A 404 Media investigation, based on internal contractor documents, reveals that Microsoft's Copilot AI image editing is not as private as users might assume. A workforce of at least hundreds of human reviewers is shown the original prompt, the uploaded picture, and two AI-generated edits. Their job is to judge which edit better follows the instruction, preserves the original image, and looks natural.
The prompts and images reviewers encounter are often sexual or disturbing. Contractors reported requests to shorten skirts, enlarge breasts, and put people into sexual positions. Some saw upskirt photos, pro-anorexia content, or prompts involving children's cartoon characters. One contractor described being asked to create an image of Ariana Grande with anorexia. Another said they recoiled at the material.
These workers are not moderators flagging harmful content. According to the documents, Microsoft wants their human intuition to help improve Copilot's responses. The instructions tell contractors to trust their gut reaction when comparing two edits. This differs from earlier reports about OpenAI's contractors, who also review real prompts, but the key point here is that the goal is quality, not safety. Users who upload private images likely do not expect them to be seen by human reviewers, yet that is exactly what is happening.