What happened

An investigation by Swedish media outlets Svenska Dagbladet and Göteborgs-Posten revealed that thousands of data annotators working for Meta’s subcontractor, Sama, in Nairobi, Kenya, are routinely reviewing raw footage captured by Ray-Ban Meta smart glasses. 

The footage includes accidental and intimate recordings of users in bathrooms, undressing, or engaging in sexual activity, and watching pornography. In some instances, reviewers say they can see sensitive data like bank cards, including account numbers.

In addition, non-consenting third parties such as partners, family members, and bystanders are being captured on footage and having their images reviewed by strangers without any knowledge or agreement. 

While Meta claimed to use AI-powered blurring to protect identities, workers reported that the anonymisation frequently failed, leaving faces and private environments clearly visible.

Why it happened

The root cause lies in a combination of inadequate transparency, weak informed consent, and the structural realities of AI model training.

If a wearer wants to make use of the AI features, they must agree to Meta's terms of service that allow any data captured to be reviewed by humans, because Meta's large language models often require people to annotate visual data so the AI can understand it and build training models. 

However, contractors said Meta failed to properly inform users that their footage was being seen by humans, not just automated systems.

Buried in Meta's AI terms of use, the company reserves the right to review user interactions with AI, including conversation content. The document also warned that users shouldn't share information they don't want the AI to use. But, given the kind of footage being reviewed, many users appeared wholly unaware of this. 

Whistleblowers said that many of the videos appear to be moments captured when users were not aware they were being recorded. 

Compounding the problem, while Meta stated that data is first filtered to protect privacy and that faces appearing in annotation data are automatically blurred, Kenyan workers claimed the anonymisation tools did not always function properly, with faces and bodies sometimes visible, especially in difficult lighting conditions.

There was also a structural suppression of internal concern. Contractors described feeling pressured to process disturbing content without questioning ethical implications, with one worker explaining: "You are not supposed to question it. If you start asking questions, you are gone." This created an environment where legitimate privacy concerns could not be raised by the very people witnessing them.

Once material has been fed into the models, the user in practice loses control over how it is used, according to a data protection lawyer at the non-profit None Of Your Business.