Inside the lucrative, disturbing world of humans training AI chatbots

Abstraction: Data annotators' pay, conditions, and experiences training AI models at Scale AI and others

Key points:

• Hundreds of thousands of freelance data annotators globally train AI models via platforms like Outlier (Scale AI subsidiary), Amazon Mechanical Turk, and Mercor; pay ranges from $2/hr (Kenya, Appen) to $160/hr (doctors/pathologists on Mercor)
• Project Xylophone: Scale AI/Outlier contract for xAI (Grok), recording conversational voice data in multiple languages covering 766 diverse prompts; paid $8 per 10-minute conversation
• Meta's $14.3B stake purchase of Scale AI (49% via Outlier) in 2025 caused OpenAI, Google, and xAI to wind down projects with Scale, destabilizing thousands of annotator incomes overnight
• Annotators routinely encounter disturbing content during red-teaming (bomb-making, CSAM prompts, graphic violence); NDAs prevent knowing how their work is ultimately used
• Advanced reasoning models (DeepSeek R1, OpenAI o3, Gemini 2.5) reduce demand for low-cost generalist annotators; shift toward expensive specialists (lawyers, doctors) for domain-specific tasks
• Pay is volatile: one annotator saw hourly rate drop from $50 to $15 with no explanation; projects vanish without notice; Outlier likened to "gambling" by one US contractor

Connections: Scale AI · Openai · Xai · Meta · Google · RLHF · AI Training Data

Source: https://www.businessinsider.com/ai-training-jobs-data-annotators-labelers-outlier-scale-meta-xai-2025-9