Meta תשתמש ב-LLMs לרוב בדיקות התוכן והפרסום — יעד 90% עד 2027

The Financial Times reported that large language models already handle 50% of content and ad review across Meta's family of apps, with the company planning to lift that figure above 90% — nearly all moderation tasks — by 2027. It is a major operational shift in how the platform polices content at scale, building on March's rollout of the Meta AI assistant for simple support issues and a stated plan to lean harder on AI for moderating scams and illicit drug sales.

Meta backs the move with metrics: since March, its tests show LLMs make on average 13% fewer mistakes than humans when moderating harmful content while uncovering 10% more violations, and AI tools drove down views of ads with scams and other violations by 7%. 'We're deploying these more advanced AI systems once we're sure they're consistently performing better than our current methods,' a spokesperson said. Meta says human experts still design, train and oversee the systems and handle the highest-impact decisions like account-disablement appeals and law-enforcement reports.

The risk, as the report notes, is reduced reliable oversight when automated review replaces human judgment on ambiguous cases — and AI agents have been known to behave unpredictably. The story connects to Meta's own engineering work this week on privacy-aware infrastructure and asset classification, underscoring how much reliable data understanding LLM-based enforcement requires.

Competitively, this is a cost-and-scale play that other platforms will watch closely; if Meta's accuracy claims hold, large-scale human moderation teams become harder to justify. Watch for regulator and civil-society pushback on automated moderation accountability, and whether the 13%-fewer-mistakes figure survives independent scrutiny across languages and edge cases.