The "Fire and Ice" Divide in AI Copyright Between China and the U.S.: What Should You Do with Your AI-Generated Content When Going Global?
On March 2, 2026, the U.S. Supreme Court declined to hear the appeal in the *Thaler* case—meaning that, in the United States, works generated entirely by AI are not eligible for copyright protection. On the same day, on the other side of the globe, Chinese courts had already held in a series of rulings that AI-generated images can constitute works protected under copyright law.
The same technology, two completely opposite answers. This is not merely a talking point for legal academics—it is a real minefield that every Chinese company using AI to generate product images, marketing materials, software copyright code, or even brand IP characters is stepping into.
A designer used Midjourney to create a "transparent butterfly chair." When a peer copied it for mass production, she sued for RMB 200,000 in damages. The court dismissed the claim entirely. The reason was not that her drawing was insufficiently good, but rather—she had not saved screenshots of the creative process and could not prove the original intellectual input of "from where to where." Without original records, an AI-generated work is tantamount to having no copyright. This case (the *Fantasy Wings* case) became final in the second instance in April 2025, making it the first effective judgment in China holding that AI-generated content "does not constitute a work."
At the same time, in another case—the *Spring Breeze* case, China's first copyright case involving AI-generated text-to-image works—the plaintiff preserved 24 positive prompts, 120 negative prompts, and records of 4 parameter adjustments. The court found that this "reflected personalized expression and intellectual input" and awarded RMB 500 in damages.
Both involved AI-generated images—one won RMB 500, the other lost RMB 200,000. The only difference: did you leave a paper trail?
---
I. The U.S. Logic: Pure AI Output Is "Unprotected," Human Control Is Your Shield
The U.S. Copyright Office's *Copyright and Artificial Intelligence* report, Part 2, issued in January 2025, set forth eight conclusions. The core message boils down to one sentence: existing law is sufficient—no AI-specific legislation is needed; the existing "human authorship" standard can resolve the issue.
Specifically:
- Purely AI-generated output: not protected. If you type "draw a girl" into Midjourney, that image has no copyright in the United States, and anyone can take it and use it commercially for free.
- AI-assisted tools: do not affect protection. If you use Photoshop's AI features to edit an image, the copyright still belongs to you—because you, as a "human," are the one creating.
- Prompts themselves do not constitute "creative control": The U.S. Copyright Office's logic is that prompts are more like "suggestions" than "commands"—you cannot predict what the AI will ultimately output, so you cannot "control" it.
But there is one narrow path: proof of "human creative control."
The U.S. Copyright Office and courts recognize the following pathways in practice:
- You first draw a hand-drawn sketch, then have AI generate based on it → the Copyright Office may recognize the "perceptible human-authored expression" portion
- You make substantial modifications to the AI-generated original using Photoshop → in the *A Single Piece of American Cheese* case, Keirsey made more than 35 inpaint revisions to the AI-generated original, and the Copyright Office found this demonstrated human creative control and granted copyright protection
For Chinese companies going overseas, this means: If you list AI-generated product images on Amazon, use AI-generated brand visuals on your independent site, or submit software copyright materials to the U.S. Copyright Office—without a chain of evidence showing "human creative control," these materials are in the public domain in the U.S. market. Anyone can copy them, reuse them, or even turn around and sue you for infringement.
⚠️ There is also a hidden risk: pure AI-generated output not only lacks copyright protection in the U.S., but may also give rise to reverse infringement liability. Getty Images is currently suing Stability AI, alleging that approximately 12 million licensed images were used to train Stable Diffusion. If it is later discovered that a particular AI tool's training data included copyrighted works of others, your commercial content generated with that tool could expose you to "contributory infringement" liability in the United States.
---
II. China's Logic: Case-by-Case Determination, but "Paper Trail" Is the Decisive Factor
China's approach is less about "protecting AI creation" and more about "protecting human choices."
