Photos uploaded to Microsoft Copilot are shown to human reviewers with faces uncensored

According to a report published on September 28 by US technology news site 404 Media, based on internal company documents, outside contractors working to improve Microsoft’s AI assistant Copilot can see the photos users upload for editing and the prompts they write. The documents include instruction guides, real Copilot prompts and images, and posts from an internal message board for contractors.

According to the report, in each task a reviewer is shown the user’s original prompt, the uploaded photo and two edits generated by Copilot, and chooses which is better. One person who works on the prompts told 404 Media: “Faces are always uncensored, and many of the prompts are sexual in nature and dubiously consensual.” The reviewers are not tasked with flagging harmful content, only with rating the technical quality of the edits.

Contractors described frequently encountering sexually explicit and potentially illegal material, including covertly taken upskirt photos and requests for sexualised edits of real women. At least one company hiring the contractors is Prolific. According to Windows Central, Microsoft’s guidance to contractors advises them to trust their intuition in unclear cases.

Microsoft says the updated consumer Copilot app, available since August 18, does not use prompts, responses or uploaded file contents to train foundation models. The reporting does not show that every Copilot prompt or uploaded image is reviewed by a person; it shows that some contractors saw user-submitted material during quality evaluations. Earlier this month, 404 Media reported that OpenAI contractors also read real ChatGPT conversations.