A report published September 28, 2026, said human contractors may review some Microsoft Copilot image-editing results—and may see the prompt, uploaded image and generated output while doing so. The reported evaluations concern whether edits follow instructions and meet visual-quality standards, not whether content should be blocked.
What contractors may see in image-editing evaluations
The reported task can include a user’s prompt, the original image and the AI-generated result. Contractors reportedly said faces in source images were not blurred and identifying details in prompts were rarely obscured, even though the data was supposed to be anonymized.
Contractors also described sexualized requests and other disturbing imagery. The reported review concerns some image-editing evaluations; it does not establish that a person reviews every Copilot conversation.
Copilot Studio has a separate human-input feature
Microsoft documents a different human-review workflow in Copilot Studio. Its configurable Request for information action pauses an agent flow to collect input from an assigned reviewer, then uses the first response. This is a step an agent builder can add to a flow, separate from the reported contractor evaluation of Copilot image edits.
Microsoft’s customer-data statement
Microsoft said it uses customer data as described in its terms, including to improve its products and enforce its code of conduct.
