Japan's AI Regulation Stance
Japan takes a soft-law-leaning approach combining sector-specific laws + cross-cutting guidelines, not the EU-style "bind tightly with a single hard law." Behind this is the policy intent of "not wanting to dull industrial competitiveness with over-regulation."
Cross-Cutting Guidelines
AI Business Operator Guidelines (April 2024)
Jointly published by MIC and METI. They organize common guidance and specific items by dividing operators into "develop," "provide," "use" of AI.
- Common: human-centric, fairness, privacy, security, transparency, accountability, education
- Developers: data-quality management, technical verification, information provision
- Providers: responsibility allocation in contracts, explanation to users
- Users: appropriate operation, result confirmation
Not legally binding, but a de facto standard for government procurement and listed-company governance.
Copyright Law and AI Training
Japan's Copyright Act Article 30-4 stipulates "use without rightsholders' consent is allowed when provided for information analysis," making Japan one of the world's most permissive countries for AI training-data use. However, these issues remain.
- Interpretation of enjoyment purpose: output merely intended to be appreciated is out of scope (can be infringement)
- Similarity judgment: if output has "expressive commonality" with the training source, normal copyright-infringement judgment applies
- March 2024 ACA view: imitating a style isn't infringement, but if it shares an existing work's expressive features, it is
- Creator-protection debate: opt-out systems and scope expansion continue to be debated
Sector-Specific Regulation
Medical AI
Diagnostic-support AI needs PMDA approval as a "medical device program (SaMD)" under the Pharmaceuticals and Medical Devices Act. X-ray image analysis, ECG analysis, endoscopy support apply. A scheme for continuous-learning types to submit a "change-management plan" and receive sequential approval is being prepared.
Financial AI
The FSA's materials guide. Explainability and prevention of discriminatory treatment are emphasized, and judging credit by black-box AI alone is effectively not allowed.
Autonomous Driving
Amendments to the Road Traffic Act and Road Transport Vehicle Act started Level 4 (full automation under specific conditions) operation. The design assigns accident liability to the operation manager.
Personal Information Protection
The Act (APPI) requires "proper acquisition" and "specifying purpose of use" for AI training data. Acquiring sensitive personal info (medical, thought, race, etc.) without consent is in principle prohibited.
2025 Trends
- AI Promotion Bill: debated in the 2025 ordinary Diet session. Mainly clarifying operator responsibility and promotion. Not as binding as the EU AI Act
- Elections and deepfake measures: spurred by the 2024 lower-house election, debate on AI-generated-content display obligations during election periods
- Creator protection: legislative debate on style protection and training opt-out
What to Do in Practice
- Set internal policy along the AI Business Operator Guidelines
- For regulated sectors (medical, finance, transport) organize sector-specific requirements with legal
- Document a copyright policy for training data/prompts
- Parallel response with the EU AI Act (if you provide extraterritorially)
Summary
Japan's AI regulation's pillar for now is "guideline-centric + sector-specific application of existing laws." Slow legislation doesn't mean no response is needed; the pitfall is that existing per-sector regulation naturally applies to AI too. Operators in medical, finance, hiring, and autonomous driving especially need care.
