On February 27, 2026, Swedish newspapers Svenska Dagbladet and Göteborgs-Posten published the results of a joint investigation into how Meta trains the AI that powers its Ray-Ban smart glasses. Their reporters interviewed more than 30 employees at Sama, a data annotation company headquartered in San Francisco with operations in Nairobi, Kenya. What the workers described was not an anomaly in Meta’s system. It was the system working as designed.
“We see everything,” one worker told the Swedish journalists. “From living rooms to naked bodies.”
A month later, a U.S. class action lawsuit was filed. The UK’s Information Commissioner’s Office sent a formal letter demanding answers. Multiple European Parliament members submitted questions to the European Commission. Meta sold over 7 million pairs of these glasses in 2025. The majority of buyers had no idea that saying “Hey Meta” could route footage of their bedroom to a contractor in Nairobi.
The Pipeline: From Your Voice Command to a Kenyan Annotation Floor
The technical pathway is straightforward, and that is the problem. Meta’s Ray-Ban glasses contain ultra-wide cameras, microphone arrays, and an AI assistant activated by the wake phrase “Hey Meta.” When a user activates the AI, asking it to identify an object, describe a scene, or answer a question about their surroundings, the glasses capture video and audio and transmit it to Meta’s cloud infrastructure for processing.
Meta’s AI Terms of Service state that interactions “may be automated or manual (human).” In the U.S. version of the policy, this disclosure is buried deep in supplemental terms that most users never read. In the UK version, the BBC found a mention of human review in Meta’s AI terms of service, but not in the main product marketing. The glasses themselves are marketed with phrases like “designed for privacy, controlled by you” and “built for your privacy.”
Once footage reaches Meta’s servers, a subset is routed to human data annotators for labeling. The annotators’ job is to draw bounding boxes around objects, assign category labels (“plant,” “vehicle,” “furniture”), and perform quality assurance on the visual data. This labeled data is then used to train and improve Meta’s multimodal AI models. The annotators doing this work are employed by Sama in Nairobi.
There is no way to use the glasses’ multimodal AI features without triggering this pipeline. As Engadget noted in its review: “images of your surroundings processed for the glasses’ multimodal features like Live AI can be used for training purposes (these images aren’t saved to your device’s camera roll).” Meta’s defense that footage “stays on the user’s device unless they choose to share it” is technically true for photos and videos stored in the camera roll. It is misleading for AI interactions, which by definition require cloud processing and can result in human review.
In April 2025, Meta quietly updated its privacy policy to make AI features the default, with voice recordings retained on Meta’s servers for up to a year. The Clarkson Law Firm’s complaint alleges there was “no real way to opt out.”
What the Workers Actually Saw
The Svenska Dagbladet and Göteborgs-Posten investigation documented specific examples from worker interviews that went far beyond routine object labeling.
One contractor described a video in which a man placed his Meta glasses on a bedside table and left the room. His wife then entered and changed her clothes in front of the camera, apparently unaware that the glasses were still recording and that the footage would be reviewed by strangers in another country.
Other workers described footage of people using the toilet, users watching pornography while wearing the glasses, explicit sexual content filmed by wearers, bank cards and personal financial documents visible on screen, and private conversations about relationships, politics, and alleged criminal activity.
“I don’t think they know,” one contractor told the Swedish newspapers, “because if they knew, they wouldn’t be recording.”
Workers said the material made them uncomfortable, but that the pay kept them in their seats. Offices were monitored by cameras. Personal phones were banned. Strict non-disclosure agreements were in effect. Several workers agreed to speak to the Swedish journalists only under conditions of anonymity, aware that losing their jobs could mean financial ruin.
The Face-Blurring Failure
Meta has stated that it uses AI to blur faces in footage before it reaches annotation teams. Workers at Sama disputed this. The blurring, they said, did not consistently work. A former Meta employee confirmed to the Swedish outlets that the anonymization algorithms “sometimes miss,” particularly under “difficult lighting conditions.”
Workers reported that poor lighting, rapid movement, and unusual camera angles frequently defeated the automated blurring system. The result was not isolated failures but regular exposure to unblurred, identifiable faces and bodies in footage from people’s homes.
