Is Your Favourite Political Influencer Real? 'Ghost Creators' Are Paid to Read AI Scripts Attacking Democrats
Political operatives paid to perform AI-written scripts while posing as authentic influencers shows how synthetic content can evade detection in high-stakes public discourse.
On this page
AI-Written Scripts With a Human Face Beat Every AI-Content Filter You Own
Researchers at Semafor and the AI-ecosystem research group Riddance AI spent 2026 tracing a US political YouTube network in which real, on-camera presenters read scripts that read like AI-generated or AI-assisted news copy, aimed mostly at Democratic figures including Zohran Mamdani, Gavin Newsom and Alexandria Ocasio-Cortez. By mid-August the connected channels had drawn 45 million views and 90,000 comments. Casting posts tied to a company called Virelox and a broker called Casgains offered on-camera presenters $26 a video for what one listing called long-term, high-volume work. Researchers pulled 622 transcripts from the network and found titles and claims that were often wildly exaggerated versions of real events. YouTube subsequently removed channels tied to the reporting for violating its spam policies on coordinated, repetitive content, though it's still unclear who actually commissioned the scripts.
Why couldn't AI-content detectors catch this?
Because the deception sits in who's paid to say the words, not in the words themselves — no text or video scanner audits commissioning relationships.
Tools this site covers regularly — GPTZero, Originality.ai, Copyleaks, Turnitin — score linguistic fingerprints: perplexity, burstiness, phrasing tics that survive in a raw document. A human reading a script on camera erases almost all of that signal. Pacing, ad-libs, vocal inflection and edits scrub the statistical trace an LLM leaves behind, so even if the underlying script was drafted by a model, what viewers actually watch carries none of the surface evidence detectors are built to find. Video-authenticity tools aimed at synthetic avatars or voice clones are equally blind here, because the presenters — "Omar," "William," Victoria Foster and the rest — are real people, not synthetic media.
What does the audience data actually tell a buyer?
Four analysed channels skewed heavily toward men over 55, roughly 80% on both measures — a demographic long targeted by persuasion campaigns generally.
That skew matters for anyone weighing provenance and watermarking pitches: an older audience is statistically less likely to check channel ownership, upload history or shared branding before trusting a news-style presenter. If your detection strategy stops at "does the text look machine-written," you're ignoring exactly the audience segment least equipped to spot commissioned content on their own.
Should platform "AI slop" rules have caught this sooner?
Not automatically — spam and impersonation policies target repetitive posting patterns, not who wrote or funded the script behind a human presenter.
This is the same gap this site flagged when covering YouTube's AI-slop crackdown missing the difference between a directed AI film and a bot farm: enforcement follows behavioural signals, not authorship. Here, channels ran up 45 million views before removal, and it took outside journalists cross-referencing 622 transcripts and a named vendor, Virelox, to establish the pattern — not platform detection at scale.
Frequently asked questions
Does this mean AI text detectors are worthless for this kind of content?
Not worthless, just out of scope. Detectors like Originality.ai and Copyleaks read documents; once a script is voiced by a paid human presenter, that text layer is no longer what the audience sees or hears.
What should I ask my own detection or brand-safety vendor?
Ask whether coverage extends to script provenance and commissioning networks, not just surface linguistic markers in a file you already hold — that's the layer this network exploited.
Is this only a political-advertising problem?
No. The same "paid presenter, AI-assisted script" model works for product reviews, testimonials or local news — any category where a human face currently substitutes for verification.
Sources: International Business Times UK, drawing on the original investigation by Semafor and Riddance AI.