Wu Yue targeted by AI face-swap livestream selling: account ban came after public outcry—why does governance always lag behind?

(The author of this article is Fan Zhongchen, Associate Professor at the School of Economics and Finance, Xi'an Jiaotong University)

In October 2026, actress Wu Yue posted on social media, calling out a video account by name: the account had used AI face-swapping and voice synthesis technology to impersonate her image and sell cosmetics via livestream. The account had also generated a video featuring "Deng Yaping" endorsing products on camera. After the incident came to light, the account in question changed its name overnight and was subsequently permanently banned by the platform.

The outcome looks satisfactory, but the sequence deserves attention: the ban came after the person herself spoke out and after media reports, not at the moment the infringement occurred. The same situation had already appeared two years earlier. In December 2024, an AI-synthesized "Zhang Wenhong" was hawking protein bars in a livestream late at night; the real person complained repeatedly with limited effect and could only lament that fake news "happens again and again, like a locust plague." Celebrities have studios, lawyers, and public attention, yet even they can only issue statements; ordinary people whose faces are stolen often do not even notice.

The cost of infringement is too low, and the cost of defending rights is too high

AI face theft for livestream selling persists despite repeated bans. Rather than talking idly about technology ethics, it is better to first do the cost accounting. Generating a convincing face-swapped video costs only a few dozen yuan and a few minutes; if an account is banned, it can simply be registered anew and continue operating, with negligible losses. On the rights-defense side, victims must first discover the infringement, then secure evidence, then complain to the platform; once the platform finishes handling it, the other party moves to another platform and continues broadcasting; and once litigation begins, it often takes more than half a year from filing to judgment. The gains are calculated by livestream sales revenue, the cost is only a few dozen yuan, yet defending rights takes months. As long as this pattern remains unchanged, "whack-a-mole" governance will not end.

Proving "I am me" is difficult in three respects.

First, the subject is hard to find: account real-name information is held by the platform, and individual inquiries and police retrieval must both follow procedures; by the time the identity is ascertained, the payments have often already been split up and transferred away;

Third, the chain is too long: livestreaming, clip editing, distribution, and payment collection belong to different entities, and each link can claim it "only provides technology" or "is merely a channel."

What needs to be clarified further is: what exactly do face thieves steal? Not simply a face, but the credit accumulated over decades behind that face. The value of Wu Yue's image stems from her works and reputation; AI has merely found a channel for rapid monetization of this slow variable—in essence, costless arbitrage of reputational capital.

Different characterization leads to completely different methods of calculating losses: if it is treated only as a personality rights dispute, compensation is calculated on moral damages and often settled at tens of thousands of yuan; if it is recognized that what was stolen is a monetizable credit asset, then livestream sales revenue is the proceeds of conversion, and the compensation base is entirely different. The same is true for legitimate businesses: they spend real money to hire endorsers and build reputation, while face-theft accounts borrow that credit at zero cost and seize orders—a classic case of bad money driving out good. Livestream e-commerce ultimately runs on trust, and AI face theft ultimately overdrafts the credit foundation of the entire industry.

Platforms must intercept in advance, not ban after the fact

To be fair, the platforms already possess all the means needed for interception.

At the registration stage, opening an account requires real-name verification and facial recognition, and risk-control systems are fully capable of identifying who is registering sales accounts in bulk. At the broadcasting stage, the Provisions on the Administration of Deep Synthesis of Internet Information Services, effective in 2023, already clearly stipulate: generating or editing faces or voices shall require separate individual consent; where deep synthesis service providers provide deep synthesis services that may cause public confusion or misidentification, they shall add prominent labels at reasonable positions and areas of the generated or edited information content.

Who actually operates a livestream room and whose image appears on camera—the platform knows this better than any other party. At the distribution stage, when a newly registered account suddenly broadcasts frequently using a celebrity's face and sales rise abnormally, identifying such an account is no harder than identifying fake orders.

At the monetization stage lies the most critical gate, and also currently the weakest one. Accounts can be registered repeatedly, but the payment recipient is hard to replace: livestream sales income must be settled through merchant accounts and payment channels, so as long as the on-camera person, the account, and the payment recipient are bound and verified together, a face-theft account cannot withdraw funds even if it goes live. For verified impersonation sales, merchant accounts should be frozen, payment blacklists applied, and settlement suspended—the mature experience of "delayed arrival of funds" in anti-fraud governance can be fully transplanted to livestream rooms.

