Meta deploys an LLM to flag ads that signpost child abuse material hosted off its platforms

Meta's Wednesday blog post defines "signposting" as ad content that appears benign but is strongly suspected of directing people to illegal content ^[1][4]. The LLM has to judge intent from an ad that looks clean. A classifier that scores only the ad's image and text has nothing to flag, because the material sits on hosts outside Facebook and Instagram, according to Meta ^[6]. The company paired the model with AI-driven sweeps meant to surface material its earlier systems missed ^[13].

Meta did not disclose how many ads, accounts or links were involved, where the material was hosted, or how often the model flags a legitimate ad ^[12][4].

The detection rates in the same post cover content Meta's filters could already see. Over 97% of the global total was found before any user report ^[5]. That leaves fewer than about 1 million items that someone else flagged first ^[15]. India accounted for about 16% of the global count, and more than 98% of it was found proactively ^[16][3], so fewer than about 106,000 Indian items reached Meta as reports first ^[17].

The ad scheme came in through reports. By Meta's account, its investigation started from ads that had been reported to it ^[12]. In July, India's IT ministry issued Meta a notice over such ads on Instagram, ordered them disabled and summoned the company's global executives to India ^[9]. Hindustan Times reported that the disclosure comes amid a government crackdown on child sexual abuse material ^[11].

The case for a sustained campaign rests on Meta's word. The company describes the ads as a change in predators' tactics ^[6]. Blocking a link stops a click from Meta's apps. The material itself is hosted off its platforms ^[12][6]. Meta says it has also strengthened detection of banned users who create new accounts, a control aimed at operators who come back ^[14].