Meta says it caught over 97% of 33.2 million pieces of child-exploitation content before anyone reported them

Meta’s newsroom post credits two new artificial intelligence tools for part of its latest child-safety enforcement results on Facebook and Instagram. India gets its own line in the post. Meta says it actioned 5.3 million pieces there, with over 98 percent caught proactively.

What the 33.2 million and 97 percent measure

Across both apps, worldwide, the post says the company actioned 33.2 million pieces of child sexual exploitation content in the first half of 2026, and that over 97 percent was “found and proactively addressed before anyone reported it.” The wording matters: the unit is pieces of content, not people, and “actioned” describes enforcement taken by Meta rather than a count of arrests or prosecutions.

Because the global number covers every country, the Indian total is a subset of it, not an addition to it, and the post presents all of these numbers as the company’s own tallies. The figures come from Meta’s own newsroom post, published on its India newsroom page. Meta reported on October 7 that it actioned 33.2 million pieces of child sexual exploitation content on Facebook and Instagram between January and June 2026, and that over 97 percent of it was found and dealt with before any user reported it.

The LLM and the red-team agent

The newer part of the story is in the ad system. Meta says it widened an investigation beyond the specific ads that had been reported to it, disabled related accounts and blocked links that led off the platform. It then added a large language model built to spot what it calls signposting: advertisements that look harmless on the surface but covertly point people toward illegal content or harmful activity somewhere else. The post describes a companion step of analysing where an ad leads, rather than judging only what the ad shows.

A second addition is a red-teaming agent, an AI system that Meta says is set to probe the company’s own defenses and hunt for weaknesses and new tactics before they spread widely. The post also lists AI-driven sweeps that surface material earlier systems missed, and stronger detection of repeat offenders who open fresh accounts after being removed. Meta does not publish, in this post, a separate success rate for the LLM or the agent, so how much of the 97 percent they account for cannot be read from the document; the post presents them as additions to a layered system rather than as the reason for that rate.

Hash matching, link blocking and the human team

Much of the 97 percent still rests on older machinery. Meta says it has used Microsoft’s PhotoDNA across its apps since 2011 to catch identical or near-identical copies of known abuse imagery, alongside its own photo and video matching, which the company released as open source in 2019. The same matching is applied to link previews that point off the platform, and when a link to a site hosting or creating violating material is blocked, other posts, ads or comments carrying that link are removed and ads with such links are rejected at upload.

Behavioral signals combine account metadata with patterns of activity to flag suspicious profiles, and heuristic rules that blend classifiers, AI models and other signals are meant to catch new violation patterns, including in ads. People remain in the loop: Meta describes a team of experts, including former FBI investigators, who run manual investigations to uncover, remove and report networks of predatory accounts. The post names none of them.

Sharing signals and reporting to authorities

Meta says it sends reports of the content it removes to the National Center for Missing & Exploited Children, in line with applicable law, and that the center also helps pass reports to law enforcement. It also contributes hashes to Lantern, a signal-sharing program that Meta described in 2023 as a way for companies to share signals about accounts and behaviors that violate child safety policies so that predators who move between apps and websites can be tracked. According to that post, the Tech Coalition launched Lantern with Meta as a founding member and infrastructure provider. In the pilot described there, the file-sharing service MEGA contributed URLs it had already removed for policy violations; Meta used them to run a wider investigation, removed more than 10,000 violating Facebook profiles, Pages and Instagram accounts, reported them to the center, and fed its findings back into the program. That example shows the intended pattern, in which one company’s takedown becomes another company’s lead.

In India, the post adds, Meta announced in September 2026 that it will report child safety cases directly to the National Cyber Crime Reporting Portal, which is run by the Indian Cybercrime Coordination Centre, building on reporting it already does through the portal with help from the National Center for Missing & Exploited Children.

The post closes on the company’s stated priority of keeping young people safe on its apps and describes the work as continuous. What the document does not supply is an outside check: the 33.2 million and the 97 percent are Meta’s figures, so how many pieces slipped past detection entirely remains outside the reported numbers.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

More from Morning Overview