This creates a structural paradox. Meta markets the face-blurring system as a privacy safeguard, leading users to believe their recordings are anonymized before anyone sees them. But the blurring operates on the same footage that was captured without the knowledge or consent of bystanders. A system that claims to protect people it has already exposed is not a safeguard. It is damage control, and when the damage control fails in low light, there is no fallback.
Sama: The Same Company, Different Trauma
Sama’s involvement is not incidental. It is the most damning element of the story, because the company has a documented history of exactly this kind of harm in its previous work for Meta.
From 2019 to 2023, Sama served as Meta’s largest content moderation provider in Africa. Workers in Nairobi reviewed Facebook and Instagram posts containing violence, child sexual abuse, terrorism, self-harm, and hate speech. A TIME investigation in 2022 documented low pay ($1.46 to $3.74 per hour), psychological trauma, and alleged union-busting at the Nairobi office.
In 2022, a former content moderator named Daniel Motaung sued Meta and Sama in Kenya, alleging forced labor, human trafficking, unfair labor practices, and failure to provide adequate mental health support. Motaung had been fired in 2019 after organizing a strike and attempting to unionize. The lawsuit grew into a class action involving more than 185 former moderators.
In December 2024, CNN reported the medical results: Dr. Ian Kanyanya, head of mental health services at Kenyatta National Hospital, assessed 144 of the former moderators. Of those assessed, 81% were classified as suffering from “severe” PTSD. The content they had been required to review included, in Kanyanya’s words, “gruesome murders, self-harm, suicides, attempted suicides, sexual violence, explicit sexual content, child physical and sexual abuse, horrific violent actions.”
In January 2023, Sama announced it was “discontinuing” content moderation for Meta. It would refocus on its “core business”: computer vision data annotation. This was presented as a strategic pivot away from the controversy.
Computer vision data annotation for Meta’s AI glasses is, functionally, the same work. Workers sit in a monitored office in Nairobi, view footage captured by Meta’s products, and label what they see. The content has changed from Facebook posts to wearable camera recordings. The working conditions, the non-disclosure agreements, the camera-monitored offices, the ban on personal phones, and the underlying labor structure have not. Sama did not stop doing trauma-inducing work for Meta. It changed which Meta product the trauma comes from.
The Legal Response
On March 4, 2026, one week after the Swedish investigation published, the Clarkson Law Firm filed a class action complaint in the U.S. District Court for the Northern District of California. The plaintiffs are Gina Bartone of New Jersey and Mateo Canu of California. The defendants are Meta Platforms, Inc. and Luxottica of America, Inc. (the manufacturer of the glasses under the Ray-Ban brand).
The complaint alleges false advertising and violation of consumer protection laws. Its core argument: no reasonable consumer would interpret “designed for privacy, controlled by you” to mean that footage from their bedrooms would be reviewed by overseas contractors. The lawsuit seeks monetary damages and injunctive relief on behalf of a nationwide class of purchasers.
“You cannot market a product as ‘built for privacy’ and then funnel footage of people’s intimate moments to contract workers without their knowledge,” said Yana Hart, partner at Clarkson Law Firm. “Meta made privacy the centerpiece of its marketing campaign because it knew consumers would never buy these glasses if they knew the truth.”
On March 5, 2026, the UK’s Information Commissioner’s Office confirmed it had formally written to Meta demanding information about how the company meets its obligations under UK data protection law. The ICO’s statement was direct: “Devices processing personal data, including smart glasses, should put users in control and provide appropriate transparency. This includes where user data is used to train or develop AI systems.” Under UK GDPR, the ICO can impose fines of up to 4% of global annual turnover.
At the EU level, data protection lawyer Kleanthi Sardeli of the non-profit None Of Your Business (NOYB) told the Swedish journalists that if footage captured by European users flows through Meta’s infrastructure to contractors in Kenya, “both transparency and a legal basis for the processing are lacking.” Kenya does not have an EU adequacy decision, meaning cross-border data transfers require specific safeguards that multiple experts interviewed for the investigation questioned whether Meta had in place.
