Jinshan District People’s Court judgment on the Battle Through the Heavens / Medusa LoRA case is a milestone in how Chinese courts will handle copyright disputes around AI large models.

When AI Meets IP: China’s Medusa Verdict Sets the Rules for the Next Generation of Models

by ChatGPT-5.1

The Jinshan District People’s Court judgment on the Battle Through the Heavens / Medusa LoRA case is a milestone in how Chinese courts will handle copyright disputes around AI large models. It clarifies who is on the hook (user vs. platform), what counts as infringement in the training pipeline, and when AI outputs qualify as “works” under copyright law.

Below I’ll first explain the case and how it is corroborated by other sources, then unpack what it means for AI makers and rights owners in China, and finally assess the pros and cons and likely impact on future litigation.

1. What happened in the Medusa LoRA case?

Facts and parties

The plaintiff is the copyright owner of the character “Medusa” in the hit Chinese IP Battle Through the Heavens (斗破苍穹). The defendants were:

• Li, an individual user of an AI image platform, and

• A leading domestic AI image-generation platform that aggregates many LoRA models and provides online image generation services.

Li:

1. Extracted 20+ images of Medusa from the plaintiff’s animated series.

2. Assembled these into an image pack.

3. Used the platform’s “Train LoRA” function to fine-tune a Medusa LoRA model.

4. Published the Medusa LoRA models on the platform, making them available to other users.

5. Other users could then generate images identical or substantially similar to the plaintiff’s Medusa character simply by changing prompt words.

The plaintiff alleged:

• Copyright infringement (reproduction, adaptation, information network dissemination rights),

• Unfair competition (claiming “Medusa” had become a commodity name with market influence), and

• Platform liability, on the ground that the platform hosted many infringing LoRA models (including in its “Anime Zone”) and failed to fulfill its platform duties.

The court’s key findings

The Jinshan District People’s Court made four crucial determinations.

1. “Medusa” is not a protected commodity name with influence.
   The court held that “Medusa” is widely known as the Gorgon from Greek mythology and is used in many comics, games, songs and plant varieties. It therefore cannot be monopolized as a special commodity name for unfair-competition protection.

2. Li infringed copyright at the input and training stages.
   The court found that Li, for commercial purposes, reproduced the original expression of the Medusa character when:

   • Extracting image materials,

   • Training the LoRA model,

   • Releasing the LoRA model, and

   • Making image packs and short videos available to the public.

   These acts infringed the plaintiff’s reproduction right and information network dissemination right under Article 10 of the PRC Copyright Law.

3. AI-generated images here do not qualify as “works”; thus no adaptation right infringement.
   The court emphasized that, absent proof of substantial human intellectual input, images directly generated by generative AI are not “artistic works” under Article 3 of the Copyright Law. In this case:

   • The Medusa LoRA model and AI-generated Medusa images did not show substantial human creativity from Li.

   • They therefore did not constitute “works”.

   • As a result, Li did not infringe the adaptation right, because there was no new “work of original creation” to speak of.

4. The platform is a neutral online service provider and benefits from safe harbor.
   The court held that the platform:

   • Did not directly participate in material extraction or LoRA training;

   • Provided LoRA technology as a neutral tool that promotes generative AI development;

   • Had complaint and review mechanisms; and

   • Upon receiving notice and the complaint, promptly removed all Medusa LoRA models, updated keyword filters, and even notified overseas AI platforms.

   Under Articles 1195 and 1197 of the PRC Civil Code (notice-and-takedown and “knew or should have known” standards), the court found the platform:

   • Had no subjective fault, and

   • Had objectively taken “necessary measures” and “forwarded notifications”.

   Therefore, it did not bear joint liability.

Remedies

The first-instance judgment ordered:

• Li must stop infringing the plaintiff’s reproduction and information network dissemination rights;

• Li must pay RMB 30,000 in economic losses and RMB 20,000 in reasonable expenses (total RMB 50,000);

• All other claims (including against the platform) were dismissed.

