Shanghai's first AI voice infringement case upheld on appeal, platform ordered to pay 50,000 yuan in compensation
In Shanghai's first AI-synthesized voice infringement case, the appellate court rejected the appeals of both parties and upheld the first-instance judgment, finding that the platform operator constituted voice infringement, ordering it to compensate the voice actor 50,000 yuan for economic losses and to bear necessary litigation expenses. The court drew on the principle of access plus substantial similarity to allocate the burden of proof; after the plaintiff completed preliminary proof, the platform, which possessed the training data, was required to explain the source of the voice, but the platform responded only that it could not verify because the employee had left, and thus failed to discharge its burden of proof. Because the plaintiff was unable to provide credible comparable licensing prices, the claim of 300,000 yuan was not supported; the court determined the compensation at its discretion by comprehensively considering the duration of use, the scope of dissemination, and the voice's contribution to commercial revenue.
Why read this
AI score 65
The case fully presents the judicial considerations in AI-synthesized voice infringement regarding the selection of expert appraisal, allocation of burden of proof, and discretionary determination of damages, and can serve as a reference for similar disputes.
Topics
Body · AI translation
In a conference room on the fifth floor of the Shanghai Changning District People's Court, more than a dozen judges and judge assistants sat around together. The question they were discussing sounded simple to outsiders: should an appraisal be conducted? This was the first professional judges' meeting during the trial of Shanghai's first AI voice infringement case. The plaintiff is a voice actress. She believed that her voice had been used without permission after being synthesized by a company's AI for promotional activities. The defendant is a company that admitted that the voice involved was generated by the AI it developed, but argued that the training corpus came from a massive speech database, that there was no evidence proving that the original feeding materials involved the voice of the plaintiff Wang Lin (pseudonym), and that the relevant technical employees had already left, making it impossible to confirm the source of the materials. The first-instance court, based on all the evidence in the case, found that the accused conduct infringed Wang Lin's voice rights and interests, and issued a judgment in June 2026. After the first instance, both parties appealed. On September 29, the Shanghai No. 1 Intermediate People's Court issued a second-instance judgment, rejecting both parties' appeals and upholding the original judgment: the platform operator constituted voice infringement, was ordered to compensate the rights holder 50,000 yuan in economic losses, and to bear the relevant necessary litigation expenses. With this, Shanghai's first dispute over the protection of a natural person's voice rights and interests caused by AI-synthesized speech came to a conclusion. AI technology is developing rapidly, and the commercialization of AI voice has run ahead of judicial practice. Deng Xin, the judge handling the first instance, told reporters that in such disputes, leaning toward the technology side may sacrifice the rights and interests of individuals, while leaning toward the rights-protection side may inadvertently harm the industry. In the absence of sufficient similar cases, it is truly difficult to strike the right balance. When ready-made legal provisions are difficult to apply directly, judicial personnel must find the point and boundaries for expansive interpretation. This requires both a sense of responsibility and a sense of proportion. A voice recognized by a friend The dispute began around January 2025. Wang Lin is a voice actress with real-name verification on a certain App, who normally earns remuneration by posting voice-over works. One day, a friend told her that in a company's internet promotion and new-user acquisition campaign, there seemed to be her voice. In the campaign, announcements such as "Congratulations, you've won" and prompts telling users to proceed to the next stage all sounded like her. The friend asked her whether she had made an advertisement for the other party. Wang Lin discovered that the voice was highly similar to her own, and that she had never said such things; the voice was suspected to have been synthesized by AI. She notarized the relevant audio and filed a lawsuit with the court in March, demanding that the infringement be stopped and 300,000 yuan in economic losses be compensated. In May 2025, Deng Xin, deputy chief judge of the Commercial Trial Division (Internet Case Trial Division) of the Shanghai Changning District People's Court, received the case. After reading the materials, his first reaction was that, for the judiciary, "this is a 'good' case." In his view, there are two kinds of "good" cases. Some have clear difficulties and concentrated disputes and are worth in-depth study; others are novel enough that adjudication not only resolves the immediate dispute but may also provide direction for similar cases in the future. Wang Lin's case belongs to the latter. Novelty also means that there is limited experience to draw on. For cases involving generative AI infringement, there are no written legal provisions that fully apply, nor are there fully unified adjudication rules. The team began to divide the work of searching for similar cases, consulting literature, and hearing the views of both parties. In their search, they noted that in April 2024, the Beijing Internet Court had already concluded the country's first case of infringement of personality rights by AI-generated voice. The voice of voice actor Yin was used to train a text-to-speech product, and the court ordered the defendant to pay 250,000 yuan in compensation. As for voices processed by AI technology, if the public can identify a specific natural person based on timbre, intonation, and pronunciation style, then that voice falls within the scope of protection of a natural person's voice rights and interests. Using a voice processed by AI technology without the natural person's permission constitutes infringement of the natural