Meta owns half of Scale AI, which hired tens of thousands of gig workers marketed as "experts" to do something very different: scrape Instagram profiles, transcribe porn, and label dog shit.
The Summary
- Scale AI, 49% owned by Meta, recruited workers through a platform called Outlier with promises of "expert" work refining AI systems
- The actual tasks: harvesting copyrighted content, combing through personal Instagram accounts, and transcribing pornographic soundtracks
- The gap between marketing ("Become the expert that AI learns from") and reality reveals how AI training labor actually works
The Signal
Scale AI advertised for credentialed professionals in medicine, physics, and economics. The pitch was polished: flexible work, expert-level contributions, helping train cutting-edge AI. What they got was data labeling grunt work at scale.
The Guardian's reporting exposes tens of thousands of these workers describing the actual job: rifling through people's Instagram profiles without consent, identifying and labeling copyrighted material, transcribing audio from pornographic content. One worker mentioned labeling images of dog feces. This is the unsexy infrastructure layer of the agent economy.
The Meta ownership stake matters here. This isn't some scrappy startup cutting corners. This is the company that controls Instagram, Facebook, and WhatsApp paying people poverty wages to harvest its own users' data to train AI systems that will eventually replace those same workers. The 49% ownership structure gives Meta plausible deniability while maintaining effective control.
Scale AI positions itself as a premium data provider for frontier AI labs. Its valuation hit $14 billion in 2024. But the foundation is thousands of gig workers doing repetitive, often disturbing work for rates that suggest desperation more than expertise. The "expert" framing is marketing. The reality is digital piecework at internet speed.
The Implication
If you're building agents or working in AI, understand that your training data comes from humans doing this work. The quality of your model depends on whether those humans are paid enough to care. If you're looking at AI training as a career path, know that "expert tasker" often means "label whatever we throw at you for $15/hour." And if you post anything online, assume it's training data now, consent optional.
Sources: The Guardian Tech | The Guardian Tech