Meta员工警告AI内容审核部署过快

Meta employees warn AI moderation rollout is too fast

AI 导读

Meta在2025年已用大语言模型替换约一半人工审核请求，计划年底前将部分内容类型的AI审核比例提升至90%以上，每年节省数十亿美元。Meta否认成本动机，称自3月测试显示其模型错误率比人类低13%，且多捕捉10%违规。但员工指出模型仍会移除或限流无害内容，缺乏足够监督，快速部署已导致外包裁员。此外，Meta已从使用Google Gemini转向自家新基础模型Muse Spark，该模型基于人工审核员的历史决策训练。

推荐理由

Meta用LLM替换人类审核已过半，年底目标90%以上，员工警告误删和裁员同步发生。AI审核的规模跃迁背后，质量控制争议可能成为社会信任的节点。

正文 · 原文

Meta has already replaced roughly half of all human moderation requests with large language models in 2025 and plans to push that share above 90 percent for some content types by the end of the year. The shift is expected to save the company billions annually, according to the Financial Times. Meta disputes the cost argument and points to quality instead, saying that since March, tests show its language models make 13 percent fewer errors than humans when enforcing content policies while catching 10 percent more actual violations. Unlike traditional ML classifiers that struggle with satire or evolving language, the language models are supposed to better grasp nuance and cover more languages.

Employees paint a different picture. One insider says the models still remove or shadow-ban harmless content, and there isn't enough oversight for such a rapid rollout. The transition is already leading to layoffs, especially among external contractors.

There's also a model swap happening behind the scenes, the Financial Times reports. 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. The models are trained on past decisions made by human reviewers.