China’s top court just put civil teeth on deepfakes. On September 7, 2026, the Supreme People’s Court put out its first guide for AI fights in court. The 24-article text treats cloned faces and voices as rights you can own and defend. It also puts notice-style duties on generative services. U.S. platforms will feel that civil liability model even though federal U.S. law covers far less ground. Chinese civil courts now share one script for cases that used to arrive as one-off fights.
The timing matters. China still has no dedicated national AI statute, per reports that cite People’s Daily. Judges instead use the Civil Code, the Cybersecurity Law, the Copyright Law, and the Personal Information Protection Law through this new guide. The result is not a full code. It is a playbook. Playbooks move faster than statutes when platforms already work across borders.
What the 24-Article Deepfakes Guide Covers
China Daily says the guide answers years of public anger over AI face-swapping and voice cloning. Doxing made the list too. Global Times also lists AI hallucinations, big-data price bias, self-driving cases, and model training. The court cast the release as both rights protection and orderly industry growth. Zhou Jiahai, head of the SPC research office, said the aim is a balance between growth and safety.
That balance talk is familiar. The hard parts are clearer. The guide tells lower courts how to spot personality-rights harm when the tool is generative AI. It also tells them when providers and users share fault. For firms that ship face and voice models into China, or host Chinese users, those lines are the product brief.
Guideline Snapshot
Issuer: Supreme People’s Court (SPC)
Date: September 7, 2026
Form: 24-article judicial guide on AI disputes
Core rights: Name, likeness, voice, privacy, reputation
Provider duty: Act after notice of infringing AI output
Fast relief: Court orders when serious harm is hard to undo
Faces and Voices Are Rights You Can Sue Over
The rule on likeness is blunt. If you use AI on a person’s name or likeness without permission, then make and spread a clear digital avatar, you can violate name and likeness rights. The rule on voice is just as direct. If you use someone’s voice without consent to make fake speech that copies tone and style, you can infringe voice rights. Both lines treat a person’s look and sound as something you cannot take for free.
Si Yanli, deputy head of the SPC research office, said AI face-swapping makes ordinary people potential deepfake targets. Face-swapping can pull a likeness from almost anywhere. Voice cloning can steal a voice at low cost with results that sound real. That mass risk is why the court wanted one national standard rather than scattered local rulings.
Chinese courts already had a preview case. In 2024, the Beijing Internet Court found that AI voice copy without consent infringed personality rights. A voice artist’s audiobook tracks had been used, without her consent, to build a commercial text-to-speech product. The court awarded damages. The new guide turns that logic into a countrywide template.
China Daily Hong Kong repeats the same rights frame. Global Times adds that close kin of a dead person may seek civil liability when the dead person’s rights are hit. That clause matters for estates and for memorial deepfakes that bring back a dead person’s voice. It also shows how far the court will stretch civil cover around identity.
AI face-swapping technology can indiscriminately harvest anyone’s likeness, making ordinary people potential targets of deepfakes.
Si Yanli, deputy head of the SPC research office, quoted by China Daily
Deepfakes Plus False Claims Raise Reputation Risk
Identity theft is only half the story. The guide also reaches reputation. Misuse of AI-made avatars or voices to spread false claims or harm reputation can trigger liability. The Star, carrying a Xinhua report, stresses how little raw material it takes. A few photos or a short clip can build a strong replica. Once that replica carries a false claim, the harm is both likeness and reputation at once.
That double hit is why U.S. platforms should care even though Congress has acted only on intimate deepfakes. A Chinese plaintiff can sue under Chinese civil rules when the service or the harm ties to China. Global apps already moderate for local law. A clearer Chinese liability map raises the cost of slow takedowns on deepfake impersonation.
Users are not off the hook. A person who uses prompts on purpose to push an AI system to make infringing content, and then causes harm, can be held liable. The court draws a line between curiosity and intentional abuse. Intent still matters. So does the output that leaves the chat window.
Notice Duties Push Generative Services Into a Takedown Frame
The provider piece is the part that travels. AI service providers can be held liable if they are told their systems made content that infringes a person’s rights and then fail to act in time. That is a notice-and-act model, not a full safe harbor. It also stops short of strict liability for every hallucination.
Hallucinations sit inside the same package. Generative systems can invent false or wrong claims. When those claims injure rights, and notice arrives, silence becomes a legal choice. Product teams that treat deepfake complaints as niche safety tickets now face a clearer civil price for delay.
Zhou Jiahai also spoke to consumer protection. Firms that use algorithms to offer different prices or terms for the same goods without a fair reason may face liability if consumers are harmed. He said the law must step in because not every buyer can spot a trick. Price steering is not a deepfake. It shows the court’s wider view: opaque AI that hurts people should meet civil duties, not only PR lines.
