Microsoft pays humans to check whether Copilot is behaving, and it turns out some users have very different priorities.
According to a report from 404 Media, human contractors review real prompts sent to Microsoft's Copilot assistant, checking whether its responses are accurate and appropriate. That job apparently includes reading requests that have nothing to do with debugging code or drafting emails. A number of users have asked Copilot to generate sexual imagery, according to the report. The contractors' task is quality control, not content moderation, but the two overlap the moment someone tries to get a work assistant to draw something it was never built for.
This isn't shocking so much as confirming: any chatbot with a human review pipeline has actual people, not just algorithms, looking at what you type. For a tool Microsoft markets as a productivity assistant, that raises awkward questions about what ends up in its review queue and who has to sort through it.
It's a reminder that plenty of AI work still runs through human eyes, and that some fraction of any large user base will test the boundaries, whether the product was built for that or not.