China’s top court sets liability rules for AI deepfakes and unfair pricing

China’s Supreme People’s Court has moved to set proper guidelines on who faces legal liability for AI deepfakes and voice cloning in a September 7 attempt to install guardrails around AI-generated content.

Artificial intelligence technology has developed to the point where anyone can generate convincing digital doubles by simply scraping a few photos or short recordings from the internet or social media and propagating that content for financial gain or to drive narratives.

China draws the line around consent 

According to China’s Supreme People’s Court, the line separating AI content based on a person’s replica from legal to illegal is consent. You infringe on someone’s rights when you create or share an AI recreation of someone, their face or voice, without that person agreeing to it. 

The level of infringement grows when that AI replica is used to spread false claims or run false narratives. 

When the intent of the fake content becomes malicious to the extent of false and defamatory sexual claims, the victim can fast-track action by asking for an injunction instead of waiting for a long, drawn-out court process.

The AI providers that people use to generate such content also face liability, especially if they have received notice of potential rights violations due to their platforms’ capabilities. 

Liabilities still apply if the AI model was spreading false or inaccurate information due to AI hallucinations.

Notably, per local reports, actors such as Joey Wong, Lawrence Ng, and even ordinary people have now begun to issue likeness rights to AI platforms, as part of a growing market to formalize AI replication. More than 95% of the 128,000 microdramas released in China in the first quarter of 2026 reportedly used AI during production. 

Who else faces liability for misusing AI? 

The court also mentioned fairness in business dealings as part of the same guidelines. For example, a business cannot use algorithms to charge different customers at different rates without any reasonable basis for the price discrepancy. 

Any firms caught in the practice will now face liabilities for the harm they cause to the affected buyers.

“We cannot expect every consumer to become an expert at spotting deception,” Zhou Jiahai, who heads the SPC’s research office, said in the Xinhua report. “The law must step in promptly to protect consumers’ legitimate rights and interests.”

How did China handle AI deepfake cases? 

Before today’s guidance from the apex court, lower court judges had to issue discretionary rulings on a case-by-case basis. 

The Beijing Internet Court issued the city’s first AI deepfake ruling in June 2024 when it fined an app operator 3,500 yuan (about $482) for allowing users to use swap templates to generate content of two models without obtaining due consent. 

Per the National Business Daily, the Guangzhou Internet Court has ruled on about 700 AI face-theft cases in just three years. 

A regulatory push on several fronts

The SPC guidelines are one piece of a wider clampdown. On September 2, the Cyberspace Administration of China said it had removed more than 5.61 million pieces of harmful or illegal content and roughly 49,000 accounts across 2,400 sites and apps, including Douyin, Kuaishou, RedNote and WeChat, in a drive against AI “slop,” fake news and impersonation, the South China Morning Post reported.

ByteDance has also taken down over 85,000 videos involving unauthorized AI reproductions of people’s faces and voices since January.

Earlier campaigns targeted the same problem in commerce. In November 2025, Xinhua reported that authorities scrubbed over 8,700 non-compliant items and dealt with more than 11,000 accounts that used AI to mimic celebrities in shopping live streams, including forged clones of actress Wen Zhengrong and Olympic champions pushing everyday goods.

Since September 1, 2025, China’s content-labeling rules have required AI-generated material to be marked, though enforcement has lagged as violators hide or strip the identifiers.