← Back to blog
AiAbout 6 min read

After 47,000 posts and 3961 accounts, AI impersonation needs a gate at publish time

Published Nov 8, 2026
After 47,000 posts and 3961 accounts, AI impersonation needs a gate at publish time

On November 6, the cyberspace administration published the results of a concentrated campaign: more than 47,000 pieces of illegal content removed, 3961 accounts handled. The notice was specific about what it was after, namely "the use of AI to impersonate well known public figures for fake product pitches, follower farming and misleading users".

The notice named four kinds of cases. Account names were masked, but the platform spread is clear. On WeChat, "花\*\*记", "时\*\*叙" and "秀\*\*8"; on Douyin, "微\*\*天" and "S\*\*物"; on Xiaohongshu, "小\*\*C" and "瓦\*\*款". What they did was copy celebrity portraits without authorization, clone celebrity voices, or lift photos of public figures' relatives to promote goods. WeChat accounts "山\*\*话" and "旧\*\*故" used AI effects to invent a rumor that a certain public figure had Alzheimer's. Douyin's "才\*\*河", Xiaohongshu's "秋\*", Bilibili's "老\*\*迷" and Kuaishou's "云\*" impersonated the person directly, posting opinions, life stories and emotional scripts in the first person to pull people toward following them. Kuaishou's "康\*" generated vulgar dance and intimate interaction footage, farming followers that way and then monetizing them.

The commercial motive behind these four groups varies a lot. The second and third want traffic; the first and fourth end in a transaction. And one phrase in the first group deserves to be pulled out on its own: lifting portrait photos and image variants of public figures' relatives. That route sidesteps the problem of scarce material on the celebrity himself, since one family photo plus a face swap is enough to assemble a product video that looks credible.

The penalty lands on the account, the problem sits before publishing

The notice says "handled accounts", not "deleted a video". That difference is not small. A single post can be reuploaded once deleted; only a banned account cuts the payment chain and the follower pool. From a governance standpoint, this is the cut that hurts more.

But account handling is an after-the-fact move. A companion commentary on CCTV's Xinwen Lianbo laid out what platforms should do as a sequence: identify early, warn early, intercept early, clean up early. Three of those four "earlies" happen before anything is published.

The trouble is that generating an AI face swap is now cheap enough to try at volume, while detection is still expensive. A 15 second fake can be produced in minutes with off the shelf tools; a human reviewer needs tens of seconds before daring to call it; automated review still has to deal with compression, cropping and re-recording. The cost curves for generation and detection currently run in opposite directions.

The Measures for Labeling AI Generated Synthetic Content (《人工智能生成合成内容标识办法》) took effect on September 1, 2025, issued jointly by four departments, and require both explicit and implicit labels. Explicit means a notice users can see; implicit means a mark buried in file metadata. The design works inside a normal publishing pipeline. The problem is that bad actors never walk the normal pipeline: metadata can be stripped, explicit labels can be cropped out, and the screen can be re-filmed.

So what platforms can actually grip is not the watermark but matching people and things. The faces and voices of public figures are an enumerable small set. Matching against a known list is far more realistic, technically, than deciding whether a given video is AI generated. This is also where "prevention in advance" becomes practical: one interception at the publishing step costs less than tracing a thousand posts afterwards.

A phone screen showing a short video feed with a masked account name in the corner.

The courts have already shifted the burden of proof

The voice actor case at Shanghai No. 1 Intermediate People's Court on November 2 sent a very concrete signal. Of 28 formant acoustic indicators, the infringing audio deviated from the person's own voice by less than 10% in 24 of them, 16 of those under 5.36%, and "fairly similar" plus "highly similar" accounted for 90% in total. The court ordered the platform operator to pay 50,000 yuan.

More important is how the burden of proof was allocated. The rights holder only has to show two things: that there was a plausible channel through which their voice could have been obtained, and that the synthesized speech closely resembles them. That completes the initial showing. It is then up to the platform to prove its training data came from lawful sources, or that the synthesized speech has nothing to do with the rights holder. The judgment states the point of this rule plainly: to answer the structural problem of evidence sitting on one side in AI infringement cases, and to stop platforms from pushing liability away with excuses like "technical black box" or "the employee left and we cannot confirm".

The reasoning here maps onto Article 4 of the Supreme People's Court Opinions on the Trial of Cases Involving Artificial Intelligence (《最高人民法院关于依法审理涉人工智能纠纷案件的意见》), dated September 7, 2026. Applied to this latest campaign, the idea is the same: whoever holds the data has to be able to show where the data came from.

The practical meaning for platforms: source records for training corpora, proof of authorization for assets, logs from the review process. Day to day these look like compliance paperwork. When something goes wrong they are the only material that can show you were not liable. The value of a paper trail is not in ordinary times, it is on the day you are the defendant.

For merchants and creators, compliance comes down to three things

The first is the authorization chain. Using a real person's likeness or voice, whether you shot it yourself or a partner supplied it, requires a written scope of authorization. In the Shanghai case the platform argued it "did not know the other side was a voice actor". The court did not accept that, because not knowing does not mean you may use it.

The second is labeling. Explicit labels go in the picture, implicit labels stay in the file. This is now a hard requirement, not an option. Plenty of teams shrink the label to almost nothing because it spoils the look, which in real enforcement is quite likely to be judged non-compliant.

The third is keeping records of where assets came from. This is the one most often skipped, and the one most likely to be fatal in a dispute. Where the material came from, who supplied it, which uses the authorization covers: best archived alongside the finished video, not scattered across a few people's chat histories.

This campaign is not a one-off

The last line of the notice reads: "Next, we will keep increasing the intensity of the campaign" and "further consolidate the primary responsibility of website platforms". Wording like this has appeared many times in AI governance documents over the past two years. It points to routine enforcement, not a short burst.

For people running AI content businesses, the test actually got simpler: can you say who the person in the frame is, where the authorization came from, and whether this piece of content is labeled. Answer all three and you need not worry about your account vanishing one day. Fail to answer, and the bigger you scale, the faster risk accumulates.

For platforms, the competitive focus will move from "can we generate it" to "can we stop it in the second before publishing". 3961 accounts is this round's scorecard. The next one probably will not count how many were handled, but how many were stopped.

Related articles