The photos and messages you send to Copilot can be seen by human reviewers, and some are horrified

Photos that users upload to Copilot and the instructions they write to modify them can end up in the hands of human reviewers hired to improve Microsoft AI. An investigation by 404 Media reveals that these workers receive material that users may consider private and that some are horrified by the sexual content they must examine.

The media has consulted internal documents, evaluation guides and conversations between workers that describe how image edits are reviewed. Their testimonies expose both the privacy problems of this process and the discomfort it causes among those who must evaluate the results.

Photographs with unblurred faces

Evaluators receive the original photograph sent by the user, their request and two versions modified by Copilot. They must choose which one best complies with the instructionspreserves the elements that should not change and offers greater visual quality, without deformations or strange textures.

Access to this material allows read the user’s request and see the person in the photo. According to testimonies collected by 404 Media, faces appear without blurringso the people portrayed may be recognizable.

This exposure also affects those who appear in photographs that someone else has uploaded to the service, including requests for sexual retouching that, according to reviewers, may not have the consent of the people portrayed.

Evaluate the quality of disturbing content

Workers’ complaints describe requests for enlarge the breasts of real women, shorten their skirts or place them in sexualized positions. They also meet fetish requests about cartoon characters, images that promote anorexia and requests to create suggestive content with minors.

The evaluations they carry out on the quality of Copilot results, even if the content is disturbing, They are not intended to identify inappropriate material in order to teach the system to block it.. This leads them to examine images that the workers themselves describe as potentially illegal while they are asked to judge how well the AI ​​has done the touch-ups. As ExtremeTech warns, the possession of certain content could constitute a crime in some countries, depending on the material and the applicable legislation.

‘It’s hard to take things seriously when I have to focus on which model has produced the appropriate breast size or which middle-aged woman has been placed in the appropriate sexually suggestive position.’wrote one of the workers in an internal forum to which 404 Media had access.

Some reviewers also report that accepted tasks without knowing that they contained explicit or potentially illegal images. One reported receiving eight upskirt photographs and asked for an unsafe content warning to be added, while another described images of what looked like an animal sacrifice.

Microsoft’s response to the media’s questions is limited to a general explanation about the processing of information. ‘Microsoft uses customer data as described in our terms of use‘, among other things to improve our products and enforce our code of conduct,’ a spokesperson said.