Microsoft uses third-party human contractors to evaluate user prompts and uploaded pictures processed through its Copilot artificial intelligence system. Internal documents obtained by investigative outlet 404 Media show that workers regularly review side-by-side image edits to judge system accuracy. These tasks include rating sexually suggestive requests, non-consensual image modifications, and private personal photographs submitted by users with unblurred faces. The workflow exposes external staff to sensitive customer files while testing how closely the software follows instructions.
Key Takeaways
- Microsoft employs outside human contractors to manually review text prompts and uploaded images submitted to Copilot.
- Contractors review uncensored user photographs without facial blurring to assess side-by-side editing accuracy.
- Leaked documents confirm reviewers encounter sexually explicit requests, upskirt prompts, and altered body features.
- The evaluation process aims to verify if the output matches prompt instructions rather than removing offensive content.
Contractor Guidelines and Evaluation Tasks
Microsoft provides Copilot, an artificial intelligence assistant built on large language models and diffusion image generators, to consumer and business accounts. Users routinely upload photos to edit clothing, backgrounds, or facial traits. Internal documentation directs workers to compare two generated outputs side by side, instructing them to trust their intuition to decide which visual result fulfills the user text request better.
The documents show reviewers assess whether Copilot followed specific image editing prompts. These requests involve shortening skirts on real women, modifying physical body proportions, generating upskirt angles, or placing subjects into sexual scenarios. The contractors do not act as safety moderators tasked with deleting harmful files; instead, they grade the accuracy of the algorithm in executing the user command.
Privacy Exposure and Data Handling
The reports highlight privacy risks for people who upload personal photos to cloud-based artificial intelligence services. Reviewers see uploaded faces completely unblurred. This means contractors can view identifiable family members, coworkers, and acquaintances that users feed into the software.
Microsoft maintains safety filters designed to block harmful and explicit material, though administrative settings in enterprise setups can alter filter controls. The company stated that customer data usage aligns with published service agreements. Even so, the discovery that real people view raw consumer inputs contradicts common user assumptions that AI interactions stay confined to automated software servers.
Frequently Asked Questions
Q1. Why does Microsoft hire human contractors to review Copilot images?
A1. Microsoft hires contractors to evaluate output quality. Reviewers look at two generated options side by side and rate which image better executes the user prompt.
Q2. Do human reviewers see private photos uploaded to Copilot?
A2. Yes. Leaked company instructions show uploaded photos reach human reviewers without facial blurring or identity masking.
Q3. Are contractors acting as content moderators to delete bad content?
A3. No. The contractors grade image accuracy and prompt compliance, not content policy violations.
Q4. What types of inappropriate requests appear in the review queues?
A4. Reviewers report seeing requests to shorten skirts on photos of real women, generate upskirt views of characters, enlarge physical features, and build sexually suggestive pictures.
Q5. How can users prevent humans from seeing their uploaded photos?
A5. Users can avoid uploading personal, sensitive, or identifiable photos to public artificial intelligence tools, as standard terms of service often allow vendor review for quality assurance.
