Human Copilot Reviewers Are Seeing Users’ Disturbing Image-Editing Requests, New Report Reveals

Microsoft Copilot users may not be the only ones seeing the prompts and images they submit to the AI assistant. Internal contractor documents obtained by 404 Media show that at least hundreds of human reviewers are sometimes being shown Copilot prompts, uploaded images, and the AI-generated edits created from them.

The reviewers are not simply checking whether Copilot produced something unsafe. According to the documents, they are being asked to judge the quality of Copilot’s image edits, meaning a user’s original prompt and uploaded photo can appear alongside two versions generated by the AI.

Copilot reviewers are seeing users’ uploaded images

The review process is fairly straightforward. A contractor receives the user’s original editing request, the uploaded image, and two Copilot-generated versions. They then choose which edit is better.

The instructions reportedly ask reviewers to check whether Copilot followed the user’s request, preserved parts of the image that were not supposed to change, avoided visual artifacts, and produced a high-quality result. Microsoft also tells reviewers to rely on their own judgment when comparing the images. The instructions noted the following:

When in doubt go with your first impression. Human intuition is good at catching subtle quality differences that are hard to articulate.

The problem is the material appearing in those queues. Contractors told 404 Media that they regularly encountered sexually explicit requests and disturbing images, including requests involving sexualized edits of real women, upskirt photographs and other graphic material.

Contractor discussions reviewed by the outlet also included reports of pro-anorexia prompts, sexualized depictions involving young people and other potentially unsafe content. One contractor reportedly asked supervisors for an unsafe-content flag after receiving a set containing eight upskirt photographs.

Microsoft says customer data can be used to improve Copilot

At least one company involved in hiring contractors for the work is Prolific. The documents also reportedly contain Prolific guidance warning that generative AI research can expose participants to disturbing or explicit material.

Microsoft did not dispute that customer data is used in this way. A company spokesperson told 404 Media that the company “uses customer data as described in our terms of use, including to improve our products and enforce our code of conduct.”

That does not mean every Copilot prompt or uploaded image is being manually reviewed by a person. The report specifically describes a contractor review process involving selected tasks and says a workforce of at least hundreds of reviewers can encounter this material.

In other AI news, Satya Nadella has said the industry is too focused on itself while also admitting that he doesn’t blindly trust AI and instead interrogates its answers. Meanwhile, Claude Sonnet 5.5 is now available in Microsoft Foundry.