Meta says every advertisement is automatically checked before it runs. That claim makes what happened on its platforms worse.
WIRED reported on August 5 that researchers at the Tech Transparency Project found more than 50 paid image and video ads containing alleged AI-generated child sexual abuse material or sexualized images of minors in Meta’s own Ad Library. The ads appeared across Facebook, Instagram, Messenger, and Threads. Some promoted AI undressing apps. Meta removed them after WIRED asked about them.
A user upload tests content moderation. A paid advertisement tests the product Meta sells. The company accepted the advertiser, reviewed the creative, approved the placement, delivered it to an audience, and measured the result. Meta’s safety system failed inside a commercial process it controls from beginning to end.
Paid ads pass through Meta’s gate
Meta described that process plainly in a July statement about child exploitation ads in India. Its system automatically checks ads before publication. Reviews cover images, video, and text. Ads remain subject to later review, while Meta can restrict the business accounts, ad accounts, pages, and user accounts behind them.
The company also says its rules categorically prohibit ads that sexually exploit or endanger children. Those rules did not stop the material WIRED examined. The ads ran between November 2025 and early August 2026, sometimes for several days. At least one reached 2,563 accounts across eight European countries. Meta’s public library does not disclose comparable reach data for the United States or every country, so the full audience is unknown.
Most of the ads had little measured reach, according to Meta. Limited distribution matters because fewer people exposed means less harm. It does not erase the approval. A company that charges for targeted distribution carries more responsibility than a passive file host. Meta sold the placement and then left outside researchers to find what its review missed.
The failures kept repeating
A July 26 investigation showed the same enforcement pattern at much larger scale. The Tech Transparency Project examined GatherOne, one of a small group of agencies authorized to place Meta ads for Chinese clients seeking audiences abroad.
Using Meta’s transparency data, TTP identified 210 Facebook pages linked to GatherOne that ran about 30,800 ads for face-swap and other AI apps. Forty-three pages ran more than 7,600 ads for apps that TTP verified could create nonconsensual sexualized images. Meta later removed policy-violating ads from 179 of the 210 pages. At least two dozen pages kept advertising after Meta had already removed some of their ads.
Independent reporting from Bloomberg, republished by the Los Angeles Times, confirmed Meta’s commercial relationship with GatherOne and the broad shape of TTP’s findings. Meta said GatherOne was one of 11 official agents in China. Meta also told investors it made $18.4 billion in China sales in 2024, or 11 percent of its global revenue.
The record shows a commercially important market, an approved advertising partner, and thousands of prohibited ads that the enforcement process failed to stop before users saw them. GatherOne occupied a formal place in Meta’s sales structure and repeatedly appeared across the ad network.
The pattern had already drawn government action. In early July, India’s electronics ministry ordered Meta to remove offending Instagram ads and explain within seven days how they passed review. The Hindustan Times reported that the notice followed a BBC Eye investigation that found roughly 30 paid ads directing people toward Telegram channels allegedly selling child sexual abuse material. Those ads were a separate set and were not necessarily AI-generated. They were another warning that Meta’s paid distribution system was being used as a route to abuse.
Meta’s numbers miss the paid channel
Criminal advertisers adapt. Meta says they use benign images, new domains, fresh accounts, and coded language to evade filters. Meta also told WIRED that many of the ads in the latest report were disabled before the newsroom shared them, and that newer AI detection technology was launched after many of the ads first ran.
The company has real enforcement work to point to. It says it removed 36 million pieces of child exploitation content and more than 4 million suspicious accounts last year. In 2025, Meta sued the company behind CrushAI, blocked search terms and violating links, shared more than 3,800 URLs with other technology companies, and built detection meant to find nudify ads that use benign images.
That work deserves to be counted. It also makes the repeated advertising failures harder to excuse. Meta already knows this is an adversarial network problem. Its own account describes advertisers creating replacement domains and accounts after enforcement. Yet TTP found repeat pages continuing to buy distribution after earlier ads were removed, and WIRED found more ads close to publication, including some that were still live after the newsroom contacted Meta.
The 36 million figure measures enforcement across an enormous stream of platform content. It does not tell the public how many severe paid ads were blocked before publication, how many ran first, how many came from repeat advertisers, or how often Meta found them before a watchdog did. A broad content-removal total cannot answer a narrow question about a product Meta directly sells.
Advertising gives Meta more ways to intervene
Ordinary posts can arrive from billions of users with little warning. Advertising creates additional evidence before distribution starts. Meta sees the payer, account history, landing page, targeting choices, creative, and connections to related business assets. Its own description of ad review says it can act on the ad, business account, ad account, page, and associated user accounts. It can examine those signals together and deny access to the advertising system.
The abuse is growing fast. The National Center for Missing and Exploited Children received more than 400,000 CyberTipline reports with a generative AI connection in 2025. More than 182,000 involved possessing, generating, or attempting to generate AI child sexual abuse material. The Internet Watch Foundation separately identified 3,443 AI-generated child sexual abuse videos in 2025, up from 13 the year before.
Those numbers describe real victims, including children whose existing images are manipulated into new abuse. They also show why dismissing these ads as a tiny fraction of Meta’s business is backwards. Severe harms need stricter controls precisely because ordinary percentage metrics can make them disappear inside a platform with billions of users.
Meta knows how to identify an advertiser, target an audience, collect payment, and measure delivery. It can use the same operational discipline to stop repeat abusive advertisers before taking their money. Until it does, every safety boast carries the same ugly footnote: watchdogs keep finding the ads first.