Abusive AI ad buyers reached 36,000 EU and UK accounts on Meta for under $5,000

Under $5,000 in payments across more than 350 creatives works out to under about $14 an ad ^[15][2][1], and that price is the mechanism. Meta's ad library discloses reach of more than 29,000 accounts across the European Union and around 7,000 in the United Kingdom ^[10], so 36,000 accounts ^[2] bought for at most $5,000 comes to roughly fourteen cents an account ^[3], and that is before the United States, Australia and India, where Meta publishes no granular performance data ^[11].

The ads also travelled. Country counts of more than 250 in the US, 120 in Australia and 80 in India sum to 450 placements against a catalogue of 350-plus ^[11][2][4], so single creatives were running in several markets at once.

David Pham, who investigated nonconsensual imagery at Meta until last year, says advertisers routinely beat new detection models within a day to five days by switching domains and obscuring who posted the ads ^[19]. That is a shelf life shorter than most release cycles, and it makes the company's statement that it reviews all ads before they run ^[12] hard to test against any single tool launch. TTP says live ads reported through Meta's own tools sometimes took a week to come down ^[13]; one ad using an identified teenager had reached four accounts in Europe when it was reported and more than 330 users by the time it came down days later, an eighty-fold gain earned during review ^[14][5].

The cheapest control available here is bookkeeping rather than machine vision: hash what you removed, refuse the re-upload. TTP found ads identical to ones Meta had just taken down ^[4], and fixing that takes bookkeeping alone, not a new model.

Meta's account of what was in the videos and TTP's account describe two different objects. Meta says the AI-generated videos featured adults, with clips of minors brief or blurred and therefore harder to detect ^[16]; the researchers describe a template that opens on a photograph of a preteen or teenage girl and morphs her face into a graphic sexual video ^[7], with four minors identified, one of them a royal whose officially released photograph was used ^[5][6].

My read, and the numbers only partly carry it: with roughly 300 of the 350 ads pointing at AI image or video generators, most tied to China-based developers ^[8], the buyers were acquiring app installs at a cost per account reached that no ad-review budget competes with, and the enforcement that bit came at the app layer, where Apple has now acted on 27 apps ^[17][6]. The counter-thesis is Meta's own: most ads under 200 impressions and under $5,000 collected ^[15] describes a tail that any review system leaves behind. What would settle it is provenance on the identical re-runs, because if those files were queued before the new tools shipped, latency is the failure, and detection worked as intended.