Microsoft’s contractors are looking at the photos you upload to Copilot, and some don’t like what they see

Hundreds of human contractors working to improve Microsoft’s Copilot are reading users’ image prompts and looking at the photos people upload, 404 Media reported on Monday, citing internal instruction guides, samples of the work and the contractors’ own message boards. Many say they are faced with a steady stream of sexual and disturbing requests.

What the contractors do

The reviewers aren’t moderators. According to the documents 404 Media obtained, their job is to rate how well Copilot carried out an image request, for example by comparing two edited versions of an uploaded photo and picking the better one. Microsoft’s guidance tells them to rely on their “intuition”. Whether a request should have been allowed in the first place is outside their remit.

That leaves them judging the quality of work they find troubling. Contractors described being asked to evaluate edits that made women’s clothing more revealing, sexualised pictures of real women and cartoon characters, pro-anorexia content, and a set of eight “upskirt” photos that one worker asked supervisors to flag as unsafe. Another said a prompt appeared designed to produce a sexualised image of a group of young girls. Workers complained that they often got no warning before a disturbing image appeared on screen.

One contractor summed up the frustration on an internal board:

It’s hard to take things seriously when my focus has to be what model generated the appropriate bust size or which middle-aged woman was put in the appropriate sexually suggestive position.

A Copilot contractor, via 404 Media

“Faces are always uncensored”

The report sits awkwardly next to Microsoft’s own Copilot privacy FAQ. It says images and files you upload are kept for up to 18 months, and that before using them to train its models Microsoft takes steps to de-identify them, “such as removing metadata or other personal data and blurring images of faces”.

One worker told 404 Media: “Faces are always uncensored, and many of the prompts are sexual in nature and dubiously consensual.” It isn’t clear whether Microsoft treats this rating work as model training under its policy, or as the separate “product improvement” review the FAQ also mentions. Microsoft hasn’t commented publicly on the report.

Can you stop humans seeing your Copilot chats?

Not completely. Microsoft’s FAQ says “some Copilot conversations are subject to both automated and human review for product improvement and digital safety purposes”, and that “an opt-out of human review is not available”. What you can control:

• Model training: signed-in users can turn off model training in Copilot’s privacy settings. It only applies to future conversations, and Microsoft says it doesn’t switch off human review.
• Who is excluded anyway: Microsoft says it doesn’t train on conversations from people who aren’t signed in, signed-in users under 18, work accounts using Entra ID, Copilot inside Microsoft 365 Personal and Family apps, or users in Brazil, China (not including Hong Kong), Israel, Nigeria, South Korea and Vietnam.
• What you upload: the simplest protection is not to upload photos of yourself or other people. Microsoft’s FAQ itself asks users not to share images of others without their consent.

Microsoft isn’t alone here. Most big AI companies use some human review, as our comparison of what each chatbot does with your data found, but few spell out that reviewers may see your uploaded photos with faces intact.

Why it matters

Microsoft is pushing Copilot into everything, from Windows to the new “OS for work” super app it relaunched last week, and most people assume a chatbot conversation stays between them and the machine. This report is a reminder that behind the AI there are people seeing what users type and upload, and that the privacy promises in the FAQ don’t cover everything they see.

Sources: 404 Media, Microsoft Copilot privacy FAQ.