Meta says strengthened systems, deployed new LLMs to prevent, detect CSAM

Meta says a dedicated language model, AI sweeps and a red-teaming agent will help detect harmful ads, identify gaps in its defences and stop banned users returning

Social media giant Meta on Wednesday said it has improved its child sexual abuse material (CSAM) detection systems to improve the “efficacy” of its systems in preventing such content from being posted across its platforms.

 

The company listed a series of improvements, such as deploying a new large language model (LLM) to flag advertising content that may look relatively harmless but shows indicators of redirecting users to illegal and harmful content off Meta’s platforms.

 

In a blog post, Meta said that while it continued to detect, investigate and remove CSAM from its platforms, bad actors had resorted to newer methods, such as using advertisements to redirect people to other third-party platforms.

 

 

“Once we understood the design of this malicious advertising scheme, we didn’t stop at the ones reported to us. We widened our investigation to find related activity across our platforms, disabled the violating accounts behind it, and blocked the off-platform links involved,” the company said in a blog post.

 

In addition to deploying the dedicated LLM, Meta has also launched new artificial intelligence (AI)-driven “sweeps of ad content” to look for child sexual abuse and exploitative material that its systems did not catch, the company said in the blog post.

 

A red-teaming AI agent has also been deployed by the company to find loopholes in Meta’s defences against CSAM, so that these weaknesses can be found and corrected, the company said.

 

Meta has also strengthened its systems for identifying bad actors who try to return to its platforms after their accounts are removed, the blog post said.

 

These actions by the social media conglomerate come within a month of the company’s top executives being summoned by the Indian government over alleged failures to contain CSAM and allowing advertisements that promoted such material across its platforms.

 

Last month, Meta became the first social media platform to agree to share details of CSAM cases and repeat offenders with Indian law enforcement agencies. Until then, all US-headquartered social media platforms, including Meta, shared instances of CSAM only with the National Centre for Missing & Exploited Children (NCMEC), a private US non-profit organisation.

 

The NCMEC would then share the details of repeat offenders or other such online accounts with the law enforcement agencies in their respective domains.

 

Google soon followed suit and said that it would also start sharing these details with Indian law enforcement agencies.