Remote‑Work Automation Overpromise

A new Remote Labor Index test (Scale AI + Center for AI Safety) gave hundreds of real paid freelance tasks to leading AI systems and found the best model fully completed only ~2.5% of assignments, with roughly half producing poor quality or leaving the work incomplete. Failures included corrupt outputs, wrong visual handling, missing data, and brittle memory — concrete limits on current automation capacity. — If replicated, this should temper near‑term job‑elimination narratives, redirect policy toward augmentation, verification standards, and targeted retraining, and shape who bears liability when AI is deployed on real economic tasks.