Smart glasses with cameras raise an obvious privacy question for the people around the wearer. In early 2026, a different question came to the fore: what happens to the footage after the wearer asks the built-in AI about what they’re seeing? An investigation found that some of it was being watched by human contractors, and some of what they saw was deeply private.
Here’s what the reporting found, how Meta responded, what’s happening legally, and what users and developers should take from it.
What the investigation found
Two Swedish newspapers, Svenska Dagbladet and Göteborgs-Posten, investigated how footage from Ray-Ban Meta smart glasses is used to improve Meta’s AI. According to Help Net Security’s summary, they found that data annotators at Sama, a Meta subcontractor in Nairobi, Kenya, review footage from the glasses to label it for AI training.
Workers described seeing material that users almost certainly never expected a person to see: people in bathrooms, people undressing, sexual content and bank card details. They also reported faces that weren’t blurred despite anonymisation policies. One worker told reporters, “You understand that it is someone’s private life you are looking at, but at the same time you are just expected to carry out the work.”
The journalists also found that staff in Swedish eyewear shops selling the glasses often believed the data stayed on the device, which suggests buyers weren’t getting a clear picture at the point of sale.
What Meta says
Meta’s position, as reported by Fortune, is that human review applies when people share content with Meta AI. A spokesperson said that “like other companies we sometimes use contractors to review this data to improve people’s experience,” that this is disclosed in its privacy policies, and that “data is first filtered to protect people’s privacy.”
Meta’s policies do say content “may be subject to human review.” What they don’t spell out, according to the investigation, is where that review happens, who does it or what safeguards apply. The investigators said Meta took two months to reply, with a written statement that didn’t directly answer their questions.
The legal fallout
- US class action: On 4 March 2026, a class action was filed in federal court in California against Meta and Luxottica of America. It alleges that users weren’t told footage could be sent to overseas contractors for manual labelling, despite marketing that said the glasses were “built with privacy at their core.”
- Kenya: Kenya’s data protection regulator opened an inquiry into the glasses’ data practices, as Techweez reported.
- Elsewhere: privacy regulators in other countries, including the UK, were also reported to be looking at the issue.
None of these had reached a final outcome in the sources cited here, so this story is still developing.
Why human review happens at all
It’s worth understanding this without assuming bad intent. Nearly every AI assistant relies on some human review. People label examples so models learn, check outputs for errors, and flag harmful content. Voice assistants went through the same controversy around 2019, when it emerged that contractors listened to some Siri, Alexa and Google Assistant recordings. Companies responded by adding opt-outs and clearer disclosures.
What makes smart glasses different is the nature of the data. A voice query is usually short and deliberate. A camera on your face sees whatever you’re looking at, including other people who never agreed to anything, in places like homes and bathrooms. The privacy stakes are simply higher.
What owners can do
If you own Ray-Ban Meta glasses or similar devices, a few practical steps reduce your exposure:
- Review the AI and data-sharing settings in the companion app. Look for options related to sharing data to improve Meta’s products or AI, and for how long voice and media are stored. Menu names change between app versions, so read through the privacy section rather than relying on a fixed path.
- Understand what triggers sharing. Asking the AI about what you’re seeing sends images or video to the cloud. Photos you simply take and keep are handled differently from content you share with the AI assistant.
- Be deliberate about where you use AI features. Avoid asking the assistant about scenes that include private spaces, documents, cards or other people who haven’t agreed.
- Delete stored interactions regularly if the app lets you.
- Take the glasses off in sensitive places. Folding them or putting them in the case is the only certain way to know the camera isn’t capturing anything.
The lesson for developers building AI features
If you build products with AI in them, this case is a clear example of how “technically disclosed” isn’t the same as “understood.” A few principles are worth applying:
- Say plainly when humans may see data. A line buried in a privacy policy doesn’t count as informing users. Say it at the moment of sharing, in plain language.
- Make human review opt-in for sensitive data types. Images and video from personal devices should default to not being used for training or review.
- Filter before review, and prove it works. If you promise to blur faces or remove sensitive content before humans see it, test that pipeline and monitor its failure rate.
- Know your vendors. If annotation is outsourced, your contracts should cover where data goes, who can access it, retention and worker protections. Annotators exposed to disturbing content need support too.
- Train the people who sell your product. If retail staff can’t explain where data goes, customers can’t give meaningful consent.
The bottom line
Smart glasses are becoming genuinely useful, and on-device and cloud AI will make them more so. But a camera on your face creates a data stream more intimate than almost anything else we carry. The Ray-Ban Meta controversy shows that users’ assumptions about who sees that data can be very different from reality. Until companies close that gap clearly and by default, the safest assumption for users is simple: anything you ask the AI about may be seen by a person.
Related reading
- The Courtsdesk Shutdown: Open Justice, Court Data and an AI Vendor
- Age Verification in the Operating System
- Can Terms of Service Be Updated by Email?
Sources: Help Net Security, Fortune, Techweez.
