Meta AI system is now responsible for half of all moderations

Half of all content moderation at Meta is currently automated via language models. Already in 2025, Meta has achieved significant AI-driven moderation: half of all content moderation decisions are now made by AI rather than humans. And they’re not done yet. By the end of 2026, Meta will have exceeded 90% AI moderation of certain types of content. The company behind billions of daily interactions is letting AI take over content moderation.

Meta’s data supports this strategy. They claim that their AI models are 13% less likely to make mistakes and 10% more likely to detect real violations than human moderators, based on tests starting in March. Language models excel at handling nuance, such as sarcasm, changing slang, and contextual meanings, which are difficult for rigid classification algorithms to process. They’ve also switched from using Google’s Gemini model to their internally developed Muse Spark model.

1Employees are sounding the alarm

But this is Meta reporting on itself. Employees told reporters that AI is being deployed too fast, without enough time to identify potential issues before reaching actual users. Their models continue to erroneously remove or shadow ban content that doesn’t violate any rules. On a platform with billions of users, even a slight misclassification rate can have a major impact on millions of users whose posts get removed for no apparent reason.

And there’s another reason why Meta doesn’t want to talk about this much. Using AI instead of humans for moderation will save them billions of dollars each year and they’re already ramping up layoffs of contractors who perform this work. They’re telling a story about improving quality while analysts are seeing it as a way to cut costs. It’s both.

2What it means when AI becomes the judge

This story matters because it’s an example of what happens when we put AI in control of important decisions that impact the speech of billions of people. Content moderation mistakes don’t just inconvenience users; they silence activists, delete satire, and erase history. With human moderators, there are procedures in place to appeal decisions and correct errors. But with AI models making systemic mistakes at a 90% deployment rate, the feedback mechanisms are different: slower, more difficult to detect, and broader in scope.

If you’re developing AI products, this story shows you how the difference between measured accuracy and real-world usage can be substantial. A model that’s 13% more accurate overall may still be making serious mistakes in certain categories. If those categories are large enough, then “certain categories” becomes “millions of decisions per day.” Best practices include gradual deployment, error tracking with human oversight, and mechanisms for correction. Those are the things that Meta’s employees are asking for.