Meta Shifts to In-House AI for Content Moderation, Ditches Third-Party Vendors

Meta has announced the implementation of new AI systems for moderation and enforcement on its platforms — Facebook, Instagram, and Threads. At the same time, the company is reducing its dependence on third-party contractors who previously performed a significant portion of violation detection work. According to the company, its own AI tools outperform previous approaches on several key metrics.

The new systems detect rule violations more accurately, block fraudulent schemes more effectively, respond faster to real-time events, and produce fewer false positives — meaning they are less likely to block legitimate content by mistake. This step fits into the large-scale transformation Meta has been conducting since the beginning of 2025. In January, the company announced the closure of its independent fact-checking program in the United States and a transition to the Community Notes model — user-generated contextual notes, following the example of X.

At the same time, Meta relaxed a number of policies regarding political content and adjusted how algorithmic recommendations work. The transition from third-party vendors to in-house systems is a logical continuation of this course. External contractors in the moderation sphere have always been a weak link: they worked more slowly, cost more, and did not have direct access to the platform's internal signals.

It is precisely these signals that allow Meta's systems to detect patterns of violations significantly earlier than an external review team could. Response speed to real-world events is one of the critical metrics for platforms with a billion-user audience. During natural disasters, elections, military conflicts, and other crises, the volume of potentially harmful content increases sharply.

AI systems with continuous access to real-time data streams objectively outperform external review teams precisely in such peak situations. The stated goal of reducing overenforcement — excessive blocking — deserves special attention. This has long been a pain point for Meta: cases of removing news content, journalistic publications, medical materials, and statements from minorities regularly became grounds for criticism of the company.

The new systems, according to Meta representatives, should distinguish more accurately between actual violations and edge cases. At the same time, a complete transition to automated moderation carries its own risks. AI systems are trained on historical data and can reproduce biases embedded in the training datasets.

Without a transparent appeals mechanism and independent audit, users and regulators will be unable to assess how fairly these systems work in practice. European regulators already require major platforms to disclose information about moderation systems and provide an appeals mechanism — under the Digital Services Act (DSA). There is no American equivalent to these requirements yet, but regulatory pressure worldwide continues to mount.

For Meta, the transition to in-house AI systems is simultaneously a technological and reputational choice. Technologically, the company gains a more flexible and scalable tool that can be adapted and retrained for specific tasks. Reputationally, it demonstrates that it is taking direct control of key infrastructure on its platforms and not outsourcing it.

How exactly the new systems are structured, what models underlie them, and how human oversight is organized — Meta has not yet publicly disclosed. It is precisely the answers to these questions that will determine whether the updated approach represents a real step forward in content safety or another round of opaque management of the world's largest social platforms.