Meta replaces third-party content moderators with AI across Facebook and Instagram

Meta is cutting its third-party content moderation contracts and handing the work to AI systems built in-house. The shift affects Facebook and Instagram, where external vendor teams have handled a significant portion of human content review for years. Mark Zuckerberg framed the move as a cost and consistency play, but the practical consequences for users, regulators, and the workers who lose those contracts are considerably more complicated than that framing suggests.

Meta spent approximately $5 billion on safety and security in 2023, a figure that includes both internal staff and external vendor payments. Content moderation vendor contracts, which cover companies like Accenture, Majorel, and Teleperformance, account for a substantial portion of that spend. Replacing those contracts with AI systems reduces recurring operating costs, and at Meta's scale, even a 20 percent reduction in that budget line translates to hundreds of millions of dollars annually.

How Meta's AI moderation actually works

Meta's AI moderation systems use a combination of classifiers trained on previously reviewed content, image and video hashing to detect known harmful material, and large language models to assess context in text-based posts. The systems are not making decisions in the way a human reviewer does. They are assigning probability scores that determine whether content gets removed automatically, sent to a human reviewer, or left up.

The company has been moving in this direction for several years. In its Q4 2024 transparency report, Meta stated that AI systems proactively detected 97.3 percent of the hate speech removed from Facebook before any user reported it. That sounds impressive until you consider that the denominator is only the content Meta removed, not the total amount of harmful content on the platform, which is impossible to measure precisely. The 97.3 percent figure tells you the AI is good at finding content that looks like what it was trained on. It says nothing about what it misses.

What happens to the moderation workforce

Content moderation at Meta's scale has been largely outsourced to workers in countries including the Philippines, Kenya, India, and Colombia, where labor costs are significantly lower than in the US or Europe. Teleperformance, one of Meta's largest moderation vendors, employs tens of thousands of workers globally on content review contracts. A reduction in Meta's contract volume has direct employment consequences for those workers, many of whom handle graphic and traumatic content as a routine part of their jobs.

A 2022 lawsuit filed by former Kenyan content moderators against Sama, another Meta vendor, alleged that moderators were exposed to extremely disturbing content without adequate psychological support and were paid approximately $1.50 per hour. The case was settled out of court. Meta's shift to AI reduces its exposure to those ongoing labor and liability concerns, which is a business benefit that does not feature prominently in the company's public messaging about the change.

The accuracy problem that AI moderation has not solved

AI content moderation consistently struggles with context-dependent speech. Satire, irony, cultural references, and code-switching between languages and dialects are all categories where automated systems perform significantly worse than experienced human reviewers. A 2023 study published in the journal Proceedings of the ACM on Human-Computer Interaction found that AI moderation systems made incorrect decisions on ambiguous content at roughly three times the rate of trained human reviewers, though they were faster and cheaper per decision.

Over-removal is also a documented problem. Facebook has repeatedly removed legitimate journalism, political commentary, and public health information because AI classifiers flagged surface features of the content without understanding its purpose. The appeals process for wrongly removed content is slow, and for small publishers or individual creators, a wrongful removal can mean days or weeks without access to an audience while the appeal is processed.

Regulatory exposure in Europe and Australia

The European Union's Digital Services Act requires very large online platforms to conduct annual independent audits of their content moderation systems and to maintain human review capacity for appeals. Meta is classified as a very large online platform under the DSA, which means replacing human moderation with AI does not exempt it from accountability requirements. It creates new ones. The European Commission has already opened formal proceedings against Meta under the DSA related to algorithmic recommender systems, and the moderation shift will attract additional scrutiny.

Australia's eSafety Commissioner has separate authority to issue compliance notices to platforms that fail to take down harmful content within specified timeframes. If Meta's AI systems have higher error rates or slower response times on certain categories of content, Australian regulators have both the legal basis and the stated appetite to act. eSafety Commissioner Julie Inman Grant said in February 2025 that her office was monitoring Meta's moderation changes closely and would not accept AI inadequacy as a compliance defense.

What Zuckerberg has said publicly about the shift

Zuckerberg announced in January 2025 that Meta was ending its third-party fact-checking program in the United States and replacing it with a community notes system modeled on X's approach. The shift to AI moderation for content removal decisions is a separate but related move that happened with considerably less public announcement. Zuckerberg's stated rationale for the broader direction is that human moderation has become too ideologically inconsistent and that AI will produce more uniform enforcement of the same rules.

The consistency argument has surface logic. A classifier trained on a fixed policy will apply that policy the same way every time, whereas human reviewers bring their own cultural context and judgment to ambiguous cases. The counterargument is that consistency in applying a flawed policy at scale just produces more consistent errors, and that the cultural judgment human reviewers bring is often what catches the cases the AI was not trained to handle.

What this means for advertisers and brand safety

Advertisers pay a significant premium to avoid having their ads appear next to harmful or controversial content. Meta's brand safety tools rely on content classification to ensure ad placement exclusions work correctly. If AI moderation produces higher rates of misclassified content, ads that should be excluded from certain content categories may not be, and advertisers have contractual grounds to demand refunds or pull spending.

The Global Alliance for Responsible Media, a cross-industry body that includes major advertisers and media agencies, issued a statement in March 2025 saying it expected Meta to provide transparent performance data on its AI moderation systems within 90 days of full deployment, and that member organizations would review their Meta spending commitments pending that disclosure.