Contractors Report Disturbing Content in Microsoft Copilot Prompts

At least hundreds of contractors are currently reviewing prompts for Microsoft's Copilot AI, with thousands more engaged in similar activities for OpenAI's ChatGPT. In March 2024, the U.S. Senate approved Copilot for use, but recent revelations from 404 Media highlight troubling practices in prompt generation.

Contractors reported that many prompts contain sexual content, with one anonymous contractor stating, "Faces are always uncensored, and many of the prompts are sexual in nature and dubiously consensual." This follows earlier reports documenting Copilot's response behavior to user inputs.

The internal guidelines for contractors reportedly encourage subjective evaluations of image edits, with instructions like, "Trust your intuition — when you glance at the two edited images side by side, which one immediately feels like the better edit?" Another contractor expressed horror at pro-anorexia content, saying, "I recoiled."

Prolific, the company hiring these contractors, emphasizes the need for caution around disturbing or explicit content in generative AI. Microsoft maintains that it uses customer data to improve its products and enforce its code of conduct, as stated by a company spokesperson. These findings raise significant ethical questions about the content being generated and the oversight of AI systems like Copilot.