On August 5, 2026, Meta stated that sexual exploitation is “horrific” and that it works “aggressively” to keep it off its platforms. But that same week, its public ad library was still showing paid advertising pieces with AI-generated child sexual abuse material (CSAM), in some cases still running. Between those two statements, the declaration of principles and the evidence in its own catalog, lie nine months in which the advertising machinery that generates 98% of Meta's revenue charged money for content the company itself defines as intolerable.
What the investigation by WIRED, based on data collected by the Tech Transparency Project (TTP), revealed is not an isolated incident, but the third warning in six weeks of the same structural failure. And indeed, on July 3, an investigation by BBC Eye documented that Instagram was showing ads in India promoting the sale of abuse material, with links to Telegram channels where it was sold for just over a dollar.
Later, on July 27, the TTP linked GatherOne, one of Meta's main Chinese advertising partners, to thousands of ads for “nudify” apps. And on August 5, the same organization revealed the most serious finding: more than 50 ads that directly contained synthetic images and videos of minors in sexual situations, published between November 2025 and the first days of August.
A door that opens by itself
According to Meta's own documentation, every ad on Meta's platforms goes through a review process that, per the company's policies, relies “primarily” on automated tools. That is the only real filter. When content passes review, it enters a system designed to maximize delivery speed: it is targeted at audiences, optimized based on performance and charged by results. And here comes the first harsh reality: safety is an upfront check; money is a continuous function.
This should not surprise us; it happens at Meta, it happens at Google (and its well-known spam links that generally lead you to scams), and it happens on X. Practically, wherever there is advertising, this kind of thing can be seen to a greater or lesser extent, depending on the platform's reach. But the responsibility on such large platforms is enormous, and entrusting that “filter” to a specific action in time, without subsequent reviews, is what leads to things like this happening.
Visa Direct adds stablecoins to operate 24/7 with Zero Hash
Without any disguise
After all, the TTP researchers describe the ads discovered on Meta as something that “made no attempt whatsoever to disguise its purpose”. For example, one of the videos discovered used as its thumbnail the image of a child sitting on the floor with text inviting viewers to “act out deep fantasies” and, when played, overlaid the minor's face onto sexual acts between adults.
Another ad showed a young girl lying down with her legs open and the message “I can show you more”. A third began with the image of a girl applying lip gloss and transformed, when played, into a sexual act. All three passed the “review” and once they passed, they were forgotten and served to the public.
The reach reinforces the severity. Although most of the pieces reached a handful of accounts, at least one reached 2,563 accounts in several European countries, and the ad library offers no data for the United States, so the real impact is probably greater.
The most disturbing data point is the pace of the failure: researchers first located about two dozen ads and, hours before the investigation became public, found about 30 more, several published after WIRED contacted Meta. The filter was not only opening; it was opening while the company was being told it was opening.
A supply chain designed not to see
But the case also has a specific geography. Meta is blocked in China, but Chinese advertisers need Meta to reach global audiences, ironies of life, and of network censorship. The solution was to create a network of authorized intermediaries that manage advertising for Chinese companies on the platform.
That is the channel through which part of the problematic content enters. In one of the documented ads, Meta identified Meet Social, a Chinese firm that became one of those authorized resellers and that exceeded $1 billion in sales, as the advertiser. The connection to “nudify” apps (apps that use AI to undress photographs or swap faces in sexual videos) is no coincidence: several ads linked to MaskAI, an app on Apple's App Store, removed after WIRED contacted the company, that hid its real content until after download.
The attack on Coldcard is still active and losses could reach $130 million
Mobilizing more than $1 billion
The TTP's July 27 report added one more name: GatherOne. The organization counted about 7,600 advertising pieces for “nudify” apps linked to this agency in recent months alone, with estimated spending of more than $70 million a year.
A Reuters report from December 2025, based on Meta internal documents, estimated that about 19% of the company's revenue in China in 2024, more than $3 billion, came from ads for prohibited content. GatherOne announced it was suspending the creation of new accounts for this type of service, but did not commit to stopping campaigns already underway. Not surprising, knowing it is a multimillion-dollar part of its business.
