Human contractors hired to evaluate Microsoft’s Copilot AI chatbot are routinely reviewing sexually explicit image editing requests and non-consensual photos uploaded by users, according to internal documents obtained by 404 Media. The contractors are tasked with reviewing the output quality of image generations, assessing whether the model successfully fulfilled user requests ranging from altering clothing lengths to modifying body features.

While users typically expect private photo uploads to remain confidential between themselves and the model, internal message boards reveal that a workforce of hundreds of human reviewers examines both text prompts and uploaded imagery. The leak indicates that contractors are not evaluating the material to flag or moderate inappropriate content, but rather to rate model accuracy and side-by-side output quality based on instructions to rely on human intuition.

The findings highlight privacy risks surrounding user-submitted content in generative AI applications and expose the operational reliance on human feedback mechanisms for evaluating output quality. The situation underscores ongoing labor and privacy tensions within AI supply chains.

Why it matters

  • Exposes privacy risks for enterprise and consumer users uploading confidential images to generative AI assistants.

  • Highlights operational challenges in RLHF data pipelines when human evaluators assess non-consensual or explicit content.

  • Signals potential regulatory and public scrutiny over data retention and human review practices at major AI vendors.

Source: 404media.co