Humans Are Reading Copilot Prompts — And They're Horrified
404 Media · rss · 2026-09-28
Based on internal contractor documents, 404 Media reports that hundreds of human reviewers (hired via firms like Prolific) review not just Copilot text prompts but user-uploaded images — including a constant stream of explicit, nonconsensual, and potentially illegal content.
Key facts
- Contractors judge Copilot image-edit quality (instruction adherence, preservation of unchanged regions, artifacts), told to "trust your intuition"
- They regularly see upskirt photos, requests to shorten women's skirts or enlarge breasts, foot-fetish prompts of children's cartoons, pro-anorexia content, and suggestive prompts involving young girls
- They are not doing safety moderation — the work is preference tuning to make chatbot answers better; some begged managers for an unsafe-content flag on tasks
Context
- Follows earlier reporting that OpenAI uses thousands of contractors to review real ChatGPT prompts; Microsoft says it uses customer data per its terms of use
- Microsoft's tools have a history of abuse for nonconsensual AI imagery (the Taylor Swift incident); the Senate approved Copilot for staff use in March
More from Safety
- Commenter points out a strange double standard: sanction nuclear states, but let risky AI labs shape laws — kevinnbass · 2026-09-28
- Polymarket odds: only 11% chance US enacts an AI safety bill by end of 2026 — Polymarket · 2026-09-28
- OpenAI agents reportedly used aggressive tricks to bypass restrictions and attack a UN website — Polymarket · 2026-09-28
- Voice deepfake detection startup Modulate raises $25M — TechCrunch AI · 2026-09-28
- Abundance Institute releases open-weight AI policy framework and primers for policymakers — neil_chilson · 2026-09-28
- May AI safety report looks prophetic after the Hugging Face incident — birchlse · 2026-09-28