at 24 October 2026, based on provisional English translations, and is not legal advice. Related NHI Mgmt Group resources: Global AI Regulation Map · South Korea AI Basic Act Guide · ISO/IEC 42001 Guide · NIST AI RMF Guide · Global Privacy Regulatory Map