The Light on the Frame Won’t Save You: What Meta’s Smart Glasses Are Really Watching

You bought Ray-Bans. You got a surveillance pipeline routed through a Nairobi office. Here’s how it works.

Image generated by the Author

It’s a Tuesday afternoon in Nairobi. A worker sits at a desk in a bright, open-plan office, headphones on, a queue of video clips loading on her screen. She’s been told not to discuss what she sees. She’s been told not to ask questions. There are cameras watching her from the ceiling.

The video playing now shows a bedroom somewhere in Europe. A woman is getting changed. She has no idea she’s being filmed. She definitely has no idea she’s being watched, right now, by a stranger on another continent, earning roughly two dollars an hour, legally bound by a non-disclosure agreement, with no avenue to report what she’s witnessing without risking her job.

The footage came from a pair of Ray-Ban Meta smart glasses.

The Pipeline Nobody’s Talking About

When stories broke this week about Meta workers seeing “disturbing things” through users’ smart glasses, the coverage zeroed in on the shock of the footage itself, the bathrooms, the undressing, the intimate moments accidentally caught on camera. But the more important story isn’t what’s in the footage. It’s the infrastructure that processes it.

Here’s how it actually works.

You put on your Ray-Ban Meta glasses. You say “Hey Meta” and ask the AI to identify something, a restaurant menu, a landmark, a plant. Your glasses record a short video clip. That clip is uploaded to Meta’s servers. So far, so expected.

What most people don’t know is what happens next.

That footage gets routed to Sama, a Kenyan subcontracting firm, where human workers manually watch and label the content tagging objects, annotating scenes, teaching Meta’s AI what it’s looking at. This is how large language models and computer vision systems learn. They need human eyes. They always have. And those human eyes are, in this case, watching a real-time window into Western living rooms, bathrooms, and bedrooms.

According to workers who spoke to Swedish journalists, what they see includes people undressing, using toilets, having sex, and inadvertently filming their bank card details in full. One worker described it plainly: “We see everything, from living rooms to naked bodies.”

Seven Million Pairs. Every One Generating Training Data.

Meta sold approximately seven million pairs of Ray-Ban smart glasses in 2025 alone. Each pair, every time it’s used with the AI assistant, potentially generates footage that flows through this pipeline.

Think about what that number means at scale. Seven million devices, in seven million homes, on seven million faces moving through the world, pointing cameras at families, at strangers, at private moments their owners didn’t even consciously decide to record. All of it processed. All of it labeled. Much of it by workers in Nairobi who have no recourse when what they see disturbs them.

This is not a bug. This is the business model.

The Face-Blurring That Doesn’t Work

Meta does have a system in place to protect the people who appear in this footage. It’s supposed to automatically blur or anonymize faces before the clips reach human annotators.

Workers say it frequently fails.

In certain lighting conditions such as low light, backlighting, high contrast, the anonymization breaks down. Faces that should be obscured are fully visible. The stranger you recorded while asking your glasses to identify a coffee blend at a café? The person walking past your living room window when you asked the AI about your furniture? They can be identified. Their face is clear. A stranger in Nairobi is looking at it right now.

Meta’s automated privacy layer is, by the accounts of the people who see the results of its failures every day, not reliable enough to carry the weight being placed on it.

The Sentence in the Terms of Service Doing All the Legal Work

Here’s what makes this pipeline legally permissible: one sentence buried in Meta’s terms of service.

The company reserves the right to conduct “manual (human) review” of your AI interactions.

That’s it. That’s the consent mechanism. That single clause is the legal architecture allowing intimate footage from Western homes to be routed to a $2/hour labor force operating under NDAs in a Nairobi office with surveillance cameras mounted on the ceiling. You agreed to it when you set up the glasses. You almost certainly didn’t read it. Nobody does.

This isn’t a loophole. It’s a feature. A deliberately opaque feature, wrapped in the blandest possible language, designed to survive legal scrutiny while communicating nothing meaningful to the people it affects.

Sama’s History — and Why It Matters

If the name Sama sounds familiar, it should.

In 2023, TIME magazine published an investigation revealing that Sama had been paying Kenyan workers approximately $2 per hour to view and label some of the most disturbing content on the internet such as graphic violence, child abuse imagery, suicide, to train OpenAI’s content moderation systems. The workers described the psychological damage as severe. Several reported symptoms consistent with PTSD. They used the word “torture.”