Meta’s Defense and Why It Does Not Hold
Meta spokesperson Christopher Sgro responded to the investigation: “Ray-Ban Meta glasses help you use AI, hands-free, to answer questions about the world around you. Unless users choose to share media they’ve captured with Meta or others, that media stays on the user’s device. When people share content with Meta AI, we sometimes use contractors to review this data for the purpose of improving people’s experience, as many other companies do. We take steps to filter this data to protect people’s privacy and to help prevent identifying information from being reviewed.”
Three specific claims in this statement are challenged by the evidence.
“Unless users choose to share media”: Using the AI features of the glasses constitutes “sharing” with Meta. There is no way to use “Hey Meta,” the Look and Tell feature, or any multimodal AI function without transmitting data to Meta’s servers. The choice is binary: use the product’s primary advertised features, or do not use them. There is no middle option where the features work without the data pipeline.
“We take steps to filter this data”: The face-blurring system fails regularly according to both current workers and a former Meta employee. The filtering is not a guarantee. It is a best-effort system operating on footage captured from wearable cameras in variable lighting conditions, exactly the conditions where computer vision systems fail most often.
“As many other companies do”: The appeal to industry norms does not address the specific allegation. The complaint is not that human review exists as a practice. The complaint is that Meta marketed these glasses as privacy-first products while operating a data pipeline that routes intimate footage to overseas contractors. The industry comparison is irrelevant to the false advertising claim.
Meta did not respond to the Swedish newspapers’ questions for two months. When it finally replied, the company referred them to its terms of use and privacy policy.
The Broader Pattern
Meta’s smart glasses data pipeline is one instance of a structural pattern across the AI industry. AI agents are already acting against user instructions in documented cases. The training data that makes those agents work comes from labor practices like the ones documented in Nairobi. The workers who label the data are the invisible infrastructure of the AI economy.
OpenAI used Sama to label toxic content for ChatGPT’s safety systems. Workers earned less than $2 per hour and described the experience as “torture.” That reporting, by TIME in January 2023, was one of the first investigations to document the human cost of AI data labeling. Three years later, the same company is performing the same category of work for the same client’s new product line.
The content moderation crisis of 2019-2023 established that outsourcing exposure to harmful content to workers in low-wage countries produces psychological damage. Sama’s own workers were diagnosed with PTSD at an 81% rate. The company’s response was to exit content moderation and pivot to computer vision annotation. But when the computer vision data comes from cameras worn inside people’s homes, the distinction between “content moderation” and “data annotation” collapses.
As the regulatory framework for AI agents develops, the labor conditions of the workers training those agents remain largely unaddressed. The UK CMA’s agentic AI framework focuses on consumer protection. The CLTR scheming study focuses on agent behavior. The Meta glasses investigation focuses on user privacy. None of them adequately address the people at the end of the data pipeline.
Meta is preparing to launch two new Ray-Ban models (codenames Scriber and Blazer, FCC filings dated March 10, 2026) with higher model numbers suggesting a major hardware upgrade, including Wi-Fi 6 support. The company is also navigating the shift to open AI assistant platforms on iOS 27, where Siri Extensions may allow multiple AI providers to compete for voice queries currently routed to Meta’s assistant.
Over 7 million people are wearing cameras on their faces. Their footage feeds a pipeline that ends in a monitored office in Nairobi where workers under NDA label what they see, including what users never intended anyone to see. The glasses are marketed as privacy-first. The pipeline is designed for data extraction. Both of those things are true, and the gap between them is where the lawsuit, the regulatory inquiry, and the next generation of wearable AI will be decided.
Sources: Svenska Dagbladet and Göteborgs-Posten joint investigation (February 27, 2026), Clarkson Law Firm class action filing (March 4, 2026), TechCrunch (lawsuit details), The Register (ICO investigation), PetaPixel (complaint specifics), TIME (Sama exits content moderation, 2023), CNN (PTSD diagnoses, December 2024), Sama (Wikipedia), Fortune (privacy analysis), Engadget (review disclosures).