- Copyright Ownership: Depends on Human Input
China (Key Cases):
Copyright recognized if output reflects significant user originality (e.g., detailed prompt engineering, parameter tuning, manual editing – Beijing Internet Court & Jiangsu Changshu Court cases).
No copyright if input is simple/minimal and lacks proof of creative process (Jiangsu Zhangjiagang Court case).
US Copyright Office: Pure AI output (simple prompts) is not protected. Protection may apply to human-modified elements.
EU: Focuses on protecting user creative input. Requires AI transparency on training data sources.
Platforms (MidJourney, ERNIE-ViLG, etc.): User retains copyright but grants the platform a broad, perpetual, sublicensable license.
Stable Diffusion (Local): User owns output copyright (“Licensor claims no rights in the Output”). Commercial use requires license for paid tiers if revenue >$1M.
- Infringement Risks: Sources
Training Data: Using copyrighted material without permission risks infringement claims (e.g., Udio/Suno lawsuit).
Output Similarity: Outputs substantially similar to existing works risk infringement claims (though form change, e.g., 2D -> 3D, may not infringe).
Platform Terms: Broad licenses granted to platforms enable third-party use.
Removing Watermarks: Infringes creator’s attribution right (China Stable Diffusion case – 500 RMB fine).
- Key Trends:
China: Case-by-case assessment based on originality proof. No specific AI law yet.
US: COPIED Act (proposed) mandates AI content labeling, bans watermark removal.
EU: AI Act mandates model transparency; allows opting out of AI training.
- User Recommendations:
Prove Originality: Document detailed prompts, iterations, edits.
Avoid Simple Prompts: Use specific, descriptive language.
Choose Tools Wisely: Prefer locally deployed models (e.g., Stable Diffusion) for full control.
Check for Similarity: Use reverse image search before commercial use.
Label AI Content: Disclose AI tool use (e.g., “Generated by MidJourney”).
Avoid 3rd-Party IP: Do not include trademarks/copyrighted characters.
Summary: AI image copyright depends on proven human creative contribution. Risks stem from training data, output similarity, and platform terms. Mitigate risk by documenting input, choosing tools carefully, checking outputs, and labeling content. Global regulations are evolving.