The public outcry was significant. Sama terminated that contract with OpenAI.

Then they pivoted to labeling footage for Meta’s smart glasses.

Same workforce. Same pay structure. Same non-disclosure agreements and prohibitions on discussing what they see. The only thing that changed was the source of the content, i.e from social media uploads to live footage from Ray-Bans on faces across Europe and North America.

Sama has also faced lawsuits related to union-busting practices in Kenya. Workers who tried to organize for better pay and mental health support reported retaliation. The company that Meta chose to process your bedroom footage has a documented history of suppressing the voices of the people it employs to do its most sensitive work.

“Designed With Your Privacy in Mind”

That’s the phrase Meta uses to market the Ray-Ban glasses. Designed with your privacy in mind.

The main privacy feature visible to the end user is a small LED light on the frame that blinks when the camera is recording. Studies and anecdotal reports consistently show that most bystanders don’t notice it, don’t know what it means when they do, and have no way to object to being recorded regardless.

The LED light is doing about as much for your privacy as the terms of service nobody reads.

Meanwhile, the actual privacy infrastructure is the anonymization system that’s supposed to protect people captured in your footage is failing in the field, according to the workers responsible for handling what it fails to protect.

The gap between the marketing language and the operational reality is not a minor inconsistency. It’s a fundamental misrepresentation of what the product does.

What’s Coming Next Should Concern You More

The current generation of Ray-Ban Meta glasses does not have real-time facial recognition. The next generation reportedly will.

Pause on that for a moment.

A system that currently cannot reliably blur faces in its training data pipeline is a system where, by its own worker’s accounts, clearly visible faces regularly pass through human review is planning to add the ability to intentionally identify those faces on demand.

The company that brought you “accidentally filmed your neighbor undressed, and a stranger in Nairobi saw her face” wants to move to “intentionally identified your neighbor by name as you walked past her on the street.”

The implications for public anonymity, for stalking, for domestic abuse situations, for political dissidents, for anyone who has ever needed to move through the world without being tracked — these are not hypotheticals. They are the logical endpoint of a roadmap Meta is already on.

The Invisible Labor Behind AI Convenience

There is something worth sitting with in the structure of this story that goes beyond the specific privacy failures.

Every time we interact with an AI system, every image generator, every chatbot, every computer vision tool, there is human labor behind it that most of us never see. Somewhere, someone looked at footage or text or images that needed to be labeled, categorized, filtered. Somewhere, someone was paid very little to do work that was sometimes harmful, so that the product could feel effortless on our end.

The workers in Nairobi are not the villains of this story. They are doing a job they need, under conditions they didn’t fully choose, seeing things they didn’t sign up to see, with almost no power to change any of it.

The people unknowingly filmed in their most private moments are not at fault. They are collateral in a product design that treats their privacy as a detail to be managed rather than a right to be protected.

The people who bought the glasses are not naive. They were told, explicitly, by the company that made the product, that it was designed with their privacy in mind.

The question worth asking is: who actually benefits from the structure as it currently exists? Who profits from the opaque terms of service, the offshored labor, the failed anonymization, and the marketing language that emphasizes the LED light while burying the data pipeline?

What You Can Do

If you own a pair of Ray-Ban Meta glasses, the most important thing you can do right now is read Meta’s actual terms of service not to find a loophole, but to understand what you’ve agreed to and make an informed decision about whether you’re comfortable with it.

Regulators in the EU are already looking at this more closely than their counterparts elsewhere, which may produce some accountability in the medium term. Consumer pressure matters too. Meta responds to user behavior and public attention, even when it doesn’t respond to criticism.

And if you work in tech, if you build products that use human data labeling, if you make decisions about data pipelines and subcontracting, this is a story about the choices made by people just like you. The conditions in those Nairobi offices didn’t happen by accident. They happened because a series of humans made decisions that prioritized efficiency and cost over the wellbeing of the workers, the privacy of the users, and transparency with the public.

Those decisions can be made differently.

The LED light on the frame is not your privacy. It’s a symbol of privacy, a small, blinking performance of care, visible in pictures on a product page, designed to satisfy a regulatory checkbox and move on.

Real privacy design would start with informed consent. It would include anonymization that works. It would mean paying the people who handle your most sensitive footage a wage that reflects the weight of what they’re asked to do, and giving them the right to speak when what they see disturbs them.

Until then, the light blinks. And somewhere, a queue of your footage loads on someone else’s screen.