AI face change clone voice can tell? The Supreme Court just gave the answer.

On September 7, the Supreme People's Court issued a document, Fa Fa [2026] No. 10,"Opinions on Trial of Cases Involving Artificial Intelligence Disputes in accordance with the Law", with 24 articles in 5 parts.

This is the first time that the country's highest judicial authority has specifically given rules on how to decide AI disputes.

I read the full text of the 24 articles and picked out the six articles that are most related to ordinary people. There is good news, and there is also one thing that everyone may not like to hear.

Make the nature clear first to avoid misunderstanding: this is a judicial policy document, not a law. It governs how to apply existing laws when courts at all levels try cases, based on the Civil Code, Personal Information Protection Law, and Consumer Rights Protection Law. And it talks aboutcivil liability, not criminal conviction.

!

1. Your face, your voice, and your family who have left

The title of Article (4) is very straightforward: Regulate in accordance with the law the use of generative artificial intelligence to infringe on personal rights and interests such as "AI changing faces" and "AI resurrect the dead".

This article is divided into four layers, and I will explain it according to the original text.

Face and name. Without consent, you took your name and portrait to create a virtual image that could identify you, and also used it and made it public. You sued him for infringement of the right to name and portrait, and the court supported it.

voice. This layer is written in more detail than I expected:

uses the voice of a natural person as training material without the consent of the natural person to imitate the timbre, tone and pronunciation style of the natural person to generate a synthetic human voice that can recognize the natural person. If the natural person claims that the perpetrator infringes on his voice rights and interests, the people's court supports it in accordance with the law.

Pay attention to the five words"as training corpus". It doesn't just care about "you synthesized a piece of my voice", but it takes a step forward: the act of you using my voice as training data is itself within the scope of regulation.

Speak. If you manipulate the synthesized image or voice to do inappropriate behavior or make false remarks, you will lower the social evaluation of this person and consider it an infringement of reputation rights. In other words, changing your face is one thing, using this face to say bad things is another thing, and the latter also adds a layer of reputation.

Dead. If the virtual digital image of the deceased is made or used without authorization, resulting in the infringement of the deceased's name, portrait, and reputation,the deceased's close relatives may sue in accordance with Article 994 of the Civil Code.

In the past two years,"AI resurrects relatives" has provided a lot of services, some of which were done by family members themselves, and some were not. The line drawn here is: Family members have the final say.

2. The most practical one, many people have not noticed

Article (8) talks about the prohibition on infringement of personality rights.

I think this is the most useful article in the entire article for ordinary people, but it doesn't appear much in the discussion.

has evidence to prove that the perpetrator is using artificial intelligence to commit or is about to commit illegal acts that infringe on his personality rights and interests. Failure to stop it in time will cause irreparable damage to his legitimate rights and interests. It applies to the people's court to order the perpetrator to stop the relevant acts or If the relevant network service providers and generative artificial intelligence service providers stop providing relevant services, the people's court may issue a ban on infringement of personality rights in accordance with the law.

The key word is "in progress or soon to be implemented."

Put yourself in your shoes: what's your biggest rush when you find out your face has been made into a video and posted online? It's not about getting a few thousand dollars in compensation in three months, it's aboutstopping it now and stopping spreading it.

How long does it take to fight a lawsuit? By the time it's decided, things have already spread. There is no need to wait for a verdict to ban the ban. It is suppressed while fighting. Moreover, the application target is not only the person who publishes it, but can also directly request the platform and AI service provider to stop providing services.

!

3. Open box

Article (5) governs "Internet box opening" and "human flesh search".

There is a detail worth mentioning about the writing of this article. It identifies violations of privacy rights as:

For the purpose of spying on privacy, > uses artificial intelligence to track and analyzepublic informationsuch as telephone numbers, online accounts and social media of specific natural persons to obtain private information

The word used is "public information."

This explains something that was originally difficult to explain clearly. People who open boxes often have a saying: Every piece of information I use is public online, you sent it yourself, I didn't hack into your system, how can it be considered a violation of privacy?

The answer to this article is: A single disclosure does not mean that it is still public in summary. Using AI to track, analyze, and assemble public fragments scattered everywhere to piece together something you didn't intend to let others know. This assembly action itself is a violation of privacy.

