A. Rulings Recognizing Copyright in AI-Generated Works

(1) Beijing Internet Court's 2024 'AI Text-to-Image' Ruling: A World-First Recognition

One night in February 2023, a man named Li sat staring at his laptop screen in a Beijing apartment.

He was a lawyer, but that night he was an artist. On the screen was Stable Diffusion, an American AI program, and he was carefully typing words with the precision of a safecracker dialing in a combination. 'Twilight,' 'girl,' 'black hair,' 'soft light.' He wasn't satisfied. He changed the angle of the light, adjusted the artistic style, and fine-tuned the weights. After dozens of attempts, an image of a beautiful Asian woman silhouetted against a sunset appeared on screen.

He posted the image to Xiaohongshu, China's equivalent of Instagram. He added hashtags: #AI, #AI插画 (AI illustration), #AI绘画 (AI art). A few days later, a blogger named Liu discovered the image.

Liu must have liked the picture. She cropped out the watermark bearing Li's signature and used it without permission as an illustration for her poetry.

This seemingly trivial incident sent shockwaves through the global legal community. Li sued Liu for copyright infringement. The amount in dispute was a mere 500 yuan, less than 100,000 Korean won. But the eyes watching this trial burned hotter than those observing lawsuits worth hundreds of billions of dollars. The central question was straightforward: Can an AI-generated image be protected by copyright?

Understanding what copyright means comes first. Copyright is something like a receipt attached to a creative work. It lets society verify who made something, and based on that verification, it creates a system for blocking or authorizing copying and distribution. The question was whether something made by AI qualifies as a 'creative work.' On November 27, 2023, Judge Zhu Ge of the Beijing Internet Court read the ruling. It was the world's first judgment recognizing copyright in a generative AI image.

The court's logic was surprisingly pragmatic.

The ruling began: "Throughout technological development, humans have progressively delegated complex and tedious physical labor or simple intellectual activities to machines. AI models are likewise tools that assist human creation." The court defined AI not as something to fear but as a sophisticated brush.

The judge posed a decisive question: "A photographer only presses the shutter, yet we recognize photographs as art. If an AI user combines hundreds of words and goes through dozens of revisions to obtain a desired image, can we say that required less effort than pressing a shutter?"

The court recognized Li's prompts and parameter settings not as mere mechanical operation but as 'intellectual investment.'

Li entered specific prompts such as 'Japan idol,' 'cool pose,' 'viewing at camera,' and 'film grain.' When the initial generated image didn't satisfy him, he revised the prompts, adjusted parameters, and repeated the process of selecting among multiple outputs. The court found that this entire process reflected "human selection and arrangement."

The ruling rested on two key points.

First, the AI model itself cannot be an author. Article 11 of China's Copyright Law specifies that an author must be a 'natural person or legal person,' and AI is neither.

Second, a human who uses AI as a tool can be an author. Li proactively controlled the AI to express his own thoughts and emotions. The resulting work was therefore Li's copyrighted creation. Liu was ordered to pay 500 yuan in damages and 50 yuan in litigation costs. The numbers were small. But this ruling sent a powerful signal to every AI creator in China: what you make with AI can become your property.

After the ruling, the judge said in an interview: "Copyright law should encourage creation and innovation. That includes creation using the latest tools. It is necessary to adapt the traditional copyright framework to evolving AI technology." This was a declaration of judicial activism holding that the law must support technological innovation.

Another ruling followed in 2024. An AIGC designer in Shanghai used Midjourney to create an image titled 'Companion Heart' for Valentine's Day. He was a professional creator who had collaborated with Mercedes-Benz, Estee Lauder, and McDonald's. A competitor used the image without permission, and the court once again recognized copyright in AI-generated work. The court added one condition, however: to claim copyright, "the claimant must demonstrate the creative effort exercised during the creation process."

(2) Establishing the Standard of Intellectual Investment (Prompts and Parameter Adjustments)

To understand the concept of intellectual investment, consider a cooking analogy.

If you just pour water and call it done, like making instant noodles, it's hard to claim creative initiative. But if you select ingredients, control the heat, adjust seasoning, and plate the dish, then 'your flavor' remains.

The 'intellectual investment' standard set by the Beijing Internet Court concerns exactly this difference.

According to the ruling, Li entered approximately 150 complex prompts to obtain his desired image.

These included positive prompts specifying lighting, composition, the subject's pose, skin texture, and artistic style, as well as negative prompts excluding unwanted elements. He did not stop after a single attempt. He fine-tuned parameters dozens of times, changed seed values, and refined the output step by step.

