A. Protection of Voice Rights and Portrait Rights

(1) 2023 Beijing Internet Court 'AI Voice' Ruling: Recognition of Voice Rights as an Independent Personality Right

One day in May 2023, a voice actress surnamed Yin, living in Beijing, froze while browsing the internet. She heard her own voice. It was reading sentences she had never recorded. The timbre was the same, the intonation was the same, even the timing of her breaths was the same. But she had never seen that text before.

Her voice had become an AI voice product called 'Mo Xiaoxuan' and was being sold on an app called 'Moyin Workshop.' 3.2 billion plays. In someone's presentation, in someone's YouTube video, in someone's audiobook, a voice that was hers yet not hers was playing.

Yin went to see a lawyer.

This is where things get complicated. Yin had previously signed a contract with a media company and recorded audiobooks for them. The contract stated that 'copyright in the recordings belongs to the company.' The company handed these audio files to another software company, which used Yin's voice as AI training data. Then yet another platform sold the resulting product commercially.

The defendants' logic was straightforward: 'We legitimately purchased the audio files. What we do with those files is our business.'

But on April 23, 2024, the judge at the Beijing Internet Court posed an entirely different question: 'Are the audio files and the voice the same thing?' To understand why this question matters, we need to briefly examine the structure of China's Civil Code. China implemented its new Civil Code in 2021. This code contains an independent chapter dedicated to 'personality rights.' Personality rights, put simply, are rights over 'the things that make me who I am.' My name, my face, my reputation. These are different from my property. They cannot be sold. They cannot be transferred. They cannot be separated from me.

The court placed voice squarely in this category. Article 1023, Paragraph 2 of the Civil Code explicitly states: 'The provisions on the protection of portrait rights shall apply mutatis mutandis to the protection of a natural person's voice.' The judge relied on this provision to recognize voice as an independent personality right.

The key sentence in the ruling read: 'Even if a voice is generated by AI, if the general public can hear it and associate it with a specific person, it constitutes that person's voice right.'

The court called this 'identifiability.' Timbre, intonation, pronunciation habits, breathing patterns. When all these elements combine to point to a specific individual, that is not mere 'data'; it is 'part of a person's identity.'

The critical implication of this logic lies in the separation of copyright from personality rights. It is true that Yin transferred the copyright of her audio files. But copyright is a right over 'expressions.' Personality rights are rights over 'oneself.' Selling a file I recorded does not mean I sold my voice itself. Just as selling the copyright to my photograph does not grant the buyer the right to use my face anywhere they please.

The court went one step further. It traced the chain of liability. The media company that provided the recording data, the software company that developed the AI voice, the platform that sold it, the business that purchased and operated it. The court held all of them jointly liable. It assigned different proportions of responsibility according to each party's role and profit, but no one could escape by saying 'I only provided the technology.'

The damages award was 250,000 yuan (approximately USD 35,000). The amount itself might be pocket change for large corporations. But its significance could not be measured in money. Every AI company in China now had to ask itself before training on a voice actor's recordings: 'Did this person truly consent to this? Does the recording contract cover AI training?'

This ruling came two months before Scarlett Johansson expressed her outrage over OpenAI's 'Sky' voice in the United States. The Chinese court moved first. While technology raced across borders, the law followed at its own pace in each jurisdiction. The voice actress in Beijing reclaimed her voice. And the ruling became a milestone for how far the protection of personal identity extends in the age of AI.

(2) 2025 AI Virtual Human Portrait Rights Ruling

On September 10, 2025, the Beijing Internet Court published eight 'typical AI-related cases.' The eighth case drew particular attention. A celebrity surnamed He had been brought to life as an 'AI companion' inside a mobile app, without his knowledge.

The app was a household budget tracker. It looked ordinary on the surface. But hidden inside was a special feature. Users could create their own 'AI companions.' They could give them names, set profile photos, and define relationships. Boyfriend, girlfriend, sibling, mother. Anything was possible.

The problem was that countless users had set the real-life celebrity He as their companion. They uploaded his photos. They entered his name. They defined relationships like 'boyfriend' or 'older brother.' The app's algorithm analyzed this information. And it began recommending the 'He' character to other users.

He himself knew nothing about this. He had never consented.