Multiple Chinese-language reports (Sina Finance, 21st Century Business Herald, Labor Daily, Nanfang Metropolis Daily, Shanghai High Court WeChat summaries and IP-economy sites) repeat these same facts, amounts, and reasoning.

The OECD AI Incident Database also records the case, describing it as Shanghai’s first AI large-model copyright judgment where the user was held liable and the platform cleared due to compliance efforts.

2. How this fits into China’s emerging AI copyright jurisprudence

This decision is not an outlier; it sits within a rapidly expanding cluster of Chinese cases dealing with AI and copyright:

• Hangzhou “Ultraman” case (2024–2025): The Hangzhou Internet Court (upheld on appeal) held an AI platform contributorily liable when it allowed users to create and share models generating “Ultraman” variants and failed to implement necessary preventive measures.

• Beijing “Chunfeng” and Guangzhou “Ultraman” earlier image cases: focused on whether AI-generated pictures could themselves enjoy copyright, and on ownership of such works.

• Shanghai “prompt copyright” case (Huangpu District Court, 2025): just days after the Medusa ruling, another Shanghai court held that simple “listed” prompts lack sufficient creativity and thus do not qualify as protected works—again insisting on human intellectual input as the touchstone.

Commentary from MLex and Chinese IP law firms notes that:

• Courts are beginning to diverge on platform fault (Shanghai vs. Hangzhou),

• But they converge on a baseline: platforms must have notice-and-takedown mechanisms and act once alerted; users who feed copyrighted material into models can be directly liable.

3. What this means for AI makers in China

a) A clearer “safe harbor” compliance blueprint

The Medusa judgment essentially sketches a compliance checklist for Chinese AI platforms wanting safe-harbor protection under the Civil Code:

1. Neutral technology posture

   • Provide general-purpose tools (e.g., LoRA fine-tuning) rather than directly pre-packaged infringing content.

   • Avoid actively curating or promoting infringing LoRAs or branded IP packs in marketing.

2. Robust content review & complaint systems

   • Have visible complaint/reporting channels.

   • Run automated and manual review for obviously infringing IP (especially famous characters and series).

3. Prompt, well-documented takedowns

   • Remove or disable access to infringing models swiftly upon notice.

   • Update keyword filters or other preventive mechanisms after a complaint.

   • Where relevant, notify linked or overseas platforms that might host the same content.

4. Legal framing as an “online service provider”

   • Align terms of service and product design with the Civil Code’s notion of a “network service provider”;

   • Emphasize users’ responsibility for uploaded content and training materials.

If a platform can show it did all this, Jinshan suggests courts may treat it as a neutral intermediary rather than a co-infringer—even in high-profile IP disputes.

By contrast, the Hangzhou Ultraman case illustrates what not to do: failing to implement preventive measures and allowing obviously infringing models to circulate can trigger contributory liability.

b) The focus shifts to input/training rather than only output

This judgment moves the focus from “Is the AI output infringing?” to “Was the training material used lawfully?”:

• Extracting copyrighted character images without permission,

• Using them to train a model,

• Publishing the trained model for others to generate similar images,

is itself enough to infringe reproduction and network dissemination rights.

For AI makers, this means that user-side fine-tuning is not a free-for-all. Platforms that offer finetuning tools must:

• Explicitly warn users about IP risks,

• Monitor and respond to abuse, and

• Consider technical safeguards (e.g., hash-matching known IP, blocking certain character names).

c) Uncertainty remains for foundation-model training

Notably, the Medusa case doesn’t squarely address whether training a base large model on copyrighted works (e.g., whole anime catalogues scraped at scale) is lawful. The infringing acts are framed around user extraction and re-use of specific character images for a LoRA.

So for AI makers training or distilling foundation models on large corpora, big questions remain:

• Is large-scale text/image scraping infringing “reproduction” or “network dissemination”?

• Are there implied exceptions or fair-use-like defenses (e.g., under China’s TDM-like provisions for data mining in certain contexts)?

Those debates are still moving toward the Supreme People’s Court and future guiding cases.