But in reality the order of handling is often reversed: first a victim's complaint, then media reports, and only then a platform ban. This ban by the video account platform is not improper in itself, but judging by timing, it is closer to responding to public opinion than to routine governance.

The platform issued an announcement in March this year on strengthening governance of AI-generated content. Now that the system has been established, its enforcement effects should be subject to scrutiny: how much impersonating content was intercepted before broadcasting? Were the payment recipients behind banned accounts referred to regulatory authorities? Such data should be published regularly, rather than issuing a single "has been banned" notice only after an incident.

Where the benefits go is equally clear: for every completed livestream sales transaction, the platform takes a cut. Whoever benefits from it should bear the cost of governance. The current situation amounts to making the weakest party bear the gatekeeping responsibility on behalf of the strongest platform.

There are laws to follow; the key is enforcement

The claim that "there is no law to rely on" is not accurate. Article 1019 of the Civil Code already provides clearly: without the consent of the person holding the right to their portrait, no organization or individual may produce, use, or publicly disclose that person's portrait, and it explicitly prohibits infringing portrait rights by means such as vilification, defacement, or forgery using information technology; the Measures for Labeling AI-Generated Synthetic Content, which took effect on September 1 last year, go further: Article 4 provides that where the generative synthesis service provided by a service provider falls under the circumstances in Article 17, Paragraph 1 of the Provisions on the Administration of Deep Synthesis of Internet Information Services, a conspicuous prompt label shall be added in an appropriate position on the image; Article 5 provides that service providers shall, in accordance with Article 16 of the Provisions on the Administration of Deep Synthesis of Internet Information Services, add implicit labels in the file metadata of generated synthetic content; service providers are encouraged to add implicit labels in the form of digital watermarks and the like to generated synthetic content.

At the judicial level, precedents have also paved the way: as long as the face-swapped image can be identified as the person themselves, infringement is established, and whether profit is made is not a prerequisite; if a defendant argues that "AI randomly produced a lookalike face," the defendant should bear the burden of proof; income earned from livestream selling can be directly used as the basis for determining the amount of compensation. These rules resolve the "I am me" dilemma: the victim only needs to prove that the image in the video can be identified as themselves, and the remaining burden of explanation shifts to the infringing party.

What is truly lacking is deterrence. Labeling should not serve merely as a prompt but should serve as a gate: once a platform detects an unlabeled synthetic face going live, the measures it should take are stopping the broadcast, restricting traffic, and freezing withdrawals, rather than adding a corner label after the fact; those who maliciously delete hidden labels in files should be held liable in accordance with the law. Compensation standards likewise need to be raised. As mentioned above, what is stolen is a monetizable credit asset, not abstract personality harm; at present, the damages awarded in such cases are generally low and far removed from livestream selling revenue. Judgments should directly calculate on the basis of livestream sales turnover, and punitive damages should apply in serious cases, so that mass face theft becomes unprofitable.

Beyond individual litigation, there should also be collective remedies: AI face theft for livestream selling deceives not only the person whose face was stolen but also a large number of consumers who placed orders. The Shanghai Consumer Council has already determined in the Zhang Wenhong incident that such livestream selling constitutes fraud. Public interest litigation by consumer organizations and public interest litigation by procuratorial organs should both apply to such cases.

How ordinary people can protect themselves

Before interception mechanisms are truly in place, what ordinary people can do is mainly basic prevention: tighten the visibility and download permissions of social media accounts to reduce the leakage of high-definition frontal face material; verify more when video calls involve transfers; and after discovering face theft, promptly screen-record and preserve evidence, and complain and report to the platform and the cyberspace administration authorities. In judicial practice, there have already been precedents of ordinary people winning cases; the standard applied by the courts is the identifiability of the image, and there is no need to shrink back because of "how to prove it is you."

The goal of governance has never been to prohibit AI technology, but to establish rules: who uses each face and where it is used should all be traceable; once verified, a price must be paid. Only when interception precedes broadcasting, awards are sufficient to create deterrence, and ordinary people no longer need to prove "I am the real me" will this problem be truly resolved.

Exclusively first published on the Yicai Account of Yicai; this article represents only the author's views.