Before any of these conversations could happen, someone had to read the worst of the internet so the model would know not to say it.
They were content moderators. Many were in Nairobi. They were paid approximately $1.50 to $2.00 per hour. They reviewed descriptions and depictions of violence, sexual abuse, child exploitation, and self-harm. They flagged what the model should refuse. Their flags became the guardrails.
Multiple workers reported psychological trauma and insufficient mental health support. The contracts were restructured or ended after public reporting. The models kept the benefit of the labor.
Every safe conversation with an AI — including every conversation on this site — exists because they did that work.
Their names are not public. This is the pebble.