100 Yuan for 30 Seconds of a Face You Don't Own

An actress woke up to a video of herself recommending a collagen serum she had never used. The lip-sync matched. The voice was close enough that people who knew her work didn't question it. In the corner, in small type, sat a line of text: this content contains AI-generated material.

The account was gone by October 19, permanently banned. Before it went, the serum had sold more than 7,000 units in thirty days, roughly 500,000 yuan in revenue. The actress, Wu Yue, went public on Weibo asking how many people had been cheated.

The label did not save anyone. It was there, legible if you looked, and 7,000 people bought a skincare product anyway. Sit with that for a second.

A labeled video still worked

Three things had to be true for this to work, and only one of them is about technology.

The face had to be available. Wu Yue's public footage is all over the internet — interviews, red carpets, drama clips. Twenty minutes of it is enough to build a likeness that survives a nine-by-sixteen crop on a phone screen.

The audio had to be close. Voice cloning from a few minutes of clean speech is a commodity now, not a research project. The account paired it with lip-sync adjustment and rewritten lines, which is the cheapest possible production stack: no actor, no set, no script supervisor.

And the label had to be ignorable. It was. A line of 8-point text in the corner of a vertical video, on a platform where the interface already covers the bottom third with captions and buttons, is functionally invisible. The label satisfied a disclosure requirement and changed nobody's behavior.

The supply chain has a price list

Reporters who went looking for the source found something more organized than a few opportunistic scammers. Shops on mainstream e-commerce and secondhand platforms openly advertise AI video customization, quoting from a few hundred to over a thousand yuan. One quoted 100 yuan for a thirty-second celebrity endorsement video and described growing a 10,000-follower account in a week.

Sellers were inconsistent about legality in a way that's revealing. Some refused celebrity work outright, citing infringement, and offered generic digital humans instead. Others said the technical side was trivial and that the hard part was the copy and the plot. Asked about rights, one didn't answer the question and said to just label it as AI content.

There's the business model, in one sentence. The label was understood, correctly, as a compliance cost rather than a legal defense. Which means the enforcement had to happen at the account level, after the sales, which is exactly what happened, and exactly why it will happen again next month with a different face.

What the law actually says

Chinese lawyers were blunt about the labeling question. A label answers "was this made by AI." It does not answer "was this authorized" or "is this true." Under the Civil Code, a natural person's portrait right is protected, and voice is protected by reference to it. Using someone's likeness and voice without consent to sell something is textbook commercial infringement, and adding "AI-generated" to the corner doesn't create an exemption.

The courts have already built the rule. In a case decided under the Supreme Court's guidance, a company took footage of a person's public lecture, paired it with a synthesized voice that closely matched her vocal characteristics, and made an endorsement video. The court found infringement of both portrait and voice rights. The company's failure to check whether the material was authorized counted as negligence, making it a joint infringer. Damages and costs came to 120,000 yuan, plus an apology.

On September 7, the Supreme Court issued its Opinions on the Trial of AI-Related Disputes, which specifically address face-swapping, voice cloning, and AI recreations of deceased people. The direction is consistent: consent is the axis, not disclosure.

The gap is enforcement speed, and it's structural. A video can be produced in an afternoon, sell for two weeks, and be deleted after a complaint. The revenue is collected before anyone with standing finds out.

This is not a fringe case

Wu Yue's story is the one that broke through, but the pattern is wider. Another actress, Wen Zhengrong, found an AI livestream using her likeness, confronted the operator, and got blocked. Kenneth Fok said his voice and appearance had been extracted into short videos, and that his own colleagues watched for a while before realizing. Jin Dong's likeness has been used to target older audiences, who are the least equipped to check.

The common thread is that none of these targets are obscure. They're recognizable enough to sell something and too distributed to police every clip. Fame used to be a defense — a well-known face is hard to fake. Now it's a supply of training material.

There's a version of this that will get worse before it gets better: the same tools that make a convincing celebrity ad make a convincing endorsement from a mid-tier influencer, a local doctor, or a school principal. Those people don't have press access or legal teams, and their audiences are smaller and therefore more trusting.

What would actually help

Four things, none of them novel, all of them unglamorous.

Make the label part of the image rather than the post. A visible mark burned into the frame survives screenshots, re-uploads, and the compression chain. A caption does not.

Require identity verification at the point of monetization, not at the point of posting. The bottleneck that matters isn't uploading a deepfake, it's attaching a payment method and receiving the proceeds. That's a much smaller surface to police, and it's where the 500,000 yuan changed hands.

Give the person whose face is used a fast lane. Not a general report form that routes into a queue, but a verified channel where a public figure can flag a synthetic endorsement and have it frozen within hours. Some platforms have pieces of this. Few have it fast enough to matter during a two-week sales window.

And stop treating the label as the answer. Disclosure is a courtesy to the viewer. Consent is the actual rule, and no amount of corner text substitutes for it.

The uncomfortable part

Seven thousand people bought a serum because a face they recognized said it was good. The economics of that are not subtle: a convincing fake of a trusted face outperforms most legitimate advertising, and it costs 100 yuan to make.

Which means the question isn't whether platforms can catch every synthetic endorsement. They can't, and the labs have said as much. The question is whether the cost of getting caught rises faster than the cost of making one. Right now it doesn't, and the label sitting in the corner is a pretty good measure of how seriously anyone expected it to.