A political video network hired real actors to read its AI scripts. That was enough to beat YouTube’s detectors.

YouTube has terminated a network of channels that reached 45 million views by paying gig-economy actors to read scripts that appear to have been generated by large language models. The enforcement followed an investigation published on Wednesday by Semafor, whose media editor Max Tani reported the story with Graph Massara.

The company said in a statement to Semafor that it has “strict policies against spam and deceptive practices” and had terminated 20 channels, comprising the 13 named in the piece and seven related ones. TNW has not independently verified the number, which appears only in YouTube’s statement and in no public enforcement notice.

The workaround was a human face

Everything about the operation looks automated except the one component that would have got it caught. Scripts, thumbnails, headshots and websites carried the usual generative fingerprints, but the person on camera was real.

That was the point, according to the media intelligence firm Riddance AI, which investigated the network in parallel with Semafor. Convincing direct-to-camera delivery is still hard for video models, Mason Broxham and Jeremy Carrasco wrote in their report, which creates a market for hiring someone cheap instead.

Platform enforcement has been built around detecting synthetic media. YouTube automatically labels AI-generated video whether creators disclose it or not, and a real person reading an AI script trips none of it.

Twenty-six dollars a video

The rate is documented. A casting call posted by the production company Virelox sought a professional on-camera spokesperson for a news-related YouTube channel, and Riddance found a companion listing offering $26 per script, with 24-hour turnaround and the actor supplying their own camera and lighting.

Actors were sourced through Fiverr and Backstage, the same places brands hire spokespeople for meat thermometers and ear-cleaning devices. One of them, Frank George, told Semafor through Fiverr’s messaging service that he was paid via a broker and did not know who was funding the work.

He described the videos to Semafor as paid advertisements and himself as a spokesperson. Asked to name the broker, he replied: “Casgains.”

Who was behind it

Semafor traced the network to Caleb Chan, a University of Chicago graduate who built a finance YouTube channel called Casgains Academy and then turned virality itself into the product. He co-founded Virelox, which sells the rapid creation of creator accounts on behalf of clients.

Chan posted on X in April that he had 106 YouTube accounts running and wanted 300 by May, and in a later post claimed a network of more than 500 ghost creators generating billions of views a month. Those are his own promotional claims and have not been audited.

He also posted what he believed he had learned about the ranking system, writing that when many accounts publish the same hashtag within minutes the algorithm treats it as breaking news. He said he had tested it nine times, and signed off: “manufacturing news > reporting news.”

What the channels published

Riddance analysed 622 video transcripts across the personas, tallying mentions of individual politicians and categorising the claims attached to each. Tax burden, budget problems and business flight from Democratic-run states dominated, with New York mayor Zohran Mamdani the most frequently targeted figure.

Titles named a Democratic politician, then attached an all-caps verb to an economic story or a scandal. Most underlying claims tracked real reporting, Riddance found, but the framing often did not survive contact with the video itself.

Semafor reviewed one William Reports News video headlined around Mamdani panicking over Hilton hotel closures, and found it contained no such claim and argued close to the opposite. Another said governor Kathy Hochul had refused to approve Mamdani’s tax proposal, which she has backed.

The seams showed, and nothing happened

Coordination was not subtle. Every persona used the same ominous piano bed, and a pairwise comparison of the transcript corpus found the Omar and Julian US channels sharing 304 script segments.

Personas turned up on each other’s channels without explanation, Linktree profiles pointed at the wrong host, and one actor switched from an anti-Trump persona to a pro-Trump one within weeks. Riddance recorded the associated websites all going offline on the afternoon of 25 August, after Semafor contacted the company.

The audience was real

This should worry platforms more than the production method does. Riddance found the 90,000-plus commenters substantially overlapped with audiences on other right-leaning YouTube channels, meaning actual people were watching.

Internal analytics for one persona, found by Semafor on a dashboard reachable through a simple Google search, showed an audience 85% American, 80% male and 80% over 55. The network out-engaged the YouTube channels of the Washington Post, the New Yorker and the Boston Globe, Semafor reported.

Terminated for spam, not for being wrong

Read YouTube’s statement again. The violation cited is spam and deceptive practices, which is a volume-and-authenticity rule rather than an accuracy one.

That framing matters, because it is the enforcement lever the platform actually has. The same lever has been catching faceless creators who never used AI at all, which is what happens when a proxy signal stands in for the thing you care about.

It is also the second time this enforcement has followed publication rather than detection. YouTube removed several channels after Semafor’s first story on AI-generated blue-state collapse videos in early August, and Riddance had been tracking this network since June.

The distribution layer is still up

Terminating the source channels does not clear the clips. Riddance documented dozens of TikTok accounts recirculating moments from the long-form videos, some of them exclusively.

Those accounts sit inside a much larger ecosystem. Nearly 60% of TikTok videos served to new users are AI slop, by one study’s count, so removed material does not so much disappear as get demoted to a lower shelf.

Why the loophole holds

Regulation is heading the same way as platform policy, towards synthetic media. Brazil is currently defining what counts as a deepfake ahead of its vote, and definitions built around generated likenesses will not catch a real man at a kitchen table.

Nothing here required a frontier model. It required an LLM, a Fiverr account and someone willing to read for $26.

Chan said in a statement to Semafor that Virelox is an independent digital media publisher which takes no political clients and uses no synthetic anchors, and that it has zero affiliation with one of the channels in question. Either way, the operation reached tens of millions of people before two newsrooms, rather than any classifier, brought it down.