The court found this to be a personality rights infringement. Here, the critical issue emerged. The app operator argued: 'We simply hosted content that users uploaded.' But the court rejected this logic. The app provided a 'relationship setting' feature, recommended characters through its algorithm, and used this to drive user engagement. That was not simple hosting. It was a structural design that turned a specific person's identity into a 'commodified substitute.'

This ruling leads to a broader question: how far does 'portrait rights' extend in the age of AI?

Traditionally, portrait rights applied to faces captured in photographs or videos. But AI technology has eroded these boundaries. It is now possible to reproduce someone's 'feel.' Without a single actual photograph, a virtual character can be created that evokes a specific person. Speech patterns, expressions, gestures. All of these can become data for training.

Another ruling from August 2025 addressed the opposite side of this issue. A company that created virtual digital characters 'A' and 'B' sued a former employee who had sold the characters' 3D models without authorization. The court ruled that if a virtual digital character's image possesses originality, it can receive copyright protection as a work of fine art.

Two rights intersect here: the portrait rights of real people and the copyright of virtual characters. The more a virtual human resembles a real person, the more these two rights collide. Suppose someone creates an AI virtual human modeled on a famous celebrity. The virtual human's appearance is the production company's copyrighted work. But if it 'evokes' a real person, that real person's portrait rights become an issue.

Chinese courts have proposed 'identifiability' as the standard for this problem. If an average user can look at a virtual character and think of a specific natural person, that character falls within the scope of portrait rights protection. This is not a question of technology but of perception. The standard is social association, not code.

The implication of these rulings is clear. No matter how advanced AI technology becomes, if what it creates infringes on a real person's identity, the law intervenes. Chinese courts were not dazzled by technological sophistication. Instead, they asked the simplest question: 'Who do people think of when they see this?'

(3) Deepfake 'Face Swap' App Portrait Rights and Personal Information Infringement Ruling

On June 20, 2024, the Beijing Internet Court issued its ruling in China's first 'AI face-swap app' infringement case. The plaintiffs were two women, both models famous for short videos in traditional Chinese style. They wore classical garments resembling hanfu, styled their hair in period fashion, applied traditional makeup, and posed before the camera. Their videos had accumulated millions of views.

One day, they discovered something strange. Templates featuring their exact costumes, hairstyles, makeup, lighting, and camera angles were being sold on an app. Only the faces had been erased. Users could pay to composite their own faces onto these templates.

The plaintiffs asserted two claims. First, portrait rights infringement. Second, personal information infringement.

The court's reasoning was subtle. And it was precisely that subtlety that made this ruling significant.

First, regarding portrait rights, the court found 'no infringement.' Under the Civil Code, portrait rights infringement includes the 'creation, use, or disclosure of a portrait.' The key here is 'identifiability.' Because the plaintiffs' faces had been removed from the templates, the plaintiffs could not be 'identified' in the final product. Anyone viewing the composite video with another person's face would not think of the plaintiffs.

But the court did not stop there. Second, regarding personal information, the court found 'infringement.' This is the heart of the ruling. In the process of creating the templates, the defendant collected and processed the plaintiffs' original videos. Those videos contained the plaintiffs' faces. Facial information constitutes 'sensitive personal information' under China's Personal Information Protection Law (PIPL).

The court's logic went like this: the fact that the plaintiff's face is not visible in the final product does not make the process lawful. The defendant collected facial data without the plaintiffs' consent. It analyzed that data. It created a commercial product based on that data. The process itself constitutes personal information infringement.

This separation logic carries powerful practical implications. Deepfake service providers commonly argue: 'The plaintiff's face does not appear in the output, so portrait rights are not at issue.' But under this ruling, service providers cannot escape liability based solely on the appearance of the output. The court examines the entire process: collection, processing, and provision. If personal information was improperly handled at any stage, liability arises.

The defendant raised another defense: 'We outsourced the face-swapping technology to a third-party company. They did the actual processing.' The court rejected this as well. If the defendant determined the method and scope of the service and profited from it, outsourcing is not a basis for exemption from liability.

The damages award was 3,500 yuan (approximately USD 490). Economic losses of 2,500 yuan and emotional distress of 1,000 yuan. The amount was small. But Judge Sun Mingxi told reporters after the ruling: 'I hope this decision helps standardize the application of emerging technologies and promotes the healthy development of the digital economy.'

