Your private images and prompts are seen by workers for Copilot

When you ask Microsoft's AI chatbot Copilot to edit an image, many people probably think it stays between you and the machine. It doesn't. According to an investigation by the outlet 404 Media, based on internal documents, real people review a large share of the images and editing requests users send to Copilot.

The contractors are tasked with assessing which of two AI-generated versions of an image is better, including whether the edit follows the user's instruction and looks good. But in practice, according to the report, they are flooded with problematic content: sexualized images of real women, so-called "upskirt" images taken without consent, and in some cases content involving children. Several contractors have said they received no warning about what they might be exposed to.

A central problem is that the faces of the people in the images are not blurred. This means the reviewers can see who the people are, even though the image has been manipulated in an abusive way without the person's knowledge. Several of the contractors have themselves questioned whether the people depicted have even consented to the requested changes.

Microsoft states in its privacy terms that conversations can be stored for up to 18 months, and that uploaded images can be used in the company's work to improve the service. But it is not made clear to the user, when uploading an image, that it may be shown to a stranger for assessment. Nor is there any option to say no and opt out of having your images used that way.

For regular users, the point is simple: An image you send to an AI service like Copilot to have it edited is not necessarily private. It may end up being seen and judged by another person, without you knowing it or having given your explicit consent.