Meta reluctantly admits it trains AI models with photos taken by Ray-Ban Meta smart glasses

Last week on Meta Connect 2024, Meta announced a new feature that upgrades the Ray-Ban Meta smart glasses' capability that lets users capture pictures using the smart glasses' camera and have Meta AI answer questions about it. To make the experience more 'natural', the company is launching a real-time AI-powered video capability that will be able to analyze a stream of images rather than a single capture. As Meta puts it, the new the Ray-Ban Meta smart glasses can answer questions and describe what the users are seeing, which sounds vastly different from having Meta AI analyze an image that was previously captured using the glasses' onboard camera.

The distinction between answering a question about something users are seeing or about a picture they captured obscures (perhaps intentionally?) a very important piece of information: smart glasses can't really see in the same sense as we do. Rather, this effect is achieved by having the glasses passively capture endless streams of images of whatever the user is asking about. Since these passively captured images lack the intentionality and awareness associated with taking a picture and knowingly uploading it for Meta AI (or any chatbot, for that matter) to analyze it, users may not always realize that when using features like the new AI video capability, they are continuously uploading images to Meta's servers.

Even more concerning is the fact that by using these features, users may be unknowingly sharing everything from the content of their closets to the faces of their relatives and other data they may not want to share with Meta after all. According to a help article about Meta AI on the Meta Ray-Ban smart glasses, "all images processed with AI, voice transcripts, and related data are stored and used to improve Meta products and will be used to train, improve, and refine Meta's AI models with help from trained reviewers."

Although users can opt out from having their voice controls interactions used for training (instructions here) it turns out that there isn't even a way to opt out from uploading passively captured image streams for Meta to use for model training. The previously mentioned help article about Meta AI on the Meta Ray-Ban smart glasses provides instructions on how to view, manage, download, and delete users' stored Meta AI data, but there is no indication that deleting the stored data suffices to opt-out from data collection. This is on-character for Meta, given previous controversies, and the most recent reminder of how little consideration tech firms can have about protecting users' privacy and rights.