Full Breakdown
Microsoft Copilot Contractors Review User-Submitted Images, Exposing Disturbing Content
By Drooid · · How we work
Core Event: Human Review of Copilot Image-Editing Prompts
Internal documents obtained by 404 Media show Microsoft uses several hundred contractors to evaluate image-editing outputs from the Copilot AI tool. Reviewers see the user’s prompt, the original photo, and two AI-generated edits, then select the better version based on adherence to instructions, preservation of unchanged areas, absence of visual artifacts, and overall quality.
Background & Context
Since launching consumer-facing AI image-editing features, Microsoft has allowed users to upload personal photos and request modifications such as “shorten the skirt” or “enlarge the breasts.” Its terms of use state uploaded content may be used to improve products and enforce its code of conduct, and de-identified data can be used for model fine-tuning. Earlier reporting by 404 Media documented similar contractor practices at OpenAI, indicating a broader industry reliance on human feedback loops.
Review Process Details
Each task presents two side-by-side edits; the reviewer decides which better follows the user’s request while preserving background elements and avoiding distortions. Reviewers are not required to flag illegal or non-consensual content; their remit is limited to quality assessment.
Content Exposed to Contractors
Contractors repeatedly encountered sexualized and disturbing material, including:
- Requests to enlarge breasts, shorten skirts, or place women in suggestive positions.
- Prompts involving children’s cartoon characters in fetishized contexts.
- “Pro-anorexia” imagery, such as a pop star rendered emaciated.
- Up-skirt photographs and other non-consensual depictions.
- Images described as “some sort of animal sacrifice.”
One contractor said they “recoiled” after seeing pro-anorexia content, and another flagged eight up-skirt photos as unsafe. These accounts indicate the quality-review pipeline does not filter explicit or potentially illegal material before human eyes.
Official Statements & Responses
Microsoft’s privacy notice states uploaded images are deleted within 30 days after a conversation ends and that automated detection systems screen for child sexual exploitation, disturbing content, and non-consensual deepfakes. No comment was provided regarding the scale of contractor access or pre-screening safeguards.
On-the-Ground Reports from Contractors
Internal message-board excerpts shared by 404 Media show contractors voicing frustration and ethical concerns. One participant asked who authorizes the focus on pornographic outputs, while another lamented judging “the appropriate bust size” rather than substantive content. Reviewers requested an “unsafe content” flag for tasks containing explicit imagery, suggesting the workflow lacks clear escalation mechanisms.
Conflicting Reports & Gaps
The reporting does not establish how many uploads reach human reviewers, nor whether every request undergoes review. Microsoft’s documentation outlines automated filters for prohibited material, but the effectiveness of those filters before reaching contractors remains unclear. The role of the contractor platform Prolific, identified in the documents, has not been independently verified, and the company declined to comment.
Implications for Users
For consumers of Microsoft’s Copilot, the gap between the expectation of private, algorithm-only processing and the disclosed possibility of human review raises privacy concerns. Exposure of sensitive or non-consensual images to a largely undisclosed workforce may affect user trust and underscores the need for clearer disclosures and stronger pre-screening safeguards.