Brands pay influencers for reach on the belief that behind the profile stands a real, traceable person. A real face is supposed to guarantee authenticity, something AI cannot yet convincingly fake. The YouTube case reverses that assumption. A real face became the gap in the platform’s defences.
YouTube removed 20 channels: thirteen named in the report and seven related ones, which published political content attacking Democratic politicians. The channels looked like independent commentary creators. In reality, they were a network of so-called ghost creators, where paid actors played the personas and AI wrote the scripts.
A real face in front of an AI pipeline
The exposure was described on 2 September by the server Semafor, where Max Tani wrote the report. The independent analysis of the network was carried out by the research firm Riddance AI. The operation used AI for scripts, thumbnails, headshots and websites, but put real, paid actors on camera.
This hybrid approach got past synthetic content detectors. Platforms look for generated video, not for a real person reading a script. The Riddance report pinpoints where full automation runs into trouble. “The weakest link is often the AI-generated videos themselves. Compelling direct-to-camera delivery is still difficult for AI video generation models,” the analysis says. A human face fills exactly the gap that AI video cannot yet bridge.
Gig actors on camera, AI scripts behind
Actors received $26 per video through the gig platforms Backstage and Fiverr, with a 24-hour deadline to shoot. One Backstage listing sought “long-term African American male hosts to record news commentary scripts”. The production was handled by a company called Virelox. Actor Frank George told Semafor that he was paid through a broker called “Casgains” and did not know where the funding came from.
According to the Riddance report, by mid-August the network had gathered more than 45 million views and over 90,000 comments. Individual personas pulled in hundreds of thousands of viewers per video. According to Semafor, one video featuring the persona “William” had 474,000 views, and videos of the personas “Omar” and “Julian” had 250,000 each. Riddance also counts millions of further views on TikTok.
Spam, not falsehood
YouTube only removed the channels after Semafor and Riddance published their findings. In a statement to Semafor, the platform said it has “strict policies against spam and deceptive practices”, and it removed the channels for spam and deceptive practices. The action was not a penalty for false content. It enforced rules on account volume and authenticity, not on the accuracy of the message.
Riddance points to a deeper cause. “If an ideologically agnostic YouTube automation hustler decides that the best way to make money on YouTube is by pushing inflammatory content about a single political party, that points to something structurally wrong with YouTube’s engagement algorithm,” the firm writes in the report. The funding targeted one political party, but the model itself rewards whatever holds attention.
AI video detection is not enough for hybrid content
The case moves the question of authenticity up a level. Until now, it was assumed that artificial content could be recognised from a generated face or voice. A real actor in front of the camera cancels out that shortcut. Detecting generated video is not enough when the face is real and the script is machine-made.
For a market that buys influencer reach, this is not just an American political episode. A European advertiser who books creators by views and comments faces the same layer of trust. A view count says nothing about who commissioned the content or why it was made. And regulation is aimed somewhere other than where this case points. The European AI Act in Article 50 introduces an obligation to label synthetic content, but an operation with a real actor in front of the camera falls outside the definition of artificial output. YouTube’s intervention came through its spam rules. A tool that targets this hybrid form of content directly does not yet exist.
Comments