If you build an AI image-generation product, here's a question worth sitting with: when a user trains a model on someone else's copyrighted character and publishes it, who pays — the user or your platform?
China's first AI-model copyright case (the "Medusa" LoRA case) just gave an answer on appeal. The Shanghai IP Court upheld the original ruling.
The facts: a user took a copyrighted anime character, trained two LoRA models on it, and published them on a domestic AI image platform. The court found the user liable for infringing the reproduction right and the right of communication through information networks — and ordered damages of RMB 50,000.
But the platform walked free. No joint liability. Why?
Because it had done three concrete things:
— an in-page reporting function on every model page;
— a review mechanism before any LoRA model went live;
— prompt takedown, blocking and delinking once notified.
The court held it had met its duty of care.
I come from a computer science background, so I read these cases through the product mechanics. And the real signal here is the opposite of what most people assume. This ruling does NOT say "platforms are always liable for AI output." It applies a duty-of-care standard — and this platform was cleared precisely because it built those safeguards.
So "technology neutrality" is not an automatic shield. But meeting a duty of care that matches your capability and role can clear you. The platforms genuinely at risk are the ones that built none of this.
Three things to check in your own product now:
- Provenance of training data — keep a clear ledger (owned / licensed / publicly available).
- Output-side filtering + a complaint channel + the ability to take content down fast.
- Content labeling — China's mandatory standard GB 45438-2025 took effect on 1 September 2025; explicit and implicit metadata labels are both required.
Going global? Copyright rules on training data and AI output vary widely by jurisdiction — some stricter than China's. Confirm with local counsel.
One thing to do today: pull up your training-data source list and output-review flow, and see which of the three is missing.
Send me your specifics and I'll walk through it with you.
———
Mike Wu, China-licensed Attorney · Cross-border practice
4 languages (CN/EN/ES/JA)
Site: wzklaw.com/contact.php?source=linkedin · Book consultation: wzklaw.com/contact.php?source=linkedin
#AICompliance #AIGC #CrossBorderLaw #TechLaw #Copyright