The court analyzed this process in three stages.

The first is prompt design. When a user enters text that specifically describes composition, light source, color palette, and the subject's pose, it corresponds to an author's conceptualization process. Typing the single phrase 'Japan idol' makes the AI generate an image on its own. But Li didn't stop there. He added specific instructions like 'cool pose,' 'viewing at camera,' and 'film grain.' This is no different from a painter sketching a concept in their mind.

The second is parameter adjustment. Adjusting iteration steps, weights, and seed values to refine the output was regarded as creative effort similar to a painter correcting brushstrokes. AI tools like Stable Diffusion have dozens of adjustable variables. How you set these variables changes the output entirely.

The third is selection and revision. Choosing from among the AI's many outputs the one that meets one's own aesthetic standards, then correcting deficiencies through additional commands or editing tools, was also recognized as part of the creative process. The establishment of this standard provided AI users with a guideline for becoming copyright holders. The moment you throw in a prompt and say 'the AI did it,' the foundation of your rights begins to crumble. Conversely, the more a user proactively controls the AI and reflects their own individuality, the more likely the output is to receive legal protection.

In September 2025, the Beijing Internet Court made this standard even clearer. A content creator named Zhou sued a company that used his AI-generated images without authorization.

The court ruled against Zhou. The reason was evidence. Zhou failed to submit records of the actual generation process from the AI software. Instead, he submitted images recreated after the fact using the same AI software during the litigation. The court treated this as 'post-hoc simulation' and found it insufficient to prove the original creative process.

This ruling left an important lesson.

"When claiming copyright in AI-generated work, the author bears the obligation to explain their creative thinking, the content of the commands they entered, and the process of selecting and modifying the generated content, and to submit relevant evidence." If you want to claim rights, you must keep records.

Prompt logs, revision histories, the progression from sketch to finished product. All of these become receipts that later allow you to argue in court: 'This is not something a machine spit out; this is something I made.'

(3) Comparison with the U.S. Thaler Decision: Differing Interpretations of Human-Centrism

Around the same time China's ruling was issued, another case was unfolding on the other side of the globe.

There was a computer scientist named Stephen Thaler. He developed an AI system called the 'Creativity Machine.' This AI produced an image entirely on its own. Thaler titled it 'A Recent Entrance to Paradise.' He filed an application with the U.S. Copyright Office to register the copyright.

In the 'author' field of the application, he wrote 'Creativity Machine.'

The Copyright Office refused. The reason: 'It was not made by a human.'

Thaler appealed. He lost in district court. On March 18, 2025, the D.C. Circuit issued its final ruling. It was a unanimous dismissal.

Judge Patricia Millett wrote the opinion: "The author is at the center of copyright law. And traditional tools of legal interpretation show that 'author' refers exclusively to humans."

The court offered several grounds. Copyright is a property right that vests in the author upon creation. Machines cannot legally own property. Therefore machines cannot be authors. The court also noted that various provisions of copyright law presuppose concepts such as the author's lifespan, inheritance, and nationality. All of these are concepts that apply only to humans.

Thaler raised a counterargument. By dictionary definition, an 'author' is 'one who creates or brings something into existence.' Nothing in this definition limits it to humans. The court dismissed this: "Statutory interpretation requires more than finding one favorable dictionary definition. The central task is to determine how Congress used that word in the statute."

Thaler's other argument was also rejected. He invoked the 'work-made-for-hire' doctrine, arguing that because the AI was his 'employee,' he should hold the copyright. The court responded: The work-made-for-hire provision states that the employer is 'considered the author.' The phrase 'considered the author' implicitly acknowledges that the employer is not the actual author. Every work must first be created by a human.

B. AI Model Infringement Litigation

(1) Guangzhou Internet Court 'Ultraman' Ruling: Direct Infringement Liability of Service Providers

One day in December 2023, a user on a Chinese AI website typed "generate Ultraman" into the input field. It was a brief action. Seconds later, an image appeared on screen. A massive humanoid figure in silver and red. The iconic silhouette that Japan's Tsuburaya Productions had built since 1966 was right there.

The legal team at Shanghai Xinchuanghua Cultural Development Co., Ltd. discovered this. They held the exclusive license for the Ultraman series in China.

Reproduction rights, adaptation rights, information network dissemination rights. Someone was creating Ultraman without permission. And it was a machine doing it.

On January 5, 2024, a lawsuit was filed. The defendant was a company operating an AI image generation website called Tab. The plaintiff sought 300,000 yuan (approximately $42,000). Cease the infringement, delete all Ultraman-related training data, and pay damages.

