Meta employees have reported concerns that the company is rushing its deployment of artificial intelligence for content moderation. Reports indicate that roughly half of all human moderation requests were replaced by large language models in 2025, with plans to exceed 90 per cent coverage for certain content types by year end. Internal staff warn that the system still removes or shadow-bans harmless material while lacking sufficient oversight for such a rapid transition. This shift has already triggered layoffs among external contractors who previously handled these tasks.
The decision to prioritise speed over caution stems from a belief that automated systems can reduce errors and catch more violations than people. Since March, tests suggest these models make 13 per cent fewer mistakes while identifying 10 per cent more actual breaches. Meta also swapped Google’s Gemini for its own foundation model, Muse Spark, which is trained on past decisions made by human reviewers. This change aims to handle nuance and evolving language better than traditional classifiers. The move represents a significant reduction in human involvement for a core platform function.
* Current model share sits at 50 per cent
* Layoffs affect external contractors primarily
* New system trained on historical human decisions
