Organize the Issues in 3 Stages
The AI-and-copyright debate is easily confused, so it's clearer to split it into (1) the training-data stage, (2) the output stage, (3) the publication/use stage.
(1) Training-Data Stage
Japan
Copyright Act Art. 30-4 allows use without rightsholders' consent "when provided for information analysis." One of the most permissive countries for AI training. But the enjoyment purpose (purpose of appreciation) is out of scope, and the training act and use at output are judged separately.
US
Fair-use principle. Under individual judgment in NYT v. OpenAI, etc. If verbatim reproduction occurs, possible infringement finding.
Confirmed in 2026: case law is taking shape. The authors' class action against Anthropic settled for about US$1.5B (≈US$3,000 per work) after a ruling that training on books is fair use but storing pirated copies is not. Thomson Reuters v. Ross was found not fair use and is on appeal; Disney/Universal v. Midjourney is ongoing. Output generated purely by AI is not registrable for copyright in the US (upheld by federal and circuit courts; the Supreme Court denied cert in 2026). In the EU, AI Act high-risk enforcement begins in August 2026, and GPAI providers must publish training-data summaries and a copyright-compliance policy.
EU
The 2019 DSM Directive has a "text and data mining" exception, but rightsholders can opt out. The AI Act's GPAI obligations require publishing training-data overviews and copyright compliance.
(2) Output Stage
Similarity Judgment
If AI-generated images/text/music have expressive commonality with a specific existing work, it's the same as normal copyright-infringement judgment. Whether it was in the training data is irrelevant; if the output result is similar, it's infringement.
Style Protection
In many jurisdictions, style is out of copyright protection. Generating with a "[author name] style" prompt itself is usually legal. But separate issues may arise under unfair-competition law, personality rights, publicity rights.
(3) Publication/Use Stage
Copyright Attribution of AI Output
- US: Thaler v. Perlmutter (2023): "pure AI output without human involvement has no copyright"
- Japan: under copyright law, human creative intent is needed. Direction to recognize creativity if prompt crafting + selection + editing is acknowledged
- EU: human creative contribution needed
Disclosure Obligations
- EU AI Act: limited-risk AI (chatbots, deepfakes) must display as AI-generated
- YouTube: set an AI-use flag
- Music distribution: debate on "AI-generated song" display
- Advertising: under the labeling law, displaying generated testimonials needs care
Measures from a Creator's View
Don't Want Your Work Trained On
- NoAI tag, exclude training via robots.txt
- Anti-training processing of images with Glaze / Nightshade
- Continuously watch JASRAC, ACA debates
- Individual opt-out (apply to OpenAI, Midjourney)
Want to Use AI Commercially
- Commercially licensed AI (Adobe Firefly, Getty AI, Shutterstock AI)
- Open-weight + license confirmation (FLUX, SD license conditions)
- Similarity check of output (reverse image search)
- State an AI-use policy in client contracts
Practical Checklist
- Confirm the training-data provenance of the AI model used
- Confirm the commercial-license scope (publishing, broadcast, merchandising)
- After output, check similarity with existing works (Google Lens for images, search for text)
- AI-use clause in client contracts
- Properly display "AI use" where platforms require it
- Save logs (prompt, generation date, model version)
Ongoing Debates
- NYT v. OpenAI ruling
- Final resolution of Getty v. Stability
- Japan's creator-protection legislation
- EU AI Act GPAI training-data disclosure
- Collective bargaining in music/voice-actor fields
Summary
AI and copyright have different issues at the 3 stages of "training," "output," "publication." In creator practice, pinning down 3 points—(a) choose commercially licensed AI, (b) check output similarity, (c) include an AI clause in client contracts—greatly lowers current risk.
