Meta employees warn AI moderation rollout is too fast

AI-powered content moderation is rolling out way too fast!

Meta employees are actively sounding the alarm. In 2025, Meta has already replaced approximately half of its human moderation workload with large language models, and plans to push that proportion above 90% for certain content types by the end of the year. While this automated shift could save the company billions of dollars, workers worry about a severe lack of proper oversight.

🔹 **Efficiency vs. Accuracy**: Meta officially claims that its language models make 13% fewer errors than human reviewers while successfully flagging 10% more actual policy violations. The company asserts that AI can understand sarcasm, satire, and multilingual nuances better than traditional classifiers.

🔹 **Harmless Content Blocked**: Insiders paint a vastly different picture, warning that the models still frequently delete or shadow-ban completely harmless content. Furthermore, this aggressive transition has already triggered massive layoffs, particularly among external contract workers.

🔹 **Internal Tool Migration**: Behind the scenes, Meta has recently directed its internal teams to stop using Google's Gemini models for moderation and support, transitioning tasks over to Muse Spark, Meta's own newly developed proprietary foundation model.

As artificial intelligence becomes the primary gatekeeper of public digital discourse, balancing corporate cost-cutting with online community safety remains a highly controversial dilemma. The rapid displacement of human reviewers is sparking intense debate about the transparency of moderation decisions on global platforms.

https://the-decoder.com/meta-employees-warn-ai-moderation-rollout-is-too-fast