The intelligence is artificial. The labor is real.

Every major AI system relies on human feedback reinforcement learning (RLHF) — a process where contractors rate AI outputs to teach models what “good” responses look like. These workers are the invisible infrastructure behind frontier AI.

We spent two months investigating the RLHF workforce. What we found reveals uncomfortable truths about how AI is actually built.

The Scale

OpenAI, Anthropic, and Google collectively employ an estimated 500,000-1,000,000 contractors for data work:

  • Data labeling — Drawing bounding boxes, categorizing content, transcribing audio
  • Preference ranking — Comparing two AI outputs and selecting the better one
  • Safety evaluation — Attempting to make models produce harmful outputs
  • Fact-checking — Verifying claims in model-generated text

This workforce is larger than the combined employee count of the major AI labs.

The Working Conditions

RLHF contractors work through third-party vendors — Scale AI, Appen, Telus International — not directly for AI companies. This creates distance and plausible deniability.

Typical working conditions:

  • Pay: $1-5 per hour in Kenya, Philippines, India; $15-25 in US/Europe
  • Hours: Flexible but often 40-60 hours/week to meet targets
  • Psychological impact: Exposure to disturbing content flagged for safety review
  • Job security: None — contractors are easily replaced

The work is repetitive, poorly paid, and psychologically taxing. Attrition rates exceed 50% annually.

The Hidden Trauma

Safety evaluation contractors are specifically tasked with eliciting harmful outputs from AI systems:

  • Instructions for violence, self-harm, and illegal activities
  • Graphic sexual content and CSAM
  • Hate speech and extremist material
  • Detailed instructions for weapons and explosives

These workers spend their days trying to make AI systems produce exactly the content the models are supposed to refuse. The psychological toll is significant and largely unaddressed.

One contractor described the work: “I go home and can’t stop thinking about the things I’ve read. The AI is protected by filters. I’m not.”

The Quality Problem

RLHF depends on contractor judgment — but contractor incentives don’t always align with quality:

Speed over accuracy — Pay is often per-task, encouraging rapid rather than careful evaluation Lack of expertise — Contractors rate outputs on technical subjects they don’t understand Cultural bias — Predominantly Western contractors impose Western values on global models Inconsistent standards — Different contractors apply different criteria for “good” responses

These issues are known within AI labs. They’re accepted as necessary compromises for scaling training.

The Economic Reality

RLHF costs are significant but hidden. OpenAI reportedly spends $100M+ annually on contractor services. Anthropic and Google are likely similar.

This spending doesn’t appear in most AI economic analyses, which focus on compute costs. But human feedback is the differentiator that turns raw models into products.

The economic model depends on cheap labor in developing countries. If Kenya or the Philippines required $15/hour wages, RLHF economics would collapse.

The Contradiction

AI companies promote their technology as eliminating drudgery and empowering workers. Simultaneously, they rely on massive low-wage workforces performing the digital equivalent of factory assembly lines.

The contradiction isn’t acknowledged. Marketing materials celebrate AI autonomy while operations depend on human labor at every step:

  • Data labeling requires humans
  • RLHF requires humans
  • Safety testing requires humans
  • Output moderation requires humans

The “artificial” in artificial intelligence obscures the human effort that makes it work.

The Future

Several trends may change the RLHF landscape:

Automation — AI systems are increasingly used to pre-label data and filter obvious cases, reducing human workload Regulatory scrutiny — EU AI Act and similar legislation may require disclosure of training labor practices Unionization — Contractors in some regions are organizing for better pay and conditions Geographic shifts — As wages rise in current contractor hubs, work moves to lower-cost regions

But the fundamental dependency remains. Current AI systems cannot train themselves. They require human judgment to learn human preferences.

That judgment comes from workers earning poverty wages to enable trillion-dollar valuations.

The Question

As AI capabilities advance, we should ask: who benefits from the intelligence being built?

The technology promises liberation from drudgery. But the drudgery hasn’t disappeared — it’s been outsourced, hidden, and paid less.

The AI revolution is built on invisible labor. That should trouble anyone who believes technological progress should benefit everyone.