A failure in the paid-content gate

Meta removed more than 50 image and video advertisements after researchers identified material described as AI-generated child sexual abuse imagery or sexualised depictions of minors in the company’s advertising archive. The advertisements appeared across Facebook, Instagram, Messenger and Threads between November 2025 and early August 2026, according to an investigation by WIRED based on findings from the Tech Transparency Project.

The distinction between ordinary posts and paid advertisements is central. Advertising is not simply content uploaded into a social network: it passes through a commercial system designed to approve campaigns, deliver them to defined audiences and collect payment. Meta says every advertisement is reviewed before publication, primarily using automated tools. Its rules expressly prohibit child sexual exploitation and also bar sexually suggestive imagery.

That makes the case a test of whether a platform’s safeguards work at the point where it has the greatest operational control. Meta’s ad library recorded that the identified campaigns were ultimately removed for breaches of rules on child sexual exploitation, abuse and nudity, or adult nudity and sexual activity. Yet some of the ads had remained visible in the archive for months, and researchers found additional active examples shortly before publication.

What the investigation found

The Tech Transparency Project found the advertisements while examining promotions for so-called nudify or undressing services. These services use generative AI to fabricate sexual imagery, often from ordinary photographs uploaded by users. Some of the advertisements linked to apps or websites offering such capabilities.

The available data suggested that most of the ads had limited delivery, although at least one campaign reached more than 2,500 accounts in several European countries. The complete reach cannot be established from the public archive because it does not provide equivalent performance data for every market, including the United States.

The reported campaigns were often associated with newly created or low-follower accounts, a pattern consistent with advertisers attempting to replace accounts after enforcement action. Researchers also said some material appeared to duplicate ads that Meta had previously removed. If so, the problem is not only initial screening but the ability to recognise repeated creative material, connected advertisers and linked destinations after an earlier violation.

Meta told WIRED that sexual exploitation is horrific and that it works aggressively to keep it off its platforms. The company said many of the identified ads had minimal reach or had already been disabled, and that many predated newer AI systems intended to detect and block violating ads during upload. It also said it removed more than 36 million pieces of child sexual exploitation content in the preceding year.

Those statements describe substantial enforcement activity, but the new findings show the limits of counting removals as a measure of prevention. A system can remove large volumes of harmful material and still fail in serious individual cases, particularly where the content is monetised and promoted through its own advertising infrastructure.

Why synthetic imagery changes the moderation problem

Generative AI has reduced the expertise, cost and time required to create convincing sexualised material. It can be used to manipulate photographs of real children, produce synthetic depictions or create material used for harassment, coercion and sextortion. The National Center for Missing & Exploited Children says it identified more than 275 direct victims of AI-generated child sexual abuse material in 2024 and 2025, while also warning that offenders have used the technology to blackmail children and re-victimise known survivors.

The reporting data also demonstrate an important measurement problem. NCMEC received more than 400,000 reports with a generative-AI connection in 2025, but not every report necessarily involved newly generated abuse imagery. The organisation separately categorised more than 158,000 submitted images and videos as AI-generated between January 2023 and December 2025. Platforms themselves had labelled only a fraction of those files that way.

That gap matters because detection remains harder when an image is newly generated rather than a known file that can be matched to a pre-existing digital fingerprint. It is also harder when advertisers disguise a harmful destination behind harmless-looking creative, frequently change web domains or use fresh accounts. The Internet Watch Foundation has reported that realistic AI-generated abuse imagery is appearing on both dark-web services and mainstream commercial platforms, with AI-generated video now a growing concern.

Transparency is not the same as enforcement

Meta has expanded disclosure tools that can label advertisements created or significantly edited with its own generative AI products. It also says it is beginning to detect third-party AI signals and add AI information to an advertisement’s disclosure panel. Such labelling can help users understand how an ad was made, but it does not establish that its content is safe.

The current case illustrates the difference. A disclosure label addresses authenticity and provenance; child-safety enforcement must identify harmful content, deceptive presentation, repeat advertisers, suspicious links and coordinated account networks before an advertisement is delivered. It must also work where AI metadata is absent, stripped out or never supplied.

The public ad library has value because it enabled outside researchers to identify and document problematic campaigns. However, the investigation also highlighted a weakness: researchers said there was no direct reporting route within the library for advertisements no longer actively running. A transparency archive is more useful when it supports rapid escalation, preserves evidence and makes it easier for independent researchers to identify patterns across advertisers and destinations.

The broader accountability question

Meta has taken steps against the nudify ecosystem, including legal action against an alleged repeat advertiser and the removal of large numbers of ads. Apple also removed one app linked from the identified Meta ads after WIRED contacted the company. These measures show that action can be taken across the chain of ad delivery, app distribution and payment.

But the incentives and responsibilities are not identical. Platforms profit from advertising before it is removed, while victims and investigators bear the cost of discovery and reporting. For a category as severe as child sexual exploitation, the relevant standard cannot merely be whether a company acts after a journalist or watchdog identifies a campaign. It is whether the advertising system prevents repeat abuse at scale.

The case therefore points toward a more demanding approach: stronger screening of advertiser networks and outbound links, rapid matching of previously removed material, meaningful human escalation for high-risk signals, and reporting mechanisms that work for active and archived advertisements alike. Generative AI has made the production of harmful imagery easier. Advertising systems must not make its distribution easier too.

Sources