On September 29, 2026, the Shanghai First Intermediate People’s Court issued a final ruling addressing what is described as Shanghai’s first case concerning natural-person voice-rights protection arising from AI-synthesized speech. The case centers on the intersection of generative AI training practices and personality rights, and its principal contribution is an evidentiary framework for proving infringement when a plaintiff has no visibility into how a generative model was trained. Note though that this case has not been marked as typical or guiding at the time of writing and therefore may have minimal precedential value.
The plaintiff, Ms. Wang, is a real-name-verified voice actor who performs and publishes work on an app, earning income from that activity. In January 2025, she learned that Company A had used what appeared to be an AI-synthesized version of her voice in an internet promotional campaign. She had never worked with or authorized Company A. She notarized the disputed audio and filed suit seeking cessation of use and RMB 300,000 in damages.
Company A admitted the audio was generated by its own AI system but claimed it could not verify the specific training material or source because a former employee involved in that work had since left the company. It denied prior knowledge of Wang’s identity as a voice actor and argued there was no direct evidence that its training data included her voice.
To meet her burden, Wang submitted a forensic acoustic comparison between the notarized disputed audio and a live-sampled reference of her own voice, analyzed across 28 formant acoustic indicators. The comparison found 24 of 28 indicators had deviation under 10%, 16 of those under 5.36%, and an overall similarity of 90% between the two voices.
The first-instance court found that Company A’s conduct constituted AI-generated voice infringement causing economic loss and awarded RMB 50,000. Both sides appealed; the Shanghai First Intermediate Court affirmed, dismissing both appeals.
The appellate court’s reasoning rests on a core substantive rule directly addressed to generative AI: using a natural person’s voice as training corpus, without consent, to generate a synthetic voice that mimics that person’s timbre, intonation, and pronunciation style – such that the output is recognizable as that person – constitutes infringement of voice rights, regardless of the algorithmic method used to produce the output. The court was explicit that the legal question is not how the training or generation process works technically, but whether the personality interest embodied in the voice has been appropriated in a recognizable way.
The court’s more significant holding concerns how this is to be proven, given the structural information asymmetry inherent in generative AI. The plaintiff’s burden is met by showing (a) a plausible channel by which the defendant could have obtained her voice for training – here, her work was publicly posted and thus accessible as potential training data – and (b) that the resulting synthetic voice is highly similar to her own such that ordinary members of the public would associate the two. Once that showing is made, the burden shifts to the defendant, as the party with exclusive control over training data, model logs, and data-sourcing records, to prove either the lawful origin of its training material or that it did not in fact use the plaintiff’s voice characteristics. Company A’s claim that records were unavailable due to staff turnover did not satisfy this burden, and it bore the adverse consequences of that failure of proof.
The court framed this explicitly as a shift in analytical focus for generative AI personality-rights cases: away from requiring proof of what specific data trained the model (something plaintiffs structurally cannot access) and toward whether the output is recognizable as the plaintiff (something that can be established through forensic comparison of the generated content itself). This reframing is presented as a response to the “black box” problem of generative AI – where the training corpus, weights, and generation logic are opaque to anyone outside the entity operating the model.
On damages, the court applied a two-step approach relevant to voice used as training data and later monetized: first, look to a comparable licensing fee for use of the plaintiff’s voice, if provable; absent that, weigh factors including duration of the infringing use, scope of dissemination, manner of use, degree of fault, resulting harm, and the proportional contribution of the personality element to commercial value. Here, the synthetic voice was used briefly as a process/notification prompt rather than as a core commercial asset, and Wang’s own public profile and market activity were modest, which the court treated as limiting the independent commercial value of her voice for damages purposes. The RMB 50,000 award was upheld on that basis.
For comparison, the Beijing Internet Court’s April 2024 decision – reported as the first case nationwide on AI-generated voice personality-rights infringement – had already established that voice rights under Civil Code Article 1023 extend to AI-generated voice where the output is recognizable, and that authorization to use a sound recording (e.g., under copyright) does not itself confer authorization to use the underlying voice as material for AI training or synthesis. Beijing’s facts involved a traceable commercial chain from original recording to licensed data transfer to AI product, so the case did not need to resolve how a plaintiff proves unauthorized training use when that chain is not documented or disclosed. Shanghai’s ruling takes up that gap directly, and expressly ties its reasoning to Article 4 of the Supreme People’s Court’s September 2026 Opinions on AI-related disputes, suggesting this burden-shifting approach to training-data opacity is intended to have broader application in future generative AI personality-rights cases.
The text of the original announcement from the Shanghai High Court is available here (Chinese only).