4. Use prompt words to induce AI to do bad things. The responsibility lies with you

The first half of Article (7) talks about platform liability: If the content automatically generated by AI is infringed, and the platform fails to stop generation in time after the rights holder notifies it, the platform will bear the responsibility. The notice must contain preliminary evidence of infringement and your true identity information.

The second half of the paragraph is for users, so I will copy the original text:

QKPFX7 If a QK network user maliciously induces generative artificial intelligence to generate infringing content and causes damage to others by entering infringement prompts, the network user shall bear infringement liability in accordance with the law.

This sentence is for those of us who use AI every day.

You carefully construct prompt words to circumvent restrictions, induce the model to generate infringing content, and then cause damage, and the responsibility lies with you. It's not "it's generated by AI, it's none of my business", nor is it "the platform is not blocked, find the platform."

"How to circumvent the limitations of AI" has been a lively topic in recent years, and many people are sharing their skills. This rule is equivalent to saying: If something generated after circumvention causes trouble, the account will be blamed on the circumvention.

5. Kill familiar and bring goods with fake stars

Article 10: Two things.

Be familiar with big data. If the same thing is sold to different people at different prices, unreasonable differential treatment using algorithms will bear tort liability. What to look at when identifying? There is a very specific sentence in the original text:

QKPFX9 Does QK form individual transaction conditions based on consumers 'consumption preferences, willingness to pay, ability to pay, browsing history and other information

The words "willingness to pay" and "browsing history" are clearly written in.

Counterfeit celebrities bring goods. The original text of this paragraph is:

When > operators provide goods or services, they use artificial intelligence to implement "counterfeit celebrity bringing goods" and constitute fraud. If consumers claim punitive damages in accordance with the provisions of Article 55 of the Consumer Protection Law, the people's court will support it in accordance with the law.

Rule 55 is to refund one and compensate three, and if less than 500 is counted as 500. In other words, buying this kind of goods is not just a refund.

Note that it has a prerequisite,"and constitutes fraud." Not all goods that use the AI image are automatically applied, and they must first be able to commit fraud.

6. One thing that everyone may not like to hear

The first five directions are all about protecting individuals. Article (6) has the opposite direction and must be presented as it is:

is an artificial intelligence model training. If it handles personal information that an individual has disclosed on his own or other legal disclosure within a reasonable range, and the individual does not explicitly refuse,it is generally not considered an infringement of personal information rights.

What you have posted publicly is used to train models, and it is generally not infringement.

!

I understand the consideration of this article: training data is basically obtained from public data. If this article is reversed, the entire industry cannot carry out it.

But it left two cuts, both worth remembering.

One is that "the individual did not explicitly refuseit." Read it the other way around: If you explicitly refuse, the situation is different. Therefore, the switches hidden in the platform settings such as "Don't use my data to improve services" are not decorations.

The other is that "if it has a significant impact on personal rights and interests, personal consent should be obtained in accordance with the law." For the big impact part, I need your nod.

Among the factors to consider when determining the "reasonable range" is one that I think is well written: the context in which personal information is disclosed and the range of uses that can reasonably be expected. What you send in the community owner group is not the same as what you send on the public platform.

by the way correct a statement

The term "AI illusion infringement" has appeared a lot in discussions in recent days.

I flipped through the 24 articles, and there wasno separate article on "illusion" in the full text. AI generates inaccurate content that leads to infringement. In Article (7), the rule follows the path of "notifying and then taking necessary measures", which is the same logic as the handling of other online infringements.

This is not a big problem, but if you look according to "There is a rule specifically governing AI hallucinations", you won't be able to find it.

few sentences boundary

The "Opinions" are judicial policy documents that guide courts at all levels on how to review them, not the legal provisions themselves. The Civil Code and the Personal Information Protection Law are still cited in the real lawsuit.

It talks about civil liability, compensation, cessation of infringement, and apology. The criminal part of the document has other statements, which are outside the scope of this article.

Specific to the case how to judge, but also look at the evidence, look at the circumstances, look at the court determination. The document gives the rules, not the verdict.

The full text of 24 articles, I only picked here with ordinary people's daily most close to the six articles. The items of liability for autonomous driving accidents, AI product defects, liability exemptions for open source software, and intellectual property rights have not been expanded. If you want to read the full text, it can be found in the release column of the Supreme Court's official website.