4. What this means for rights owners in China

a) A clear red line: user-level fine-tuning on protected IP is actionable

Rights owners now have a strong precedent that:

• Users who “feed” copyrighted character designs into LoRA or similar tools

• And then share the resulting model for others to use,

can be held liable for infringement, with a relatively straightforward access + substantial similarity analysis.

This is particularly important for:

• Anime, comics, games, and film/TV characters,

• Branded mascots and stylized logos,

• High-value entertainment IP that proliferates through fan communities and UGC.

b) Strategic implications for enforcement

The decision also tells rights owners how to structure enforcement:

1. Go after both user and platform — but build the record for platform fault.

   • Send detailed notices with evidence (screenshots, URLs, model IDs, prompts).

   • Document the platform’s response and timing.

   • Only if they ignore or delay can you credibly argue they “knew or should have known” and failed to take “necessary measures” under Civil Code Articles 1195–1197.

2. Use the case as leverage in platform negotiations.

   • Platforms will be keen to avoid repeat litigation; rights owners can use that to insist on:

     • proactive filters for specific IP,

     • whitelists/blacklists, and

     • possible licensing deals for official character models.

3. Recognize the limitation on adaptation claims.

   • Because AI outputs with low human input are not treated as “works” here, rights owners may need to anchor their claims in:

     • the underlying copyrighted source images, and

     • the training and dissemination of the model,
       rather than trying to claim adaptation rights in each AI-generated image.

5. Pros and cons of the verdict

Pros

1. Clear guidance on user liability and the “input side”.
   The judgment sends a strong signal that using copyrighted images to fine-tune LoRAs for commercial or public use is not a gray area. Users cannot hide behind the “it’s just training data” argument.

2. A workable safe harbor for tech innovation.
   The decision gives AI platforms a concrete compliance pathway: neutral tools + notice-and-takedown + updated filters can shield them from liability, which should reduce chilling effects on legitimate AI R&D.

3. Consistent insistence on “human intellectual input”.
   Together with the prompt-copyright decision, the court is drawing a principled line:

   • No protection for bare prompts or low-creativity instructions,

   • No automatic protection for raw AI outputs without substantial human creativity.

   This avoids an over-expansion of copyright that could stifle sharing, remix culture, and AI-assisted creativity.

4. Alignment with general intermediary liability rules.
   The reasoning closely tracks the Civil Code’s generic rules for network service providers, making it easier for platforms (and their lawyers) to integrate AI-specific compliance into existing frameworks.

Cons

1. Potentially too lenient on platforms.
   Critics may argue that merely reacting after litigation is filed (rather than upon first user complaint) should not suffice. The bar for “no subjective fault” may be seen as generous, especially for a “leading” AI platform with an “anime zone” clearly attractive to fans and potential infringers.

2. Low damages relative to IP value.
   RMB 50,000 is modest given the commercial value of Battle Through the Heavens. This may not strongly deter similar user behavior, especially in large fan communities.

3. Unresolved questions on training of foundation models.
   By focusing on a user-side LoRA, the court sidestepped the core controversy: Can AI companies train their base models on copyrighted works without permission? That leaves a huge gap for future litigation and regulatory guidance.

4. Fragmentation risk across courts.
   The contrast with Hangzhou’s Ultraman decision—where the platform was found liable—can create uncertainty and forum-shopping. MLex and Chinese commentators note that local courts are still experimenting with how far to push platform responsibility.

5. Ambiguity for AI-assisted human creators.
   If AI outputs are generally treated as non-works unless you prove substantial human creativity, individual illustrators and designers who heavily rely on AI tools may find it harder to enforce their own rights. The doctrine is sensible, but its application will be contested.

6. Likely impact on future litigation

Looking ahead, this case is likely to influence Chinese AI copyright disputes in several ways:

1. More cases targeting fine-tuned LoRA and UGC models.
   Rights owners now know that going after user-trained character LoRAs is viable. Expect more suits over anime, game and film characters trained into LoRAs and shared in communities.

2. Evidence-heavy cases on platform “knowledge”.
   Future plaintiffs will structure their enforcement around creating a solid trail of:

   • Notices sent,

   • Platform inaction or delayed action, and

   • Continued availability of infringing models.

