Microsoft Copilot AI Chatbot Prompts and User-Uploaded Images Reviewed by Human Contractors

Human Review of Copilot Interactions

Human contractors hired by Microsoft to improve the Copilot AI chatbot are actively reviewing user prompts and uploaded images. This includes content that is sexually explicit, such as requests for photo editing that involve lewd or sexual positions, and images uploaded by users themselves. The contractors are tasked with evaluating whether the AI successfully fulfills these prompts.

Privacy Concerns for Users

Users who upload personal images to Copilot likely expect these interactions to remain private between themselves and the AI tool. However, a workforce of hundreds of human reviewers has access to these prompts and uploaded images. This practice extends beyond text prompts, as previously revealed for OpenAI's ChatGPT, to include visual content submitted by users.

Disturbing Content Encountered by Reviewers

Contractors have reported being inundated with requests for disturbing content, including foot fetish images of children's cartoon characters, alterations to women's clothing, and images depicting pro-anorexia content. Internal documents seen by 404 Media, including instruction guides and real user prompts, confirm the nature of the content reviewed. One contractor noted that faces in uploaded images are often uncensored, and many prompts are sexual and of dubious consent.

Implications for AI Development and User Trust

The revelation highlights the extensive human involvement in AI model training and content moderation, even for seemingly private user interactions. This practice raises questions about user data privacy and the ethical considerations of exposing human workers to potentially harmful or explicit content as part of AI improvement processes. It underscores the ongoing challenges in managing user-generated content within AI systems.