Contractors See Copilot Users’ Photos While Rating AI Edits, 404 Media Finds

The reviewers judge image quality, not primarily safety. Workers say some tasks have exposed them to sexual or disturbing material.

By 3 min read
Contractors See Copilot Users’ Photos While Rating AI Edits, 404 Media Finds
Contractors See Copilot Users’ Photos While Rating AI Edits, 404 Media Finds

Listen to this story

The audio brief

About 1:32
0:001:32
Read transcript
A photo uploaded to Microsoft Copilot for editing can be seen by a human contractor—not just processed by the AI. An investigation by 404 Media found that reviewers may see the original image and the user’s prompt, then compare two AI-generated edits and choose the stronger one. The job is to judge quality: did the edit follow the request, preserve the parts of the picture that were meant to stay, avoid distortions, and look good overall? So if someone asks to add a cup to a table, the reviewer also checks that the rest of the scene remains intact. These are visual assessments, not primarily safety checks. The investigation reports at least hundreds of reviewers in the contractor pool, but that does not tell us how likely any particular upload is to be reviewed. Workers told 404 Media they had encountered sexual and otherwise disturbing material, including requests to sexualize women and content promoting eating disorders. One contractor asked for a warning on tasks involving upskirt photos. 404 Media identified Prolific as one company recruiting for the work; it did not respond to a request for comment. Microsoft said it uses customer data under its terms, including for product improvement and conduct enforcement. But it did not clarify whether users can keep uploaded images out of these ratings. Its older Copilot privacy FAQ discusses choices about model training and personalization, while users of the updated version are directed to different guidance. The unresolved issue is what notice and control users have when people assess their image edits.

Story brief

3 key points

404 Media’s review of internal instructions and worker accounts found that Copilot image-editing quality tasks can expose uploaded originals and prompts to human raters, who compare two generated edits. The reported reviewer pool numbers at least hundreds, but the investigation does not establish how often any particular upload is seen. Workers reported encountering sexual and otherwise disturbing requests....

  1. 01

    Reviewers assess prompt adherence, whether unchanged parts of an image are preserved, visible distortions, and overall quality.

  2. 02

    The reported pool is at least hundreds of reviewers; the investigation does not establish the chance an individual upload is reviewed.

  3. 03

    404 Media identified Prolific as one contractor recruiter; Prolific did not respond to its request for comment.

A photo uploaded to Microsoft Copilot for editing can end up in front of a human contractor. In an investigation based on internal documents and worker accounts, 404 Media found that reviewers see the user’s prompt and original image while choosing between two AI-generated edits. The arrangement raises a question for users: who sees the pictures they hand to an AI tool?

The image on the reviewer’s screen

The instructions obtained by 404 Media describe a task built around a real request to edit an uploaded picture. A contractor sees the prompt, the original picture and two Copilot-generated versions, then picks the stronger edit. The upload is part of what the reviewer assesses, not merely an input supplied to Copilot.

Reviewers check whether an edit follows the request, preserves parts of the image that should stay unchanged, avoids visible distortions and looks good overall. If the request is to add a cup to a table, for example, the instructions say the rest of the scene should remain intact. The comparison calls for visual judgment, not just a check that Copilot produced an image.

404 Media puts the workforce at at least hundreds of reviewers who may sometimes see Copilot prompts and uploads. That figure describes the reported pool of people, not how often an individual user’s image is reviewed.

Human intuition is good at catching subtle quality differences that are hard to articulate.

Instructions given to Copilot contractors, quoted by 404 Media

A quality task with disturbing inputs

Contractors described encountering sexual imagery and other disturbing material. On an internal message board reviewed by 404 Media, workers discussed requests to sexualize images of women and said they had seen content promoting eating disorders. One contractor wrote about receiving a set of upskirt photos and asked for an unsafe-content flag, so reviewers would know what a task contained before accepting it.

Their assignment is to rate Copilot’s edits, rather than primarily to flag inappropriate requests. That means a worker judging whether an edit followed instructions may have to inspect content they find objectionable. 404 Media identifies Prolific as at least one company hiring contractors for the work; Prolific did not respond to its request for comment.

The question about user control

Microsoft told 404 Media that it uses customer data as described in its terms, including to improve products and enforce its code of conduct. Its response did not say whether users can exclude uploaded images from the rating tasks described in the investigation.

Microsoft’s privacy FAQ says uploaded files and conversations about them are subject to users’ choices about model training and personalization. But that FAQ applies only to an older Copilot app; Microsoft directs users of the updated version to different guidance. The live question is what notice and control people have when human reviewers assess their image edits—not simply whether they can change a model-training setting.

Sources

  1. support.microsoft.comsupport.microsoft.com
  2. 404media.coHumans Are Reading Copilot Prompts — And They're Horrified

Loading discussion...

YOUR READING SPACE

Notifications