person's voice rights and interests. In other words, the legal question clarified by the Beijing case is: a voice that can still be identified as belonging to a specific natural person after AI processing may be included in the scope of protection of voice rights and interests. Wang Lin's case is Shanghai's first AI voice infringement case, and what it must answer is the legal question that arises immediately afterward: how to determine that the accused infringing audio is identical to Wang Lin's own voice—in other words, how to prove that this AI voice points to Wang Lin. Deng Xin explained that traditional voice infringement may involve someone directly intercepting an audio clip from a film, program, or the like, or extracting words and sentences from several recordings and reassembling them. Usually the original carrier can still be found for comparison. AI synthesis, by contrast, may make a person's voice say things the person never said, and there is no original recording whose content corresponds exactly. Simply put, when faced with a piece of generative AI audio that sounds like person A to one listener and not like person A to another, how should the judiciary determine that the voice comes from the same person? And what considerations should there be? Should an appraisal be conducted? The answer may differ from popular common sense. Because at the judicial level, the core criterion is not "whether we want to" but "whether it is necessary." Deng Xin explained that initiating a judicial appraisal has strict prerequisites. The court will first determine whether an appraisal is the only, necessary, and feasible means of resolving the dispute. The considerations include whether it is a fact that can be inferred based on ordinary life common sense and rules of experience, whether it is relevant to the focus of the dispute, whether the inspection materials are missing or of unknown origin, and whether the appraisal techniques and methods are reliable... AI concept image at the professional judges' meeting. In August 2025, the first professional judges' meeting in this case was held in a conference room on the fifth floor of the Changning District People's Court. The participants were more than a dozen judges, assistants, and others. After hearing the basic facts of the case and the current progress, "the main issue we want to discuss at this meeting is whether everyone believes an voice appraisal should be initiated in this case," the meeting host said. "Please express your opinions one by one." Deng Xin recalled that roughly three views emerged at the meeting. One view follows the civil litigation principle that "the party who asserts must prove": "Since the plaintiff claims the defendant used her voice, she should naturally produce direct evidence; if she cannot, the lawsuit should be dismissed." Some added that although the plaintiff had the matter notarized when filing suit, she did not obtain an independent expert appraisal; she sued purely because a friend said they "heard a similar voice" and because it sounded like her to her own ear, so she failed to fulfill the initial burden of proof and the case should be dismissed outright, with "no need for appraisal." As soon as this view was voiced, someone immediately objected: AI technology involves a "black box," and the generation process is difficult for outsiders to observe or reveal; most such data is held by companies, and individuals will inevitably struggle to obtain it. Placing the entire burden of appraisal and proof on the plaintiff is an excessive demand. This also represents the second view. They hold that the defendant is a major platform, the activity at issue was a flagship event, and its evidentiary capacity far exceeds the plaintiff's; the defendant from beginning to end only objected on the grounds that "the staff member has left and it cannot be verified," without producing any evidence. In past voice infringement cases, reference can be made to personality-right disputes such as portrait rights. Courts generally do not order an appraisal; they simply compare the image with the rights holder, and the result is obvious. Whether a voice sounds alike can likewise be judged; if the two truly cannot be distinguished, the platform should be found to have infringed. Once this view was voiced, others objected. The objectors did not disagree because they could not hear any similarity at all. On the contrary, Deng Xin recalled that the people who participated in the discussion at the meeting, as well as the people's assessors who had participated in the trial activities before the meeting, could all hear the similarities between the voices provided by the two sides. The dispute was whether such auditory perception could serve as a sufficient basis for finding infringement. The objectors argued that portrait images differ from voices: a seasoned voice actor is capable of imitating voices of various timbres and intonations, and the human ear alone may not be enough to rule out other explanations. "The key is still a question of balancing interests," Deng Xin summarized. "One view favors the rights holder, the other favors the platform; both must be considered, because one is the protection of a person's rights and the other represents the orderly development of AI technology." "If you bring me a clip of a voice and ask me to listen and subjectively judge whether it sounds like your voice, I think that is still rather arbitrary." Therefore Deng Xin and others proposed a third view: a judicial appraisal needed to be initiated. Their consideration was that, faced with a new type of case, the team hoped the judgment could be verified more objectively. But this choice also brought real costs. The judicial appraisal for this project was expensive and required the plaintiff to advance the fees; the plaintiff in this case did not live in Shanghai, and coming to Shanghai for the judicial appraisal would also mean spending time and incurring travel expenses. If the appraisal result favored the plaintiff, these costs would ultimately be borne by the defendant. Initiating an appraisal meant that whichever side lost, its own losses would increase. Each view had its merits. The meeting moved to the summary stage, and ultimately more than two-thirds of the participants favored initiating an appraisal. Wang Lin at one point felt the