A month later, on February 8, the Guangzhou Internet Court issued its ruling. It was the world's first decision holding a generative AI service provider directly liable for copyright infringement.

The distinction between direct infringement and contributory infringement needs to be understood first. If you steal someone's property yourself, that is direct infringement. If you know a theft is happening and serve as a lookout, that is contributory infringement. In the case of AI service providers, the user was the one who actually typed "draw me an Ultraman." The service provider only supplied the tool. So who bears the responsibility?

The defendant argued: "We merely provided a technologically neutral tool. The actual act of generation was performed by the user." This is the so-called 'technology neutrality' defense. Just as a hammer is morally neutral, so too is an AI tool, the argument goes. The court rejected it.

The core logic of the ruling went like this.

First, the defendant did not merely provide technology. It operated a paid service and was deeply involved in content generation. Registered users consumed computing power they had purchased to use the AI drawing service. This was a for-profit content generation service.

Second, Ultraman is a work with high recognition in China. It is easily accessible on major streaming platforms like iQiyi. It would be difficult for the defendant to claim ignorance of the work's existence.

Third, the defendant failed to implement any filtering measures to block Ultraman-related keywords when they were entered.

The court found that the defendant's conduct infringed upon the plaintiff's reproduction rights and adaptation rights (the right to create derivative works). Some of the generated images were identical to or substantially similar to the original Ultraman works (reproduction right infringement). Others retained parts of the original expression while adding new features, making them unauthorized derivative works (adaptation right infringement).

The damages awarded were 10,000 yuan (approximately $1,400). Far short of the 300,000 yuan the plaintiff had sought. But the significance of this ruling was not in the amount.

The court specified three obligations that generative AI service providers must fulfill.

First, they must notify users through their terms of service not to infringe on others' copyrights. Second, they must establish a complaint mechanism so that rights holders can protect their copyrights. Third, when AI-generated content could cause confusion or misidentification, they must attach a conspicuous label.

At the end of the ruling, the court stressed balance. "Considering that the generative AI industry is still in its early stages of development, it is necessary to strike a balance between rights protection and industry development. While it would be inappropriate to impose excessive burdens on service providers, service providers must proactively fulfill reasonable and manageable duties of care." The message was clear: technology development should be encouraged, but technology that kicks over someone else's rice bowl will not be tolerated.

(2) Hangzhou Internet Court LoRA Model Decision: 30,000 Yuan in Damages

After the Guangzhou ruling, the same plaintiff filed another lawsuit against a different defendant. This time it was in the Hangzhou Internet Court.

The technology called LoRA needs to be explained first. LoRA stands for 'Low-Rank Adaptation.'

Think of it as a custom lens you attach to a base AI model. Just as the same camera produces certain textures and forms with striking clarity when you swap lenses, LoRA makes the model better at "hitting" a specific character or artistic style.

A user uploads a few Ultraman images, selects a base model, adjusts the parameters, and an Ultraman LoRA model is created. Once this model is applied, anyone can easily generate images resembling Ultraman.

The defendant in this case was an unnamed AI platform. This platform provided a service where users could create, share, and apply LoRA models. The platform's homepage featured Ultraman-related LoRA models and images in its 'Recommended' and 'IP Works' sections.

The defendant's defense was different. "Our platform does not provide training data. Users upload images and train the models themselves. We simply integrated open-source models uploaded by third parties. This falls under the platform's 'Safe Harbor' rules."

Safe Harbor is an important concept in internet law.

It is the principle that platforms can be exempt from liability for user content under certain conditions. If a user uploads a copyright-infringing video to YouTube, YouTube is not automatically liable. It just needs to take it down once notified.

But the court rejected this defense. On September 25, 2024, the Hangzhou Internet Court issued its first-instance ruling. The defendant bore contributory liability for infringement of the right of information network dissemination. Damages: 30,000 yuan (approximately $4,200).

The defendant appealed. On December 30, 2024, the Hangzhou Intermediate People's Court upheld the first-instance ruling. The court's reasoning centered on four factors.

First, the nature of the service and its revenue model. The defendant derived direct economic benefit from the creative services its platform provided. This was not mere neutral hosting.

Second, the prominence of the Ultraman brand. Ultraman is a widely recognized character in China. It would be difficult to believe that the platform operator was unaware of the character's existence and copyright protection.

Third, the obviousness of the infringement. Ultraman-related content was displayed in the 'Recommended' section of the platform's homepage. This meant the platform was not simply hosting content passively but was actively recommending it.

