Meta is betting heavily on smart glasses to advance its artificial intelligence and augmented reality business, but newly exposed internal processes have once again raised strong questions about its privacy boundaries. Several data annotators employed by the Kenyan outsourcing company Sama revealed that they have to watch and annotate a large number of raw video clips shot by the "live AI (real-time AI)" function of Meta smart glasses every day, which not only includes daily interaction scenes, but also involves extremely private and even shocking scenes.
These annotations are a key part of training the Meta computer vision system. Every frame of video that is manually annotated will feed back the algorithm performance of its augmented reality assistant. After the "live AI" function is turned on, the smart glasses launched by Meta and Ray-Ban will continue to collect short videos and audio through cameras and microphones in order to analyze the scenes seen by the user in real time and answer their questions. The relevant data will then be uploaded to the Meta system and integrated into a huge training data set to optimize future versions of the AI assistant.
However, Sama employees say the data uploaded is actually far more private than users imagine. These data annotators said that they have repeatedly seen scenes of people wearing glasses moving, changing clothes, and even having sex in the bathroom, and the shooting perspective is from the glasses themselves. Even if explicit content is not involved, the relevant clips often reveal a large amount of sensitive information, such as fully displayed bank cards, clearly identifiable home interior layouts, and even highly private conversations. Audio clips documenting discussions about protests, alleged criminal behavior, or the privacy of one’s personal life were incorporated into training material for Meta’s algorithm.
Judging from the public terms, Meta has reserved space for this approach at the clerical level. The company writes in its AI terms of service and privacy policy that interactions with "live AI" assistants may be retained and viewed by automated systems or human review, and explicitly reminds users not to share sensitive information. However, according to reports from the Swedish media Svenska Dagbladet and Göteborgs-Posten, users who actually wear glasses generally do not seem to understand that every move in front of them may be transmitted overseas and reviewed frame by frame by unfamiliar contractors. Annotators revealed that when they complained to management that the nature of the content was too personal or that the annotation process was uncomfortable, the feedback was often dismissed out of hand.
Meta initially did not respond to repeated inquiries from Swedish journalists for several weeks. When a spokesperson later responded to the media, it was still based solely on the company's AI terms of service and privacy policy, emphasizing that media content uploaded by users when using "live AI" will be processed in accordance with these documents, and declined to provide further comments to Straight Arrow News.
Social concerns surrounding Meta wearables have continued to grow in recent months. Earlier this year, The New York Times reported that Meta planned to add facial recognition capabilities to its smart glasses, citing an internal memo. Civil liberties groups warn that when facial recognition is combined with continuous video collection, it could create a mobile surveillance network in the real world with limited oversight and transparency.
Outside of Meta, third-party developers are also beginning to try to build "defensive" tools to deal with this emerging threat. A recently emerged Android app claims to be able to detect whether someone is wearing smart glasses nearby: the app scans specific visual or wireless signal characteristics and alerts the user once it determines that there is a suspicious wearable recording device nearby.
Meta emphasized that its smart glasses will light up a small LED indicator when recording to inform people around them that they are filming. However, privacy experts point out that this design has very limited protection in real-world scenarios, especially after researchers have demonstrated that the indicator light can be easily turned off or bypassed.
In the eyes of regulators and privacy advocates, a key question is emerging: Can a reminder buried in a long terms of service be considered "effective notification" of such a large-scale and in-depth collection and manual review of personal data? For those annotators who have to "teach" Meta's AI how to understand the world every day, the answer seems to be self-evident - they know exactly what kind of content they are watching, and they are also aware of the content, and the person involved in the content may never realize it.