procedure was too complicated—she had already had the matter notarized, and undergoing an appraisal would add to the burden of defending her rights—so she hesitated about whether to proceed. Only after the court further explained the relevant requirements did she agree. The court conducted two rounds of investigation concerning the appraisal, determining the scope of comparison, the test materials, and the samples. The synthesized audio at issue served as the test material, and Wang Lin personally went to the appraisal institution to record her voice as the comparison sample. Who should clarify the origin of this voice? The institution undertaking the appraisal had experience in traditional voice appraisal but had never handled a similar AI-synthesized voice issue. Because the original training input materials could not be obtained, the appraisal method was to conduct a voice analysis and comparison between the synthesized audio and Wang Lin's own voice. Unlike traditional appraisals, the appraisal institution did not issue a conclusion on "whether the AI voice came from Wang Lin or not," but instead provided a series of data: among 28 formant acoustic indicators, 24 of the case audio's deviations from Wang Lin's own audio were less than 10%, of which 16 were less than 5.36%, and the portions that were relatively similar and highly similar reached 90%. These indicators reflect vocal tract resonance and timbre characteristics; the smaller the deviation, the more it shows that the AI audio is close to the person's own voice in timbre. Deng Xin and the team had to understand the meaning of these parameters before making a legal judgment. To this end, the judges consulted materials and communicated with the appraisers, and the appraisers later also appeared in court to be questioned. In March 2026, the second professional judges' meeting before the first-instance hearing was convened, mainly to discuss the direction of the trial and the results of the damage. This included the allocation of the burden of proof. When the voice generation process is hidden in the training data and back end controlled by the company, to what extent must the infringed party prove its case? Because of the report results, the disagreements at the professional judges' meeting were relatively smaller. Deng Xin explained that the plaintiff first proves that there is a reasonable possibility that her voice was accessed, and that the synthesized voice is highly similar to her own; only after that does the party holding the key materials have an obligation to explain how this voice came about. Deng Xin told reporters that when the court considers whether the plaintiff has fulfilled the "initial burden of proof," it mainly looks at "access + substantial similarity." "Access + substantial similarity" is a principle in copyright and other intellectual property disputes: the rights holder need only prove that the infringer actually had access to the prior work, or that there was a reasonable possibility of access, and that the two works are substantially similar; where the infringer cannot rebut this, infringement may be established. This case drew on that principle—Wang Lin's voice works were publicly released online, and the company had a reasonable possibility of obtaining these works; the voice comparison and other materials also supported the connection between the synthesized voice and her. On this basis, after the infringed party completes the initial burden of proof, the corresponding burden of proof shifts to the infringer. The infringer should produce rebuttal evidence that its generated audio did not access or use Wang Lin's voice as material; otherwise, it will bear the adverse consequences of failing to meet the burden of proof. After the second-instance judgment was concluded, some interpretations appeared online claiming that this case involved a "reversal of the burden of proof" and that it was "a major breakthrough in the adjudication of such cases." Wu Dan, the judge handling the second instance and a judge in the Case Filing Division of the Shanghai No. 1 Intermediate People's Court, explained that this was not a reversal of the burden of proof. In fact, the basic principle that "the party who asserts must prove" has not changed; what changed is that when an individual has produced the evidence they are able to obtain, a company cannot simply end its explanation of the facts with the single statement that "the employee has left and it cannot be confirmed." The technical process may be complex, but the responsible party cannot disappear because of it. AI concept illustration of the courtroom scene. During the trial, the company did not complete the corresponding burden of proof. Its response remained that the relevant business and technical personnel had left and the generation process could not be reconstructed. The first-instance court accordingly found infringement and issued its judgment in April 2026. From receipt of the case to the first-instance judgment, about 13 months elapsed. The unrecognized 300,000 yuan. After the first instance, both parties appealed. Wang Lin believed the compensation was insufficient and did not accept the 50,000 yuan award, while the company did not accept the finding of infringement or the liability to compensate. On September 7, 2026, the Supreme People's Court issued the Opinions on the Lawful Adjudication of Cases Involving Artificial Intelligence, Article 4 of which expressly provides: unless otherwise provided by law, where a natural person's voice is used as training corpus without that person's consent to imitate the timbre, intonation, pronunciation style, and the like of that natural person and generate a synthesized voice capable of identifying that natural person, and that natural person claims that the actor infringed their voice rights and interests, the people's court shall support the claim in accordance with law. The first-instance team immediately communicated with the second-instance team. They judged that this case was very likely the first case of its kind to enter second instance after the issuance of the Opinions, and that the adjudication approach was consistent with the guiding philosophy of the Opinions. Deng Xin felt that the earlier judgment of himself and his team had been confirmed. Wu Dan pointed out that the court would not, merely because