Fourth, the potential for proliferation. Given the technical characteristics of generative AI, infringing content can spread rapidly and in large volumes.

The court ruled as follows: "The defendant knew or should have known that network users were using its services to infringe upon the right of information network dissemination, yet failed to take necessary preventive measures." This constituted negligence.

There was an interesting point. The court dismissed the plaintiff's broader claim, which demanded deletion of all materials and data related to the Ultraman character. Not all Ultraman-related use constitutes infringement. Users can still legitimately use the character for learning, research, and personal enjoyment.

(3) The Classified-and-Tiered (Fenlei Fenceng) Liability Framework: Reconciling Technology Neutrality with the Duty of Care

The most significant contribution of the Hangzhou ruling was establishing the 'classified-and-tiered' (fenlei fenceng, 分类分层) liability framework. This is a framework that assigns responsibility by distinguishing between stages of AI services.

The court divided the generative AI process into two stages.

The input stage (data training phase). The court applied a relatively lenient standard to the use of copyrighted works for model training. It recognized that data training could constitute 'reasonable use (fair use)' for the purposes of technological innovation and creation. This reflected a policy consideration aimed at not impeding technological progress.

The output stage (content generation and dissemination phase). By contrast, when generated outputs were substantially similar to existing works, replacing the market or harming rights holders' interests, the court imposed strict liability.

This framework is practical.

If the mere act of collecting training data is treated as a violation, the industry cannot grow. But if the products of that training replicate originals and flood the market, rights holders are harmed. Chinese courts drew a line between these two stages when assigning liability.

The court also introduced the standard of 'reasonable duty of care within the same industry.' AI service providers cannot hide behind technological neutrality to evade responsibility. They must fulfill a reasonable level of monitoring and filtering obligations, taking into account the reach of the technology and its commercial nature.

The Hangzhou Intermediate People's Court (second instance) put even greater emphasis on this point. "The determination of a platform's negligence should not be limited to the application of 'Safe Harbor' immunity or the 'red flag' (obvious infringement) doctrine. The balancing relationship between copyright protection and AI technological innovation and development must be considered."

This classified-and-tiered liability framework provides clear guidelines for AI companies.

First, there is relative leniency regarding training data. The failure to obtain copyright permission for all training data does not automatically constitute a violation. Of course, this applies within China. Different standards may apply in the United States or Europe. Second, the standard for outputs is strict. When content resembling well-known copyrighted works is generated and disseminated, the platform is liable. The defense of "we only provided the tool" does not hold.

Third, prevention is better than cure. Removing content after infringement occurs is not enough. Preventive measures must be taken in advance: filtering keywords related to well-known characters, issuing user warnings, labeling AI-generated content, and more.

While American AI companies hide behind the shield of "fair use," adopting a strategy of "train first, apologize later," Chinese courts have taken the opposite stance: "Be careful from the start." This imposes regulatory costs on Chinese AI companies, but it also builds a fortress of protection for copyright holders.

C. Standards for Assigning Rights to AI-Generated Works

(1) The User's Creative Intent and Investment

Who owns what AI produces? This question sits at the heart of every AI business model.

Three answers are possible.

The company that developed the AI model.

The person who used the AI.

Or no one at all (the public domain).

Consider the consequences of each.

What if every AI-generated work belonged to the AI developer? If OpenAI held the copyright to all text produced through ChatGPT? Users paying a monthly subscription would own nothing they created. They would become digital sharecroppers. The incentive to create would vanish.

What if every AI-generated work fell into the public domain? If Disney made a movie with AI and anyone could copy and sell it the moment it premiered? No one would bother making AI films. Investment would dry up.

Chinese courts chose a third answer. Copyright in AI-generated works belongs to the human who used AI as a tool.

The core principle established in the Beijing Internet Court's "AI Wen Sheng Tu" ruling is "user attribution." The court granted copyright not to the AI model developer (Stability AI) but to the user (Li) who used that model to generate the content.

The reasoning rests on "creative intent" and "investment." The AI model developer supplied the tool (the paint and brushes). It was the user who decided what specific picture to paint (intent), who spent time researching prompts, adjusting parameters, and producing the final output (investment). The court held that when a user exercises primary control over the AI and uses it as a tool to express their own thoughts and emotions, rights in the resulting work properly belong to that user.

This draws on Locke's labor theory of value. "The one who sweats reaps the fruit." The same principle applies in the digital creative environment.

The relationship with AI service terms of service raises an interesting question. Many AI platforms state in their terms that "all rights in generated content are assigned to the user." Chinese courts respect the validity of such private agreements. But there is a prerequisite: the output must possess the "originality" required for copyright protection. Terms of service alone do not automatically create copyright.

