Meta is rapidly shifting content moderation work from human reviewers to large language models, a move the company presents as a quality upgrade but some employees see as a risk to users, oversight and moderation jobs.
The company has already replaced roughly half of all human moderation requests with large language models in 2025. For some content types, Meta plans to push that share above 90 percent by the end of the year.
Meta says AI moderation is improving enforcement
According to the Financial Times, the shift is expected to save the company billions annually. Meta disputes the cost argument and says the reason for the change is quality, not savings.
The company says that since March, tests show its language models make 13 percent fewer errors than humans when enforcing content policies. Meta also says the models catch 10 percent more actual violations.
That framing matters because content moderation is not only about removing rule-breaking material. It is also about avoiding unnecessary removals, understanding context and applying policy consistently across many kinds of posts.
Meta argues that large language models offer advantages over traditional ML classifiers. Traditional systems can struggle with satire or evolving language, while the newer models are supposed to better grasp nuance and cover more languages.
Employees warn the rollout is moving too quickly
Employees describe a different picture. One insider says the models still remove or shadow-ban harmless content. The concern is not simply that mistakes happen, but that there is not enough oversight for a transition happening at this speed.
That warning points to the central tension in Meta's AI moderation push. The company is relying on test results that suggest fewer errors and more detected violations. Some workers, however, are focused on what happens when the system makes mistakes at scale.
In moderation, an error can affect whether content is visible, limited or removed. If harmless content is removed or shadow-banned, the user may not see the same issue Meta sees in its internal quality metrics.
The source does not describe every type of content affected, but it does state that Meta wants AI to handle above 90 percent for some content types by the end of the year. That makes oversight a practical concern, because a larger automated share means more decisions depend on the model's judgment.
The change is already affecting moderation workers
The transition is already leading to layoffs, especially among external contractors. That detail underlines how a technical rollout can quickly become an employment issue.
Human reviewers have been central to the decision history that the models now learn from. The models are trained on past decisions made by human reviewers, which means the new system is built partly on the work of the people it may replace.
This creates a complicated handoff. Meta is using prior human moderation decisions as training material while reducing the role of human review in current requests. Employees warning about oversight are therefore raising a question about how much human judgment should remain in the loop as AI takes on more of the work.
The company and its employees are not only debating whether AI can moderate content. They are debating how fast that responsibility should move away from people, and what safeguards are needed when automated systems become the default.
Meta is also switching the model behind the system
There is another change happening behind the scenes. The Financial Times reports that Meta had been using Google's Gemini for moderation and support but recently told staff to switch to its own new foundation model called Muse Spark.
That model swap adds another layer to the rollout. Meta is not only increasing the share of moderation requests handled by large language models. It is also moving staff from an outside model to an internal foundation model.
The source does not provide technical details about Muse Spark. What it does make clear is that the shift is part of a broader move to make language models a much larger part of moderation and support workflows.
What this means for AI moderation
The facts in the source show a sharp divide between Meta's public quality argument and employee concerns from inside the rollout. Meta says its tests show fewer errors and more violations caught. Employees warn that harmless content is still being removed or shadow-banned and that oversight is not keeping pace.
Both points can shape how this transition is judged. If Meta's language models consistently improve enforcement, the company can argue that AI moderation is better suited to nuance, satire, evolving language and more languages than older classifiers. If employee warnings prove correct, the speed of the rollout may become the problem, even if the technology performs well in tests.
For now, the direction is clear. Meta has already moved roughly half of all human moderation requests to large language models in 2025, is targeting above 90 percent for some content types by the end of the year, and is replacing Google's Gemini with Muse Spark for moderation and support. The debate is over whether that shift is being governed carefully enough.