of this case, require the company to respond regarding the source of the materials and thereby indefinitely expand its review obligations over massive general data. However, where the materials came from, what use the authorization permitted, and how the model generated its output should become a full-process record that can be retained and verified. Deng Xin reminded rights holders that which rights have been harmed and how much actual loss has been incurred are the evidence that most needs to be fixed as early as possible at the outset of rights protection; and that when enterprises receive relevant infringement complaints, they should promptly conduct self-inspection and take necessary measures such as deletion and disabling links. Repeatedly denying everything and pushing disputes into costly appraisal and litigation will often only increase their own losses. In fact, before the judgment, the issue of compensation in the case had already entered discussion. At the second professional judges' meeting before the first-instance hearing, someone suggested that if the plaintiff could prove that her voice could have been accessed by the platform, while the defendant had consistently been unable to explain the source of the materials, then infringement compensation would most likely be unavoidable. Professional judges' meetings are not easy to convene, so the discussion did not stop at "whether to compensate" but proceeded directly to "how much to compensate." However, there was not much material truly available for the meeting's reference. The plaintiff submitted her own remuneration income, and the defendant produced market fee standards. At the meeting, someone pointed out that looking at these two alone was not enough; it was also necessary to consider the duration of continued use of the audio, whether the activity was national or regional, and exactly how broad the dissemination scope was, and then to consider these together with industry fee standards. Wu Dan told reporters that the establishment of voice infringement does not mean that the plaintiff's claim of 300,000 yuan in compensation will be supported. For voice actors who rely on their voices as a professional resource, licensing fees for similar uses are an important reference for determining loss. If a licensing price for a similar use could be provided, the court would have a more direct reference when determining loss; however, in both the first and second instances, Wang Lin failed to provide an admissible licensing price for a similar use, and also lacked sufficient evidence to prove the claimed actual loss, so the 300,000 yuan could not be established on estimation alone. Protection of voice rights and interests is not premised on fame, but unauthorized use of a voice differs in degree of harm from borrowing someone's influence to attract users, and the amount of compensation cannot be determined mechanically. When actual loss and infringement profit are difficult to calculate precisely, the court will comprehensively consider the duration of the infringement, the scope of dissemination, the manner of use, the degree of fault, the consequences of the harm, and the contribution of the voice to commercial revenue. Wu Dan explained that these audio clips were used for process prompts in user acquisition activities, such as repeated announcements of "Congratulations, you have won" and notices telling users to proceed to the next stage. The company used the voice to complete activity announcements and did not make Wang Lin's identity or influence the core selling point; the court also took into account that the voice was online for about four months and that Wang Lin's account had limited followers and activity, and that her role in the activity was mainly auxiliary prompting, with a relatively low contribution to commercial revenue. The professional judges' meeting at first instance also discussed the scope of compensation. Deng Xin said that there is no ready-made formula for such new types of cases, and the team listed the factors to be considered one by one before making a comprehensive judgment; after review in the second instance, the court found the 50,000 yuan determined at first instance to be reasonable and upheld it. Wu Dan emphasized that 50,000 yuan is not a uniform price for similar cases in the future—if an enterprise uses a voice for a long time, fails to take it down even after a complaint, or makes someone's voice a traffic selling point, the considerations will differ; and if a rights holder can prove a licensing price for a similar use or provide fuller evidence of dissemination and harm, that will also affect the outcome. According to Deng Xin, in addition to the 50,000 yuan, the company must also bear the relevant necessary rights-protection costs supported by the court. The second instance ultimately rejected both parties' appeals and upheld the original judgment. The conclusion of the case also leaves lessons. Deng Xin judges that with the rapid development of AI, there will only be more and more copyright disputes arising from AI-generated voices, face-swapped avatars, and videos. "There needs to be a rule—how exactly should the burden of proof be allocated" is also one of the many legal questions that will need to be answered in future adjudication of AI-related cases. The Opinions of the Supreme People's Court on the Lawful Adjudication of Cases Involving Artificial Intelligence is the first judicial adjudication rules document on artificial intelligence issued by the highest judicial organ of the state, and is "Chinese judicial wisdom" contributed to promoting the upward and benevolent development of global artificial intelligence. Amid ever-changing technological transformation, countless judicial personnel are still needed to refine the rules stitch by stitch and apply "patches" in specific cases. In the Changning District Court, from the choice of whether to initiate appraisal, to the allocation of the burden of proof, and then to the final gavel on the discretionary award of 50,000 yuan in compensation... these are both the prudence required by the specific disputes at hand and the measure by which rules are gauged amid the wave of algorithms. And this measure must protect the rights of a specific person, and also enable artificial intelligence—this winged steed—to run fast and run steadily.