Work

Permanent Judgment, Temporary Labor: Human Feedback Work, Automated Substitution, and the Law of Platform Exit in the Global South

Year: 2026
Type: preprint
Abstract: : Human feedback work builds the machine that replaces the rater. Across Kenya and the Philippines, generalist annotators rank and label model outputs at piece rates. Those comparisons train reward models, which then train AI judges to score outputs against written safety policies. Human raters disappear from that stage. The *paradox of automation's last mile* sorts the remaining human work by credential and geography. Expert evaluation pays hundreds of dollars an hour in wealthy economies, while Remotasks cut Kenyan, Nigerian, and Pakistani workers in 2024 and Outlier still bars them. Domestic labor statutes do not reach those exits. Philippine and Kenyan law require advance notice and separation pay for redundancies and labor-saving devices, yet platform terms classify annotators as independent contractors outside both statutes. Three legal instruments close that gap. Article 298 of the Philippine Labor Code and the OECD disengagement standard supply notice and mitigation duties. ILO Convention No. 193 extends those protections to independent contractors. Keywords: RLHF Labor, Reinforcement Learning from Human Feedback, Reinforcement Learning from AI Feedback, RLAIF, LLM-as-a-Judge, AI Data Annotation, Platform Disengagement, Independent Contractor Classification, Global South Labor, Kenya Employment Act, Philippine Labor Code Article 298, Labor-Saving Devices, ILO Convention No. 193, OECD Guidelines for Multinational Enterprises, Responsible Disengagement, Remotasks, Outlier, Sama, Scale AI, Data Colonialism, Ghost Work, Algorithmic Management
Language: English
Open Access status: green
OpenAlex ID: W7233467747
Cites:
Cited by:
Related to: 10