This ruling connects to China's broader AI regulatory framework. The 'Provisions on the Administration of Deep Synthesis Internet Information Services,' which took effect in January 2023, imposes obligations on deepfake service providers to obtain the consent of the person depicted and to label generated content. The court's civil ruling confirmed the effectiveness of this regulation in practice. The fact that technology makes something possible does not mean it is permissible. The faster technology advances, the stronger the duty of care becomes.

B. AI Style Imitation and Unfair Competition

(1) TikTok v. B612 Filter Case: 91.7% Similarity

In June 2020, Douyin, the Chinese version of TikTok, launched a filter called 'Transformation Manga Effect.' It detected the user's face in real time and converted it into a Japanese anime style. The person on screen suddenly acquired large eyes and a pointed chin, transformed into a cartoon character. The filter was a massive hit. Hundreds of millions of users filmed videos with their faces turned into manga.

Two months later, a nearly identical feature appeared on the competing app B612. It was called 'Girl Manga Effect.' The results were virtually the same.

ByteDance's lawyers were suspicious. This was no coincidence. They filed suit against B612's operator. But there was a problem. How do you prove 'copying'?

In a traditional copyright dispute, you compare source code. You look at how much of A's code matches B's code. But AI filters are different. Two companies may have used the same publicly available training data. They may have adopted the same neural network architecture. Similar outputs do not necessarily mean one copied from the other.

ByteDance took a different approach. Instead of copyright, it claimed 'unfair competition.'

What is unfair competition? Here is a simple analogy. Suppose you run a restaurant. You spent years and money developing unique recipes. You designed the interior in a distinctive way. Customers recognize your restaurant. Then one day, a nearly identical restaurant opens across the street. The menu is similar. The interior is similar. Customers get confused. That is unfair competition. It is free-riding on the effort and investment you put in.

The court ordered a technical appraisal. Experts analyzed the internal structures of both apps' AI models. The results were shocking. The overall network architectures were nearly identical. Of 36 convolutional layers, 33 matched exactly. Even the parameter settings matched in detail. Expressed numerically, the similarity was 91.7%. On March 31, 2025, the Beijing Intellectual Property Court ruled in favor of ByteDance on appeal. The court recognized two things. First, Douyin's AI model structure and parameters were the product of 'massive resource investment in development and training.' Second, these constituted 'competitive interests that bring market advantages and business revenue' to Douyin. Therefore, B612's unauthorized imitation to provide a similar service was 'an act of unfair competition that infringes legitimate competitive interests and disrupts market order.'

(2) Protecting Competitive Interests in AI Model Architecture

The core significance of this ruling is that 'AI model structure and parameters' can now receive legal protection.

An AI model is similar to a cooking recipe. You gather ingredients (data), apply cooking methods (algorithms), and taste and adjust countless times (tuning) until it is complete. Reverse-engineering the recipe from the finished dish is difficult, but not impossible. If someone eats at your restaurant every day and analyzes your dishes, they could eventually replicate the flavor.

In the AI industry, this is called 'knowledge distillation.' You send millions of queries to a competitor's AI model, collect its responses, and use them to train your own model. You are not stealing source code. You are not extracting data. But in effect, you are copying the other party's 'intelligence.'

Chinese courts have taken a clear position on this issue. An AI model's internal structure, parameters, and tuning outputs can be protected as a company's 'competitive interests.' Where traditional copyright or patent law struggles to provide protection, the Anti-Unfair Competition Law fills the gap.

This is not just a matter of technology. It is a matter of investment. Douyin collected hundreds of thousands of images. It hired designers. It consumed thousands of hours of computing resources to train its model. All of these costs are concentrated inside the AI model. If someone copies only the output, all of that investment evaporates.

Proof is difficult, of course. Infringers typically claim 'independent development.' Rights holders have no way to look inside the other party's system. That is why these lawsuits do not end with output similarity alone. Circumstantial evidence becomes critical: development records, data sources, parameter update histories, employee transfer records.

This ruling sent a clear message to China's AI industry. Free-riding on a first mover's results carries legal risk. Latecomers must be able to prove an independent development process. Otherwise, the number 91.7% will be waiting for you in court.

