Microsoft and OpenAI employ hundreds of contractors to review user interactions with the Copilot AI chatbot. These workers analyze both text queries and images uploaded by users to improve the service's performance and quality.
Internal documents obtained by 404 Media reveal that contractors frequently encounter explicit, sensitive, and potentially non-consensual imagery. Workers are tasked with evaluating the quality of AI-generated edits without specific instructions to filter offensive content.
The review process requires contractors to compare original user images with two AI-generated variants based on four criteria. Instructions explicitly state that researchers cannot fully control the content displayed to participants, advising them to exercise caution.
Contractors are hired through the platform Prolific, which provides feedback for AI safety and benchmarking. Some workers have reported discomfort and requested that management flag certain assignments due to the nature of the material.
Microsoft stated that it utilizes customer data in accordance with its terms of use, specifically for product improvement and enforcing its code of ethics.
