The people grading Copilot's images are reading your private photosHumans Are Reading Copilot Prompts — And They're Horrified

Microsoft pays hundreds of contractors to score Copilot's generated images, and what they see all day is the photos users assumed only the AI would see, plus a flood of explicit and sometimes illegal edit requests.

You assume the photo you upload to Copilot stays between you and the AI. In practice at least several hundred human contractors are looking at them, one by one.

Their job is picking the better-looking one

A user types a prompt and uploads an image; Copilot produces two edits. The contractor picks the better one: did it follow the instruction, did it leave untouched what should stay untouched, are there distortions or unnatural textures.

Microsoft's instructions tell them to trust their gut, saying human eyes catch subtle quality differences a model cannot articulate. The trouble is how many of those prompts are explicit.

Nobody flags this as distressing content

These contractors are not moderators. They do not filter illegal images; they rate quality. So the material just sits in front of them: shortening a real woman's skirt, enlarging her breasts, upskirt photos.

One contractor hit a set of eight upskirt photos and had to request an unsafe-content flag himself. Another saw what looked like animal sacrifice. A third reported prompts that appeared to sexualise a group of young girls.

Human review is old news, the position is new

Social networks have hired people to police posts for years. What differs is that the material here is chat logs and selfies users treat as private, and the point is not to teach the model to block things but to make its answers friendlier and clearer.

One firm hiring for this work states on its site that with generative AI you cannot fully control what participants are shown. Microsoft's response was a single line: we use customer data as described in our terms of use.