We cannot expect every consumer to become an expert at spotting deception. The law must step in promptly to protect consumers’ legitimate rights and interests.
Zhou Jiahai, head of the SPC research office, quoted via Xinhua in The Star
Doxing and Court Orders Widen the Fixes
Privacy gets its own AI chapter. The guide targets AI tools that track and parse public data on people, such as phone numbers and social profiles. If a user pulls private facts from that data and leaks them, that can infringe the right to privacy. Si called doxing a major form of cyberbullying that disturbs daily life and erodes safety online. Scraping that looks “public” is not a free pass once the pipeline is AI-aided and the output is private harm.
Speed matters when harm is hard to unwind. The Star report describes court orders for serious injury that is hard to reverse. One example is a digitally altered face used to spread false and defamatory sexual claims. Waiting for a full trial can leave damage that money cannot fix. Chinese civil courts can be asked to freeze the spread first. That fix is as important as damages for deepfake victims.
Tao Kaiyuan, vice-president of the SPC, said some issues stay open on purpose. AI moves fast. The court wants clearer rules later as experience builds. Open ends are not loopholes by design. They are an admission that judges will keep writing the next chapters case by case.
Why State Media Sources Cut Both Ways
Readers outside China should flag the sourcing on these deepfake rules. The public English record on day one runs heavily through state outlets such as China Daily and Global Times. Those outlets are state-linked. They can report the official frame with high fidelity and still soft-pedal enforcement risks. The smart move is to treat the quotes as real court messaging, not as field audits of how every lower court will act.
That caveat does not erase the text. A 24-article guide from the Supreme People’s Court is still a governing signal for Chinese civil suits. Firms that ignore it because the first stories came via state media mix up the messenger with legal force. The safer read is dual. Trust that the rules exist. Keep doubt about selective zeal.
The Speech and Innovation Chill Case
The strongest pushback is chill. Rights that cover likeness and voice can become a weapon against satire, parody, research demos, and stress tests. A notice duty can push platforms to over-remove. Startups with thin legal teams may ship weaker tools rather than risk a Chinese civil fight. Innovation slowdowns are a real cost when the field is still young.
SPC officials try to meet that critique early. Zhou spoke of balancing rights protection with healthy AI growth. Si added that tolerance should not mean a free pass, and prudence should not mean looking away. The official story is measured toughness. Critics will ask whether “careful” handling of new acts protects builders when a powerful plaintiff complains.
Selective enforcement is the harder critique. Guides can be applied firmly against commercial deepfake scams and lightly against favored actors. Civil process still sits inside a broader state system. Foreign firms may face sharper discovery and faster orders than local champions. None of that is proven by the launch-day coverage. It is the risk premium that compliance teams will price into China exposure anyway.
What U.S. Platforms Feel Without a Matching U.S. Law
Washington does have one federal deepfakes law. The TAKE IT DOWN Act, signed in May 2025, makes platforms pull nonconsensual intimate images, AI fakes included, within 48 hours of a valid request. Those platform duties took effect in May 2026. Still, no federal statute covers cloned faces and voices this broadly. State laws and platform rules fill gaps in uneven ways. China’s move does not rewrite U.S. code. It still changes the work map. Global safety queues already route China-related identity complaints into local-law workflows. A clearer SPC map makes those tickets harder to park.
Contract and insurance language will follow. Buyers who sell into China will ask vendors about notice response times for likeness and voice claims. Model hosts, for their part, will revisit prompt-abuse logs and consent capture. Those are civil-process shifts, not speech codes copied from Beijing.
Cross-border spillover also runs through people. A U.S. creator whose likeness is cloned for a Chinese audience may find Chinese civil courts more ready than a U.S. venue that still debates first principles. A Chinese user harmed by a U.S.-built tool with China service can pull the provider into the same frame if jurisdiction attaches. The guide is domestic. The product stacks are not.
Closing the Liability Frame
China’s Supreme People’s Court did not wait for a dedicated AI law. It wrote a 24-article civil guide that treats faces and voices as rights you can defend against deepfakes. The guide also extends privacy rules to AI-aided doxing and attaches notice-style duties to generative services. Court orders for serious harm complete the toolkit. The venue is Chinese civil courts. The market effect is wider.
State media carried the launch. That is a reason to read with care, not a reason to look away. Speech chill and selective enforcement remain live risks. So does the opposite risk: platforms that treat cloned identity as a side quest until a court order arrives. The SPC just raised the price of that delay. U.S. firms that sell generative faces and voices into a global stack now have a clearer civil map to ignore at their own cost.
