Meta cuts external moderation and shifts content review to more advanced AI

Meta is significantly reducing the role of external contractors in moderating Facebook and Instagram. The company wants to transfer more content checks to its own AI systems, leaving humans with the most risky and controversial decisions.

How Meta's Moderation Is Changing

Meta stated that over the coming years it will gradually deploy more advanced AI systems across all its applications. This is not just about filtering spam, but about fully applying rules to content: the company wants to more accurately find and faster remove serious rule violations — fraud, illegal content, drug sales content, and other dangerous categories. If new systems consistently show better results than current methods, Meta will reduce its dependence on third-party contractors and strengthen its own internal teams and tools.

This is a notable shift for Meta's entire operating model. Previously, the company spent years combining automated violation detection with the work of thousands of external moderators who manually reviewed disputed posts and complaints. Now the bet is that modern models will handle bulk, repetitive, and rapidly changing work better, where it's hard for humans to keep pace. For Meta, this is simultaneously a question of speed, scale, and cost of moderation on platforms with billions of pieces of content.

What AI Can Already Do

The company claims that early tests of new systems look strong enough to expand their role. Rather than general promises about efficiency gains, Meta shows specific metrics on fraud, fake accounts, sexual violations, and suspicious ads. The company builds its case for transferring part of the work previously done by external vendors to its own AI on these numbers.

• They find and block approximately 5,000 additional fraudulent attempts per day that existing review teams missed.
• They reduced complaints about accounts impersonating the most frequently imitated celebrities by more than 80%.
• They identify twice as much rule-violating content with sexual propositions aimed at adults while making over 60% fewer errors.
• They spot complex account hacking scenarios where individual signals seem harmless alone, but together indicate an attack.
• After broader testing, they reduced impressions of ads with scams and other serious violations by 7%.

Meta separately emphasizes language coverage. According to the company, the new systems work with languages spoken by 98% of people online, whereas previously coverage was around 80 languages. More important than the number itself is that the models learn to adapt faster to local context: slang, emojis, code words, and behavior of niche communities. Such details often make automated moderation a weak point, especially when bad actors quickly change tactics.

Where Humans Remain

Meta is not planning to remove humans from the chain entirely. The company explicitly states that AI will take on tasks better suited to technology: repetitive checks of graphic content, bulk searches for fraud schemes, and categories where violators constantly adapt to rules and bypass filters. But the most sensitive decisions will remain with humans — for example, appeals after account suspension or cases that could lead to interaction with law enforcement.

Within this scheme, the human role changes rather than disappears. Experts will design, train, test, and evaluate systems, monitor bias and decision quality, and handle complex cases where context matters more than speed. Meta also promises that the community rules themselves will not change due to this transition. In parallel, the company is launching an AI support assistant on Facebook and Instagram so users can more easily understand why content was removed, how to appeal, and what happens next.

What This Means

Meta is moving AI from the role of an auxiliary filter to the core of platform governance. If the transition works, the company will reduce its dependence on contractors and accelerate moderation on a global scale. But the cost of error also increases: when a controversial decision is made by a model rather than a human, any systemic inaccuracy can instantly scale to millions of users. Meta's next test is to prove that such a scheme really produces fewer errors, not just lower costs.