The Contractor Behind Your “Private” AI Prompt

Human contractors hired to improve Microsoft’s Copilot reportedly reviewed image-editing prompts and uploaded photographs, including sexual, disturbing, and potentially nonconsensual material. Their assignment was to judge how well Copilot fulfilled a request, not to decide whether the request was ethical. Workers described seeing uncensored faces, upskirt photographs, sexual alterations of real women, pro-anorexia material, and prompts involving young girls or children’s characters.

The review placed people inside an experience designed to feel like a private exchange with software. A user types into a clean box, uploads a photograph, and receives an answer from a named product. Nothing in that choreography suggests a review floor where another person may assess whether the requested bodily alteration was rendered accurately. The machine has a back office.

That does not make every human review improper. Image systems need safety testing, abuse detection, quality control, and investigation when their safeguards fail. But those purposes require different access. A moderator checking for prohibited material does not need the same mandate as an evaluator rating whether a sexualized edit followed instructions. Necessary review is not a blank permission slip.

The social code of the interface matters because it encourages disclosure before explaining the audience. People may upload partners, relatives, patients, coworkers, or strangers who never consented to appear in a training or evaluation queue. Faces reportedly remained visible. The supplied reporting does not establish whether reviewers also saw account identifiers, where all contractors were located, how long reviewed material was retained, or whether users could delete it after evaluation. Those are not technical footnotes. They define who is in the room.

Workers carry another part of the cost. Contractors said disturbing material could appear without warning, and some described recoiling from what they saw. A responsible review operation would provide advance content labels, the ability to decline categories without penalty, trained psychological support, limited exposure periods, strong access logs, and pay reflecting the difficulty of the work. Surveillance of contractors cannot substitute for care of contractors.

Before submitting intimate material, customers should check whether prompts and uploads are reviewed by people, used for product improvement, linked to account data, retained after deletion, or excluded through a privacy setting. Microsoft and other AI companies should state those answers beside the upload control, in ordinary language, before the file leaves the device. A conversation cannot be meaningfully private when its human audience is disclosed only after someone goes looking for it.