Microsoft Copilot Contractors Review Disturbing User Image Edits
Microsoft Copilot contractors are reviewing user-uploaded images and editing requests, including disturbing and abusive content, raising privacy concerns.
Microsoft's AI image generation service, Copilot, is reportedly involving human contractors in reviewing user-uploaded images and their associated editing requests. This practice has come under scrutiny following a report from 404 Media, which detailed how third-party contractors are tasked with assessing the technical merit of AI-generated image edits based on user prompts.
The scope of these reviews extends to highly disturbing and abusive content. Contractors are presented with original user photos, the editing requests, and the AI's output, with no censorship of the faces in the original images. The requests and subsequent edits have included sexually explicit material, nonconsensual imagery, and even content involving minors, according to interviews with affected contractors. Examples cited include requests to alter clothing, enlarge body parts, and place individuals in sexual positions, as well as disturbing requests involving cartoon characters and public figures.
Compounding the privacy concerns, these contractors are reportedly paid to evaluate the technical quality of the AI's output, such as how well the edits followed instructions or maintained image integrity, rather than to flag or report inappropriate or illegal content. This means that sensitive and potentially harmful material is being viewed and processed by humans without explicit user consent for such review.
Microsoft's data usage policies add another layer of complexity. While the company states that customer data, including uploaded images, may be used to improve its products, it has not clarified whether users can opt out of having their image uploads subjected to these human review tasks. The company's privacy FAQs mention that while users can delete conversations and opt out of some AI training, an opt-out from certain human review processes is not available.
This situation is not entirely new for Microsoft's AI tools. Earlier in 2024, concerns were raised about the ability to bypass safety guardrails in Copilot Designer's text-to-image tool, DALL-E 3, leading to the generation of inappropriate images. A loophole was also patched that allowed for the creation of nonconsensual intimate deepfakes of public figures.
Microsoft's updated consumer Copilot app, released in August 2026, claims not to use prompts, responses, or uploaded file contents for training foundation models. However, the core issue of human review of user-uploaded images for quality assessment remains a significant privacy question, particularly for users who may not be aware of this practice.
This practice mirrors similar initiatives by other AI companies, such as OpenAI's Project Lily, which also employs contractors to review AI-generated content. The ongoing reliance on human reviewers for AI model improvement highlights a persistent challenge in balancing AI development with user privacy and ethical considerations.
Users concerned about their data privacy with AI services are advised to consult resources on AI training data and user content, such as guides provided by cybersecurity firms, to understand their options for opting out of data collection and review processes where possible.