- Ad Failure: Meta’s pre-run checks missed paid ads containing child sexual-abuse material or promoting AI tools that sexualize ordinary photos.
- Review Gap: Tech Transparency Project researchers found delayed responses, inconsistent decisions on identical ads, and new ads appearing after earlier removals.
- Official Scrutiny: US and Australian authorities opened distinct inquiries or requested answers, but none had announced a violation or sanction by September 8.
Meta’s ad-review system allowed hundreds of paid ads containing child sexual-abuse material or promoting AI tools that create sexualized images from ordinary photographs to run after the company announced new detection measures. A Tech Transparency Project investigation published September 8, ifound that some manipulated ads used photographs of four identifiable real children; the findings prompted separate US and Australian inquiries and information requests.
The ads did not form one uniform set or represent one victim count. Prohibited material passed Meta’s automatic pre-run check, reports sometimes received delayed or contradictory decisions, copies appeared after removals, and ads sent users toward tools that can create sexualized images from ordinary photographs. Meta removed the documented sample and says it has strengthened detection, but the observed recurrence left the effectiveness of those controls unresolved.
How Meta’s Review Failed
Meta says every ad is checked automatically before it runs and may receive additional automated or manual review afterward. The company also prohibits ads that exploit children. Yet the Tech Transparency Project’s detailed investigation found paid ads that had passed that first check and were delivered across Facebook, Instagram, Messenger, Threads, and Meta Audience Network placements.
The researchers’ reports exposed a second weakness. They submitted reports on 129 active ads. Meta initially told them that 73 did not violate policy and that 56 had already been removed, although six in the latter group remained visible. Some reviews took at least a week, and identical creative material sometimes received different decisions. One reported ad using a real child’s photograph expanded from four accounts reached in the European Union to more than 330 before Meta removed it.
By September 1, 183 ads in the researchers’ 332-ad sample were no longer available in Meta’s Ad Library. The company removed the remaining 149 within hours of receiving the full findings on September 2 and corrected 113 notices that had not identified child exploitation as the reason for action. Researchers nevertheless found dozens of new ads running after that notification. Wired reported that the identified ads had all been removed by publication on September 8.
The Counts Describe Different Samples
The first investigation identified 53 ads from late 2025 through early August. The later work found more than 250 additional ads after August began and examined a detailed set of 332 ads from 2026. Wired describes a broader total of more than 350 abusive video ads since late 2025. Those figures overlap and use different windows, so they cannot be added together.
The categories also overlap without being interchangeable. The 332-ad sample contained material the researchers classified as child sexual-abuse material; 298 ads directed users to AI image or video tools. Testing and image tracing showed that some promoted nudification or sexual-video functions, while a smaller group used photographs of four real children as source material.
Meta Ad Library data available to the researchers showed at least 29,000 accounts reached in the European Union and 6,800 in the United Kingdom for the visible subset. Those regional figures do not establish worldwide reach. Meta separately said most ads received fewer than 200 impressions and that advertisers paid it less than $5,000 in total.
Why New Images Complicate Detection
Services can compare an uploaded file with a cryptographic fingerprint, or hash, of material already reviewed and classified as abusive. Known-file matching and the detection of previously unseen material are separate tasks. Newly generated images and altered versions of ordinary photographs may not match a known hash, so detection also depends on image analysis, behavioral signals, reporting, and human judgment.
That technical difference helps explain why new or manipulated material can be harder to catch, but it does not account for the entire record. The researchers saw identical ads receive different review outcomes, and copies of removed material returned. The operative failure therefore extended beyond identifying a novel file to consistent classification, report handling, copy matching, and action against repeat advertisers.
Meta Removed the Sample but Has Not Shown Closure
Meta told Wired that it does not tolerate nudification apps or child exploitation, whether real or generated with AI. It said adversaries change their methods, that brief or blurred imagery can be harder to detect, and that many of the ads had already been disabled. The company also said it reports apparent child exploitation to the National Center for Missing and Exploited Children, or NCMEC, and that it was reporting these ads. The researchers said they had also submitted their findings to NCMEC. Neither NCMEC nor law enforcement have publicly confirmed how these specific reports were handled.
Meta has previously detailed its specialized ad detection, copy matching, link blocking, signal sharing, and account-network actions. It also announced a performance-gated, potentially multi-year transition of more content enforcement to AI earlier in 2026 while retaining people for high-risk decisions. Those announcements describe available or planned controls, not proof that a particular control reviewed each ad or prevented recurrence.
The ads also exposed a downstream enforcement chain. Tech Transparency Project had previously documented nudification tools in Apple and Google app stores, and a later San Francisco demand prompted removals and suspensions of named store apps. In the September episode, Apple and Google again said they had removed or suspended flagged destination apps. Such store actions can cut off one destination, but it do not stop new ads or copies on Meta’s network.
NCMEC offers the CyberTipline for reporting suspected exploitation and Take It Down for people concerned about explicit images of themselves created before age 18. People should not save, circulate, or try to investigate suspected abusive material themselves.
Scrutiny Is Active but Unresolved
Senator Mark Warner sent Meta questions and not received a reply by the company. Florida Attorney General James Uthmeier said his office was investigating. Michigan’s attorney general is also actively looking into the reports.
Australia’s eSafety Commissioner is also assessing Meta’s compliance and has requested information at a senior level.
