Microsoft Copilot Human Reviewers Exposed to Explicit User Images and Prompts

Human contractors hired to improve Microsoft’s Copilot AI chatbot face constant exposure to lewd or sexually explicit photo editing requests and uploaded images. These uploads include upskirt photos or images placing women into sexual positions. Workers must review whether the generated images successfully fulfill prompts such as making AI enlarged breasts larger.

Users uploading pictures of themselves or others likely expect privacy between themselves and the AI tool. A workforce of at least hundreds of human reviewers routinely examines these prompts and uploaded images. Contractors face demands to create foot fetish images of children’s cartoon characters, shorten real women's skirts, and place people into sexual positions.

Internal documents reveal explicit tasks

Internal contractor documents seen by 404 Media detail instruction guides, real Copilot user prompts, and internal message board conversations. Contractors discuss prompts asking Copilot to make skirts shorter, enlarge breasts, or show more leg while wearing stilettos. She-Hulk images appear frequently alongside pro-anorexia content and prompts requesting images of Ariana Grande with anorexia.

Contractors are not hired to flag or vet offensive or inappropriate content. They receive payment to review the quality of Copilot output including sexual imagery. Microsoft focuses explicitly on human intuition regarding which image looks better. Instructions tell contractors to trust their gut reaction when comparing two edited images side by side.

Contractors report disturbing material

Tech companies and social networks have long used contractors to moderate user generated content. AI companies also employ overseas workers to train models and reduce toxicity. In the case of Copilot, contractors review material users assume is private not for safety filtering, but to make chatbot responses more informative, friendly, and clearer.

Contractors review an original prompt, an uploaded picture, and two Copilot generated edits. Workers select the better edit based on instruction adherence, preservation of unchanged elements, absence of visible artifacts, and overall quality. Some contractors complained in private forums about unknowingly encountering explicit, unsafe, or potentially illegal images including eight upskirt photos and animal sacrifice imagery.

Companies respond to review process

Prolific is at least one company hiring contractors for this work. Prolific mentions human feedback services for preference tuning, safety evaluations, and benchmarks. A contractor quoted Prolific guidelines noting that researchers cannot fully control what generative AI shows participants. Prolific did not respond to a request for comment.

A Microsoft spokesperson stated in an email that Microsoft uses customer data as described in its terms of use, including to improve products and enforce a code of conduct. Microsoft AI tools have a history of abuse for generating nonconsensual images, such as Taylor Swift pornographic images generated on Microsoft Designer. A top Senate administrator approved Copilot for Senate use in March 2024 for routine tasks like drafting documents and conducting research.