The Human Labor Behind Embodied AI Training Data
A 钛媒体 investigation by 产联社CLS examines the human labor used to collect embodied AI training data through offline centers, home-based crowdsourcing and family rental programs. Workers are typically paid by approved usable duration rather than raw hours, while quality-control failures and rework can substantially reduce usable output; the report also documents under-disclosed quotas, long payment cycles and privacy risks involving household footage and biometric data. Lightwheel Intelligence said about 800–1,600 collectors work daily on its network, while Galaxy General reported more than 1 million hours of human-action data across multiple sources.
The investigation shows that scaling embodied AI depends on a largely hidden labor and governance system in which effective-data accounting, quality assurance, worker disclosure and personal-data authorization determine both training-data supply and compliance risk.
(本文作者为 产联社CLS,钛媒体经授权发布)
文 | 产联社CLS
早上9点,35岁的陈萍在五道口附近的一间办公室扫码签到。她的工作在招聘信息里写得简单:早9点到晚6点,日薪180元。走进办公室,她看见几排长桌,数十名采集员,每人一台笔记本电脑、两个腕带式手环和一个头戴设备。她要做的,是跟着一段视频,反复完成标准化的手部动作。
中途她被叫去做体测:先测血压,再站上一台机器测量身体数据。起初准确率不高,有人过来指导。直到听说把手环挪到上臂肉多的地方,准确率能到100%,她才过了关。
下班前她才弄明白,这份工作“不仅要看工作时长,还要看采集的数量”。此前没有人告诉过她这一点,她还误删过几条数据。为了做够数量、不被扣钱,她加班40分钟才扫码签退。
42岁的“锦姐”已经居家做了10天同类工作。名义时薪25元,但真正计入收入的是平台审核确认的“有效时长”:每天开机约8小时,按约70%的有效率折算成5小时,日收入约125元。头三天她不知道这套规则,以为录制时长全都算数,有效率只有40%,“等于白忙活大半天”。
#Lightwheel Intelligence#Galaxy General#JD#Songyan Power#Mifeng Technology#China Academy of Information and Communications Technology#YDT 6771