Hand a recording to a tool and, before long, it can produce the voice of someone you know, saying things that person never said. Plenty of people only realize after scrolling past such a clip that a voice can be copied at all. The law draws a line here, and the key is not how convincing the technology is, but whether the person himself agreed.
Why a Voice Is Protected
The Civil Code does not set out a separate right called a voice right; it protects the voice by borrowing another route. The second paragraph of Article 1023 states that protection of a natural person's voice refers to the relevant provisions on portrait rights. However a face is protected, a voice follows the same logic: making, using and publishing it all have to clear the person himself first.
On September 7, 2026, the Supreme People's Court issued the Opinion on Properly Trying AI-Related Dispute Cases, made up of five parts and 24 articles in all. The document gathers scattered questions in one place, including what a person can actually claim when their voice is used for training and then synthesized.
The Opinion states that, unless otherwise provided by law, where a natural person's voice is used as training material without that person's consent to imitate timbre, tone and pronunciation style and generate a synthetic voice that can identify that person, and that person claims infringement of voice rights, the people's court shall support the claim in accordance with the law.
Training material is, plainly, the voice samples fed to a model to learn from, and a synthetic voice is new speech the model generates from those samples. By naming timbre, tone and pronunciation style one by one, the document signals that what is protected is the voice itself as an auditory identifier, not the content of any particular sentence.
Whether You Can Tell Who It Is
Whether a line has been crossed hinges on whether you can tell who it is. According to a report by Jinan Daily · Aijinan, Li Fumin, a lawyer at Shandong Longquan Law Firm, explains that as long as an ordinary listener can match the voice to a specific person, that synthetic speech falls within the protected scope; it does not have to match the original word for word.
Which means the threshold is not a percentage of resemblance but whether the listener can place the voice. Timbre, speaking pace and habits combine to let an acquaintance recognize who it is, and that can be enough for identifiability; small technical differences do not change that judgment.
Apart from fair-use circumstances otherwise provided by law, even copying, editing or forwarding without making money can still amount to infringement. Many assume only commercial use matters, but the decision belongs to the person at the outset; using the voice first has already skipped that step.
One more point is often overlooked: what is protected is the natural person, not only the voices of celebrities. Ordinary people's reading aloud, voice messages and livestream audio fall inside the same line, and a casual remark sent out one day can become an identifiable sample in someone else's hands.
Series, Ads and Livestream Rooms
Beyond the voice, the Opinion also addresses names and portraits. Unless otherwise provided by law, where a person's name, portrait and similar data are processed without consent to generate an identifiable virtual digital image of that person and it is used or published, and the person claims infringement of name rights or portrait rights, the court shall support the claim in accordance with the law.
In concrete terms, that means series, short videos, advertisements and livestream sales. Bringing an unauthorized real person's image or voice into commercial use can touch both portrait rights and voice rights, and whether money changed hands or how many views it drew does not alter the premise that consent was missing.
The Opinion further states that where a generated, identifiable virtual digital image or voice of a specific natural person is manipulated to carry out improper acts or make false statements, lowering that person's or another's social standing, the people's court shall determine that this constitutes infringement of reputation rights.
For anyone making content, it comes down to one sentence: before using a real person's image and voice, obtain authorization, and make it specific and clear. Material found online or a recording saved in passing is not the same as consent to synthesis, still less consent to commercial use.
A Deceased Person's Voice Has a Keeper Too
The Opinion also addresses the deceased specifically. Where AI is used without authorization to make or use a deceased person's virtual digital image, infringing that person's name, portrait or reputation, and close relatives ask the actor to bear civil liability under Article 994 of the Civil Code, the people's court shall support the claim in accordance with the law.
According to a report by Jinan Daily · Aijinan, the lawyer Li Fumin explains that a deceased person's voice is not ownerless material either. Using AI to reproduce a deceased person's voice for commercial promotion or malicious hype without the close relatives' permission runs against public order and good morals, and the relatives can demand that the infringement stop.
In other words, the person is gone but the layer of personality interests around image and voice remains, and the right to claim passes to close relatives. So-called applications that bring the deceased back have the same two boundaries: whether permission was obtained and whether public order and good morals were crossed.
That is also a reminder to ordinary families: a recording or a video clip left behind by an elderly relative is not theirs for anyone to synthesize. What technology can do does not mean the formalities can be skipped, and in commercial settings it is worth asking the family first.
What makes it harder is how fast such content spreads; by the time a lawsuit runs its course, the clip may already be everywhere. The document therefore offers another route, letting rights holders act before a judgment arrives: hold down what is happening first, then argue the merits later.
Pressing Pause First
The Opinion states that where evidence shows a person is using AI to carry out, or is about to carry out, an unlawful act infringing personality rights, and failure to stop it promptly would cause irreparable harm to lawful rights and interests, an injunction against infringement of personality rights may be sought from the people's court.
The court may order the relevant acts stopped, or order the relevant network service provider or generative AI service provider to stop providing the relevant service. In issuing the injunction it also weighs the type of right infringed, the manner of the act and the likely scope of harm, and does not go beyond what is necessary.
The lawyer also flags several easy assumptions. According to a report by Jinan Daily · Aijinan, audio posted publicly does not mean it can be cloned and used for training directly, and verbal agreement does not mean unlimited authorization; for commercial use it is best to put the scope of permission in writing.
The lawyer also noted that a label added to generated content is only a tracing tool and cannot replace the person's authorization. It can show where something came from, but it cannot stand in for consent, and tucking a notice into a corner is not the same as obtaining permission to use a voice.
Technology can reproduce a timbre; it cannot reproduce the step of the person agreeing. Next time a familiar voice asks for money or a verification code, it is worth slowing down and asking who is actually speaking. That invisible ID card stays in the person's own hands.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.