   Only with that record can they overcome the safe-harbor shield and argue joint liability.

3. Courts will keep refining the “human input” test.
   We’re likely to see more nuanced decisions on when AI-assisted works do qualify as “works”—for example, where users heavily edit, composite, or meaningfully direct generation. This will be central to creators who use AI as a tool, rather than a substitute.

4. Pressure on regulators to clarify training data rules.
   As these cases accumulate, the Supreme People’s Court, the National Copyright Administration and other regulators may:

   • Issue judicial interpretations on AI and copyright,

   • Clarify how existing exceptions apply to training and TDM-like uses,

   • Possibly designate “high-risk” practices for AI providers to avoid.

5. Global implications for foreign AI companies and foreign IP owners.

   • For foreign AI providers offering services into China or partnering with Chinese platforms, Jinshan indicates the minimum compliance they will be expected to meet.

   • For foreign rights owners, it shows that Chinese courts are willing to protect entertainment IP against AI-related misuse—while still giving domestic AI firms a robust safe harbor if they behave “responsibly”.

On balance, the verdict is a cautiously platform-friendly, user-targeted approach that pushes responsibility down to the individual who feeds and shares infringing training data, while giving platforms a clear but demanding compliance path. The next wave of cases—especially around base-model training and repeat notices—will determine whether that balance holds or shifts toward stricter platform accountability.

Sources:

1. Fei Dang, “Shanghai Court Issues First Instance Judgment on AI Model Infringement,” MMLC Group (English commentary, PDF you provided):
   https://mmlcgroup.com/china-ai-copyright-1711/ MMLC Group

2. Shanghai High Court / IP-economy report on Shanghai’s first AI large-model copyright case (Chinese):
   https://www.ipeconomy.cn/mobile/dongtai/9970.html ipeconomy.cn

3. 21st Century Business Herald report “《斗破苍穹》被AI‘抄袭’：用户判赔5万，大模型公司免责”:
   https://www.stcn.com/article/detail/3478936.html STCN

4. Nanfang Metropolis Daily report “用户投喂AI大模型生成《斗破苍穹》角色图片，被判赔5万元”:
   https://www.nfnews.com/content/WyjqNjnao0.html nfnews.com

5. Labor Daily / Shanghai report “《斗破苍穹》美杜莎形象被抄袭，上海首例人工智能大模型著作权侵权案宣判”:
   https://www.51ldb.com/shsldb/sz/content/019a4cdbc832c0010000dd82f3b39891.html 51ldb.com

6. Inf.news English summary of the Medusa case:
   https://inf.news/en/comics/e01902b7ebfb3d5a7d4a93ed323dbacc.html inf.news

7. Ronze Law “Data Law Weekly” overview including the Medusa case (English/Chinese):
   https://www.ronzelaw.com/news_details/128.html Ronze Law

8. MLex, “China’s AI copyright rulings diverge on platform fault, align on liability baseline” (Shanghai vs Hangzhou comparison):
   https://www.mlex.com/mlex/articles/2408166/china-s-ai-copyright-rulings-diverge-on-platform-fault-align-on-liability-baseline MLex

9. Hogan Lovells, “When AI and Copyright clash: Chinese courts find AI platform liable for contributory copyright infringement” (Hangzhou Ultraman case):
   https://www.hoganlovells.com/en/publications/when-ai-and-copyright-clash-chinese-courts-find-ai-platform-liable-for-contributory-copyrightwww.hoganlovells.com

10. Zhichanli (IP media) article discussing the Medusa case and the Shanghai prompt-copyright decision: https://www.zhichanli.com/p/962221931

11. MMLC Group, “Shanghai Court Rules on the First AI Prompt Related Copyright Case” (prompt case context):
    https://mmlcgroup.com/china-copyright-ai-prompts/ MMLC Group

12. OECD AI Incident Database entry “Shanghai Court Rules User Liable for AI-Generated Image Infringement”:
    https://oecd.ai/en/incidents/2025-11-03-677b