With the rapid iteration of artificial intelligence (AI) technology, applications such as speech synthesis, voice cloning, and text-to-speech have widely penetrated into daily life. However, while technology brings convenience, it also brings risks: when your voice is "stolen" by AI, how do you protect your rights?
On September 29, 2026, the Shanghai No. 1 Intermediate People's Court concluded Shanghai's first dispute over the protection of natural persons' voice rights caused by artificial intelligence synthesized speech in accordance with the law. It ruled that the platform operator constituted voice infringement and compensated the rights holder for economic losses of 50,000 yuan.
Ms. Wang is a real-name certified voice actor on a certain App. She usually gets paid by dubbing and publishing works on the App. In January 2025, Ms. Wang learned from a friend that Company A was suspected of using Ms. Wang’s voice in its Internet promotion and recruiting activities. After verification, Ms. Wang found that the relevant sounds used in the event were suspected to be synthesized by AI, so she notarized the audio to fix the evidence and filed a lawsuit with the Changning District People's Court.
Ms. Wang believed that she had never cooperated with Company A and had not authorized any authorization. Company A's unauthorized use of her voice constituted infringement, so she requested Company A to stop the infringement and compensate for economic losses of 300,000 yuan.
Company A argued that the audio involved in the case was indeed generated by the AI developed by the company, but the specific source and fed materials could not be confirmed due to the resignation of the former employee. The company has never known the identity of Ms. Wang’s voice actor before, has not collected or used her voice, and the existing evidence cannot directly prove that it used Ms. Wang’s voice in promotional activities.
During the trial of the case, in order to prove the fact of infringement, Ms. Wang conducted a speech analysis and comparison between the audio involved in the case and her own voice through judicial appraisal. The appraisal opinion shows that among the 28 formant acoustic indicators of the audio involved in the case and Ms. Wang’s own audio, 24 have deviations of less than 10%, 16 of which are less than 5.36%, and the parts that are relatively similar and highly similar between the two reach 90%.
The Shanghai Changning District People's Court held after hearing that natural persons have the right to voice, and the voice of the right holder may not be produced, used, or disclosed without the consent of the right holder. Under the preponderance of evidence available to the plaintiff, it is inferred according to the rules of evidence that the defendant Company A came into contact with Ms. Wang’s voice in its business activities and used her voice as material to generate the audio involved in the case without her consent. The audio is highly similar to Ms. Wang's voice, enough to make the public, especially her relatives and friends, mistakenly believe that she is the voice, infringing on Ms. Wang's right to voice and causing losses. There is a causal relationship between the two, and Company A should bear infringement liability according to law. Based on the duration and scope of the AI-generated audio involved in the case, the plaintiff’s damage results and the degree of fault of Company A, Company A was judged to compensate Ms. Wang for a loss of 50,000 yuan and bear the necessary expenses incurred by Ms. Wang in litigating the case.
Ms. Wang and Company A were dissatisfied with the first-instance judgment and both appealed.
The Shanghai No. 1 Intermediate People's Court held that using a natural person's voice as training corpus without the natural person's consent, imitating the natural person's timbre, intonation and pronunciation style to generate a synthetic human voice that can identify the natural person should be deemed as infringement of the natural person's voice rights. In this case, Ms. Wang, as a dubbing actor, has provided evidence to prove that her dubbing works are published on public online platforms, and Company A has channels and possibilities for obtaining information about the source of her voice. At the same time, after analysis of the appraisal report, the synthesized voice is highly similar to the human voice, which is enough for the general public to associate specific identities and is identifiable.
Although Company A denied it, as the controller of the original training data, it failed to provide sufficient evidence on the legality of the source of the feeding materials, the failure to utilize the characteristics of natural persons' voices, and the failure to cause public confusion and misunderstanding. It should bear the adverse consequences of failure to provide evidence. Based on this, it can be determined that Company A commercially uses the synthesized voice that can identify personal identity to the public without Ms. Wang’s consent, which constitutes voice infringement.
On September 29, 2026, the Shanghai No. 1 Intermediate Court rejected the appeal at the second instance and upheld the original verdict.
Currently, many AI voice product developers have a misunderstanding. They believe that the sound generated by AI is an algorithmic reconstruction of voiceprint characteristics and is a product of technical simulation rather than a direct copy of the original recording. Therefore, it does not involve infringement. In fact, this understanding confuses the difference between "technological generation method" and "legal evaluation object".
What the law protects is the personality identification interest carried by the sound. As long as the AI-synthesized sound can identify a specific natural person, no matter what algorithm path it is generated through, unauthorized use may constitute an infringement of the sound rights and interests.
In this case, the core training data, algorithm logs, and material source records of the AI speech synthesis model are all unilaterally controlled by the platform, and it is objectively difficult for the rights holder to obtain them. If the general rule of "whoever claims it is responsible for providing evidence" is adhered to and requires rights holders to prove that the platform uses their voice for training, it will greatly raise the threshold for rights protection.
The clear rule in the judgment of this case is that the rights holder only needs to prove that his voice has a reasonable channel to obtain and that the synthesized voice is highly similar to his voice, and the preliminary proof is completed. After that, the burden of proof shifts to the platform, which will prove that the legal source of the training data or the synthesized voice has nothing to do with the voice of the rights holder. This burden of proof allocation rule responds to the structural problem of biased evidence in AI infringement cases and prevents platforms from evading liability on the grounds of "technical black box" or "employee resignation cannot be confirmed." For the AI industry, this rule also sends a clear signal: Whoever controls the data should bear the corresponding obligation to provide compliance evidence.
The commercial use of sounds must be authorized, and technical neutrality does not exempt censorship responsibilities.
Voice rights have both personality attributes and property attributes. For groups such as professional voice actors whose core professional resources are voice, the commercial value of voice is particularly prominent. AI speech synthesis technology itself is neutral, but technology neutrality does not mean use neutrality. Business entities have the obligation to review the legality of the source of sound materials before putting AI-synthesized speech into commercial scenarios. They cannot be exempted from the basic responsibility of obtaining authorization and compliant use on the grounds of "not knowing the identity of the rights holder" or "automatically generated by AI".
Reporter Zhang Haiyan
