Humans Are Reading Copilot Prompts – and They're Horrified
9 hours ago
- Human contractors hired to improve Microsoft's Copilot AI are exposed to lewd, sexually explicit, and potentially illegal image editing requests, including upskirt photos and nonconsensual sexual positions.
- Contractors review user-uploaded images and prompts, assuming privacy, but hundreds of human workers see them to evaluate AI-generated edits for quality, not for safety or content moderation.
- Internal documents reveal contractors encounter foot fetish images of children's cartoon characters, pro-anorexia content, and prompts for nude images of minors, with no explicit flagging for offensive content.
- Microsoft instructs contractors to rely on human intuition to select better AI edits, focusing on adherence to edit instructions, preservation of unchanged areas, absence of artifacts, and overall quality.
- This practice extends to OpenAI and other AI companies, where contractors review real user prompts, differing from traditional content moderation as the goal is to improve chatbot responses rather than filter offensive material.