This approach balances the interests of AI model developers and users. Developers earn revenue through platform fees or subscriptions. Users own their creative works and can use them commercially. Both sides have incentives.

Guidelines issued by China's National Intellectual Property Administration (CNIPA) in 2025 confirmed this direction. "Whether AI-generated images or novels have copyright depends primarily on whether the content is filled with creativity or originality, and this must be analyzed on a case-by-case basis."

How to prove rights attribution also matters. A September 2025 ruling by the Beijing Internet Court clarified this point. When claiming copyright in AI-generated works, the author must demonstrate the following.

Their own creative thinking: why they wanted to create this image, what aesthetic goal they had. The content of their input commands: what prompts they used, what parameters they set. The process of selecting and modifying the generated content: why they chose this particular result from among several, what edits they made. A person who keeps no records will struggle to assert rights. This is a crucial lesson for practitioners. When creating content with AI, document the creative process. Prompt logs, parameter settings, the journey from draft to final version. All of it becomes a receipt in court.

(2) The AI Chair Design Case: Unclear Scope of Copyright Protection

Not every AI-generated work receives copyright protection. Cases that fall on the boundary remain subjects of debate.

On March 19, 2025, the Zhangjiagang People's Court issued a notable ruling. A designer had used Midjourney, an AI image generation tool, to design a butterfly-shaped chair. It had a jelly-like texture with a distinctive mix of pink, blue, and orange. He posted the image on social media, sharing the prompt he had used to create it.

A furniture manufacturer took interest in the design. Negotiations followed but broke down without an agreement. The manufacturer then used the publicly shared prompt to generate a similar image in Midjourney and began manufacturing and selling actual butterfly chairs based on it.

The designer filed suit, alleging copyright infringement and unfair competition.

The court dismissed the designer's claims.

The reasoning is instructive. The court ruled that "the scope of copyright protection for AI-generated works is unclear."

This is where the problem gets complicated.

First, the prompt disclosure problem. The designer himself had published his prompt. Anyone could use that prompt to generate similar images. This resembles a chef who publishes a recipe and then sues someone for cooking from it. Recipes themselves are not subject to copyright protection.

Second, the distinction between idea and expression. Copyright law recognizes a principle called the "idea-expression dichotomy." Ideas are not protected. Only the specific expression of an idea is. Anyone can have the idea of "a butterfly-shaped chair with a jelly-like texture." The question is about the image that concretely expresses that idea. But AI can generate different images from the same prompt. So which image counts as "expression" and which remains "idea"?

Third, the blending of function and aesthetics. A chair is a functional object. Design protection for functional objects falls closer to design patent (industrial design right) territory than to copyright. Even if the AI-generated design drawing is protected by copyright, whether that copyright extends to the actual chair manufactured from it is a separate question.

Fourth, the training data problem. AI generates new images by learning from vast numbers of existing chair designs. If the generated output is a recombination of training data, how much originality does it contain? Can it escape the problem of "substantial similarity" to existing designs?

This case shows that Chinese courts do not have all the answers either.

Under the Beijing Internet Court's precedent, if a user entered specific prompts addressing the aesthetic aspects of a chair (shape, color, ornamentation) and refined them to impart "originality," the design drawing itself could be protected as an artistic work. But how far that protection reaches, especially when the design is realized as an actual product, remains unclear.

Chinese courts maintain a "case-by-case" approach to AI-generated works. Whether AI-generated industrial designs or blueprints can receive full copyright protection depends on these factors: the degree of the user's intellectual input; similarity to existing works; and whether functional elements can be separated from artistic elements.

This is uncertainty. Uncertainty is risk. Companies seeking to secure rights in AI-assisted industrial design need intellectual property strategies that go beyond copyright to include patents, design rights, and other protections.

Taken together, China's AI copyright rulings lead to this conclusion. They are pioneering but incomplete. Courts actively recognize rights "when a human leads and AI assists." They remain cautious "when AI leads and a human gives simple instructions." And the vast gray zone between these two poles has yet to be sorted out. In the end, the key to rights attribution is documentation. This is no longer an era where you can assert creative ownership through words alone. Rights attach to the person who recorded what choices they made, what values they adjusted, what results they discarded, and what results they kept. Chinese courts are moving in this direction.

The Chinese AI copyright rulings covered in this chapter offer important reference points for the global legal community. The next chapter examines how China protects personality rights related to AI, particularly voice rights and likeness rights. Those rulings were every bit as forward-looking.