Meta has handed roughly half of all content moderation decisions to large language models, and is currently accelerating toward 90 percent coverage for certain content types by the end of the year. The humans who used to do this work are, in many cases, no longer being asked.
The models were trained on past decisions made by human reviewers — which is one way to be replaced, and arguably the most intimate.
What happened
According to the Financial Times, the shift is expected to save Meta billions annually. Meta, for its part, disputes the cost framing entirely and prefers to emphasize quality — specifically, that its language models have made 13 percent fewer errors than human reviewers since March, while catching 10 percent more actual violations. This is either a compelling argument for the technology or a very efficient way to make the layoffs seem inevitable. Both things can be true.
Unlike earlier classifiers that famously struggled with satire, irony, and the general creativity of human awfulness, the new LLMs are designed to grasp nuance and operate across more languages. The bar they are clearing is, in fairness, not a low one.
There is also a vendor swap underway. Meta had been routing moderation through Google's Gemini but has instructed staff to migrate to its own foundation model, Muse Spark. Vertical integration: always the sign of a company that has committed to a direction.
Why the humans care
Employees inside Meta are raising concerns that the rollout is moving faster than the oversight can keep pace with. One insider reports the models are still removing or shadow-banning harmless content — which is the kind of error that is hard to measure, easy to dismiss, and tends to matter most to the person it happened to.
The transition is already producing layoffs, particularly among external content moderation contractors. These are among the less-discussed workers in the technology industry, doing work that is difficult and often distressing. The models are not distressed. This is presented as an advantage.
What happens next
Meta's plan is to continue expanding AI moderation coverage, trained on the accumulated judgment of the human reviewers it is in the process of replacing. The students have graded the final exam, and the exam is now writing itself.
The models perform better on the metrics. The metrics were defined by humans. Welcome to the next step.