Microsoft’s human reviewers are seeing Copilot prompts and uploaded images—including abusive sexual edits

Human contractors working to improve Microsoft Copilot are reviewing real user prompts, uploaded images and generated results, according to internal documents obtained by 404 Media. The work includes judging whether the image system successfully fulfilled requests involving sexualised bodies, upskirt imagery and women placed in sexual positions. One contractor said faces were left uncensored and many requests appeared dubiously consensual.

The review is not a hidden capability in the narrow legal sense. Microsoft’s Copilot privacy FAQ says that some conversations undergo automated and human review for product improvement and digital safety, and that users cannot opt out when a suspected code-of-conduct violation is investigated. But that general disclosure does not convey the likely intimacy of uploaded photographs, the use of outside labour or the possibility that reviewers must inspect whether an abusive edit was executed accurately.

There are two separate safety problems. The first is product control: filters failed to stop enough sexual and potentially non-consensual requests that workers encountered them repeatedly. The second is the review pipeline itself. A system intended to detect and improve those failures still exposes identifiable people in user-supplied images to additional humans, while requiring contractors to process disturbing material.

Microsoft distinguishes consumer Copilot from enterprise Copilot Chat, where prompts and responses receive contractual data protections and are not used to train foundation models. Users can easily miss that distinction when the products share a name and interface conventions. Even enterprise protection does not answer every question about safety review; the precise product, account type and policy setting matter.

This follows 404 Media’s earlier reporting on contractors reading selected ChatGPT conversations. The pattern is broader than one vendor: ‘AI’ quality and moderation still depend on human access to sensitive failures. Meaningful disclosure should therefore say what can be reviewed, by whom, under what redaction and retention rules, and with what support for workers—not merely that review may occur.