Microsoft, like many other tech giants, hires hundreds of third-party contractors to manually analyze and evaluate text prompts and images sent by users to the Copilot chatbot. As a journalistic investigation revealed, this practice comes with serious ethical issues: moderators and reviewers regularly encounter confidential shots and explicitly intimate materials. Such checks are carried out to improve the service and fine-tune generative artificial intelligence algorithms, but users often have no idea that their personal photos are being reviewed by real people.
Moderator Operations and Document Leaks
Specialized media outlets obtained a package of internal documents from Prolific, the company Microsoft uses to hire specialists for generation quality assessment. The contractor instructions clearly state that they have to evaluate the results of editing user images. Reviewers are shown the original prompt, the uploaded source image, and two options created by the neural network. At the same time, faces in the photographs remain uncensored, and the materials themselves often shock employees since they are not subject to strict preliminary filtering for ethics or legality.
Contractor Instructions and Microsoft's Response
According to internal regulations, contractors must rely on their intuition when choosing the most successful generation option without focusing on the moral side of the issue. However, some operators complained about difficult working conditions and the need to process potentially unsafe or illegal content, asking management to flag such tasks as problematic. In response to media inquiries, Microsoft representatives stated that the company processes user data strictly in accordance with current terms of use, including product improvement goals and compliance with internal ethics codes.
Contradictory Data
While Microsoft's official stance emphasizes the legality and safety of using client data to improve artificial intelligence, independent observers and contractors themselves point to deep vulnerabilities in privacy protection. The corporation claims that all processes are automated and controlled, but the leak of internal instructions proves otherwise: real people get direct access to uncensored intimate user materials without their explicit and informed consent for this type of manual moderation.
