Microsoft Contractors Can Access Unblurred Photos and Raw Prompts from Copilot Users
An investigation by 404 Media revealed that hundreds of third-party contractors hired by Microsoft to evaluate Copilot quality can view unblurred user-uploaded photos, raw prompts, and AI-edited image results. The disclosure highlights significant privacy and data governance risks in human-in-the-loop evaluation processes, exposing personal photos and sensitive user prompts to third-party workers. While Microsoft's privacy FAQ states that faces are blurred before training AI models, this safeguard does not apply to human quality rating workflows managed via platforms like Prolific. Contractors reported reviewing unblurred faces as well as sexually suggestive or non-consensual content embedded in user prompts.
## BACKGROUND
AI developers frequently rely on human evaluation and Reinforcement Learning from Human Feedback (RLHF) to rate model output quality and safety. Platforms like Prolific connect AI technology companies with crowdsourced annotators who evaluate text, image, and video outputs produced by generative models.