C. Multi-Layered Liability Structure

(1) Users Bear Primary Liability

To understand China's AI regulation, imagine a large pyramid. At the top sits the state. It sets the rules. Below that are platform companies. They enforce the rules. At the bottom are users. They follow the rules. When something goes wrong at any level, someone on that level bears responsibility.

The first principle is straightforward. The person who uses the tool is responsible for the result.

One of the model cases published by the Beijing Internet Court illustrates this principle clearly. A user took someone else's WeChat profile photo, used AI to create a mocking image, and posted it in a group chat. The victim filed a lawsuit.

The court held the user liable. The excuse that 'it was just a joke' did not work. The fact that AI generated the image did not exempt the user either. The user entered the prompt. The user pressed enter. The user distributed the result. The decision-maker at every step was a human being.

This principle establishes a foundational legal doctrine for the AI era. AI is a tool. If you hit someone with a hammer, the hammer manufacturer is not liable; the person who swung it is. The same applies to AI.

(2) Service Providers as Normative Agents of Liability

But AI differs from an ordinary hammer in one respect. A hammer does nothing unless you swing it. AI is different. AI learns. It recommends. It optimizes. Sometimes it produces results the user never intended.

That is why Chinese courts impose a second layer of liability: the responsibility of 'service providers.'

Consider the AI voice case discussed earlier. In the case where voice actor Yin's voice was converted into AI without authorization, the court held the developer, the sales platform, and the operating company all liable. Each argued: 'I only provided technology,' 'I only ran the platform,' 'I only purchased the product.' The court rejected every one of these defenses. The court's reasoning was this: service providers are 'normative agents.' They are not neutral entities that merely provide technology. They determine the manner and scope of the service and profit from it. They must therefore bear corresponding responsibility.

This responsibility operates on two levels. First, liability as a data processor. When collecting sensitive information like faces or voices, consent must be obtained. The purpose of collection must be stated. Security measures must be in place. Failure on any of these points is immediately unlawful. Second, management liability as an information service provider. AI-generated content must be labeled as such. Systems to prevent misuse must be built. When reports come in, action must be taken.

The 'Interim Measures for the Management of Generative AI Services,' which took effect in August 2023, codified these obligations. Court rulings confirm their enforceability. Companies can no longer dodge responsibility by saying 'we are just a platform.'

(3) Platforms' Duty of Prevention: Model Optimization Training

The most distinctive feature of China's AI regulation is its emphasis on 'prevention.' Where Western regulation generally takes a reactive approach of 'punish when problems arise,' China commands: 'prevent problems from arising.'

This is called the 'model optimization training obligation.' Platforms must continuously adjust their AI to prevent it from generating illegal content. This goes beyond simple keyword filters. It means adjusting model weights, applying reinforcement learning from human feedback (RLHF), and monitoring outputs.

For example, suppose someone types 'draw Ultraman' into an AI image generation service. Ultraman is the intellectual property of a Japanese company. If the AI generates an image resembling Ultraman, copyright infringement occurs. In the Guangzhou Internet Court ruling, the court held the service provider liable. 'Why was the keyword Ultraman not filtered? Why were Ultraman-related images not removed from the training data?'

Some criticize this obligation as over-regulation. Preventing all possible copyright infringement in advance is technically impossible. The court acknowledges this. What the court demands is not 'perfect prevention' but 'measures proportionate to the level of control and foreseeable risk.' If a platform exercises reasonable duty of care, it is not held liable for individual users' infringing acts. But if a platform maximizes its monetization structure while claiming 'we are a neutral tool,' the court will not accept that argument.

This multi-layered liability structure makes China's AI ecosystem unique. Companies invest more resources in building 'safe' models that meet regulatory demands than in developing innovative ones. Thousands of workers review AI outputs, label data, and adjust models.

Ironically, this strict control system also serves as a powerful moat for Chinese AI companies. It is nearly impossible for foreign firms to enter the Chinese market while bearing this complex and heavy liability structure. In the end, the survivors on China's AI legal battlefield are not those with the best technology. The winners are those who navigate this multi-layered web of liability most skillfully, those who most diligently fulfill the 'management duties' demanded by judges and regulators.

This is the unvarnished reality of China's AI governance model. Whether this model works or suffocates innovation is too early to judge. One thing is certain. Chinese courts are prepared to hold someone accountable for every problem AI creates. That someone may be the user, the service provider, or the platform. But 'the AI did it' is not an excuse.