The piece that connects the whole mechanism is the one the TTP itself found by chance: among the August ads were pieces identical to others Meta had already removed months earlier for violating its policies. Its own systems had classified that content as infringing, and yet the copies circulated again. It is not a detection failure; it is an inconsistency in enforcement that should be interpreted with context: Meta's automated moderation behaves one way with organic and individual posts, and very differently when it is a paid ad, especially if that payment comes from a strategic partner.
The race between the generator and the detector
But beyond that, behind these ads is a technology that has become cheaper and faster than any moderation system. The Internet Watch Foundation (IWF), the British organization that tracks and removes child abuse material from the internet, documented in 2025 a more than 26,000% increase in AI-generated child abuse videos: 3,443 synthetic videos versus 13 the previous year.
65% of those videos were classified in the most severe category, compared with 43% of videos depicting real abuse, indicating that generative tools not only produce more content, but more extreme content. 97% of AI-generated images depict girls, and many are built from real photographs of victims, with models trained on already existing abuse material.
Meta has bet precisely on the long race it is losing. In March 2026 it announced it would reduce its reliance on external human moderators and would leave almost all content moderation and ad review to AI starting in 2027; the company maintains that its models make 13% fewer errors than humans. But the generator advances in parallel with the detector, and the first move is always the generator's: it takes weeks or months to detect a new pattern of AI-generated abuse, while producing and publishing it takes a few minutes.
Why the US needs the CLARITY Act even if the SEC and CFTC are pro-crypto
Meaningless excuses
Meta's response, that most of these ads “predate the new AI technology” launched to block them, only serves to admit the problem: there is a window in which infringing content circulates, is charged for and is optimized before the filter is updated, and that is precisely what abusers exploit.
Alexios Mantzarlis, former head of Trust and Safety at Google and now co-founder of the publication Indicator, sums it up clearly:
The “nudify” app ecosystem has gone from being a novelty to a permanent fixture of the online ecosystem, because platforms enforce the rules timidly and bad actors regenerate accounts and domains with the same techniques as scam networks. The only barrier between that advertiser and the user is an automated classifier trained on what it has already seen.
The damage that does not appear on the metrics dashboards
The severity of the case is not measured only in accounts reached. AI-generated abuse material re-victimizes real minors whose images are reused to train models and create new violations; it increases the risk that those who consume it progress toward contact abuse; and it fuels blackmail. And the numbers make it clear: the IWF documented a 127% increase in sextortion cases in 2025 compared with the previous year.
The spread of synthetic content also opens the door to the so-called “deepfake defense,” a legal strategy in which criminals claim that real abuse material was AI-generated, sowing doubt about the authenticity of the evidence. That a platform with more than 3.5 billion users charges money for this type of ad for nine months is not a minor operational failure, it is simply a redistribution of risks toward victims, with a legal and reputational cost that keeps growing.
And the regulatory context already reflects that cost. A New Mexico jury found Meta liable, in March 2026, for misleading users about the safety of its platforms and for allowing child sexual exploitation, with $375 million in penalties.
In April, the European Commission preliminarily concluded that Instagram and Facebook breach the Digital Services Act by failing to prevent children under 13 from accessing the platforms.
FATF warns of the industrialization of illicit use of cryptoassets
What needs to change
The solutions on the table do not require new technology, but decisions about where to draw the line. Tightening advertiser verification, especially for those who enter through authorized intermediaries, holding Chinese resellers truly accountable instead of tolerating their regenerating accounts.
That, in addition to allowing users to report ads from the public library itself (today that mechanism does not exist once a piece stops running) and systematically sharing hashes of infringing content with the rest of the industry, are concrete and applicable measures. Meta itself already shares thousands of URLs with the Tech Coalition's Lantern service; the problem is not technical capability, but the will to apply it consistently when the advertiser generates revenue.
Until that happens, the story of these nine months will function as a natural experiment: how long an automated moderation system can maintain the fiction of “zero tolerance” while its business depends on not asking too much about the content of whoever pays.
Editor's note: this article mentions child sexual abuse material and may be disturbing to some people; discretion is advised.