The current core judgment landscape:
| Case | Outcome | Key Factors |
|------|---------|-------------|
| *Spring Breeze* case (2023) | ✅ Constitutes a work; RMB 500 awarded | 24 positive prompts + 120 negative prompts + 4 parameter adjustments, with complete records |
| *Companion Heart* case (2024) | ✅ Constitutes a work; RMB 10,000 awarded | Multiple rounds of iteration + manual adjustments + prior registration with the National Copyright Administration |
| *Fantasy Wings* case (2025) | ❌ Does not constitute a work; claim dismissed in full | Only simple prompts, unable to provide original records of the creative process |
The holding of the *Fantasy Wings* case deserves to be read word by word by every enterprise investing in AI-driven creation: "Relative to the generated content, prompts are merely 'ideas' rather than 'expression'; simple prompts themselves do not constitute a work." The court's logic: if you cannot reproduce the process "from conception to selection" through original records, you do not qualify as the "author."
Professor Wang Qian of East China University of Political Science and Law offered a more vivid analogy: AI creation is like "shooting the arrow first, then drawing the target"—the arrow was shot by the AI, and you merely walked up and drew a circle where the arrow landed, saying, "This is where I was aiming." Without records of the creative process, how can you prove that the circle was not drawn after the fact?
For Chinese companies, this means: In litigating AI copyright cases in China, you are not proving "how smart the AI is," but rather "how much you invested and whether it can be seen."
---
III. Practical Checklist for Enterprises: 6 Actions for Both Overseas and Domestic Markets
Enough theory. Here is an action checklist you can execute starting today.
[U.S. Market] For the U.S. market (e-commerce, SaaS, content going overseas)
Action 1: Use "expressive input" at the input stage
Do not just type prompts. Draw sketches, write scripts, create wireframes—anything "completed by a human before opening the AI" constitutes "perceptible human-authored expression." The U.S. Copyright Office has the highest recognition for the "sketch → AI generation → refinement" pathway.
Action 2: Mandate secondary processing at the output stage
Set an internal corporate standard (e.g., "AI-generated content must undergo at least X rounds of human modification before delivery to clients/launch/registration") to ensure every finished product bears a human "fingerprint." The "35 inpaint revisions" is a real precedent and can serve as an internal reference standard.
Action 3: Proactively disclose AI materials at registration
When registering copyright in the U.S., honestly disclose the AI-involved portions and make a "no copyright claim" statement—there are already 1,000+ successful precedents, and registration itself is not rejected due to AI involvement.
[China Market] For the China market (work registration, litigation enforcement)
Action 4: Establish an AI creation paper-trail SOP
Every AI generation must preserve: full prompt text, parameter settings, iteration versions, generation timestamps, and final selection records. Use folders or internal enterprise tools for version management. The lesson from the *Fantasy Wings* case: no original records = no copyright.
Action 5: Prioritize registration
The common thread in the *Spring Breeze* and *Companion Heart* cases—both plaintiffs had registered their works with the National Copyright Administration. In 2024, the total number of copyright registrations nationwide reached 10.63 million (up 19.13% year-over-year). The registration market is enormous, and demand for AI content registration is exploding. Register first, then use—this is the lowest-cost insurance for enforcement.
Action 6: Align labeling compliance
The *Measures for the Labeling of AI-Generated Synthetic Content*, effective September 1, 2025, together with the mandatory national standard GB 45438-2025, require explicit and implicit labeling of AI-generated content. This is not merely a compliance obligation—the metadata embedded in implicit labels (generation attributes, service provider codes) is itself part of the chain of evidence for creative paper trails.
---
Conclusion
The divergent paths of the U.S. and China on AI copyright appear, on the surface, to be "one says no, the other says yes." But a closer look at the underlying logic of the precedents reveals that both countries arrive at the same destination: both require human contribution to be "visible and provable."
The U.S. looks at "control"—to what degree did you control the creative process?
China looks at "input"—can your intellectual investment be reproduced at the evidentiary level?
For enterprises, this is not merely a legal compliance issue—it is whether your AI-generated content assets can become property protected by law. Creation without a paper trail is tantamount to building a free asset library for others.
---
Next Article Preview: We will dissect "Copyright Traps in AI Tool User Agreements"—when you generate content using free AI tools, is the tool provider quietly taking away your copyright?
Follow 「纳杰觅理」 to never miss a practical article.
---
Landline: 010-65150974 | Mobile: 15321374076
Email: [email protected] | WeChat: najieip
He Zigang | Intellectual Property Attorney | Beijing Najie Intellectual Property
*This article represents only the author's personal views and does not constitute legal advice. For case-specific analysis, please feel free to contact us.*