AI smart glasses and privacy: what they capture and what the law allows

What AI smart glasses can do in 2026

When Meta introduced Meta Ray-Ban Display in September 2025, it sorted its glasses into three kinds: camera AI glasses (Ray-Ban Meta and Oakley Meta), display AI glasses (Meta Ray-Ban Display) and full augmented reality glasses, which it still showed only as a prototype. A fourth kind arrived in September 2026 with Ray-Ban Meta Audio, which Meta calls its first audio glasses. Whatever the label, the core features are the same:

• A hands-free camera for photos and video, and on Meta's glasses, livestreaming to Facebook or Instagram.
• A multimodal assistant you wake with "Hey Meta" or "Hey Google" and ask about what the camera sees. In September 2024 Meta announced continuous, video-based help, so you can keep asking as you walk.
• Live translation. Meta announced it in September 2024 for Spanish, French and Italian into English, and added German and Portuguese in September 2025.
• Open-ear audio for calls, messages and music, plus conversation focus, announced in September 2025 for Ray-Ban Meta and Oakley Meta HSTN, which amplifies the person you are talking to.
• On display models, captions, messages, turn-by-turn directions and answers shown in the lens, controlled on Meta Ray-Ban Display by an EMG wristband that reads small finger movements.

Prices are in US dollars at announcement. The direction of travel is clear from the latest launches: from answering questions toward acting on them. Meta said at Connect on September 23, 2026 that its Muse personal agent will come to its glasses in the coming months and act on what the wearer is looking at, and Google says Gemini on its glasses will handle multi-step tasks such as preparing a coffee order. Snap's announcement adds a service, SPECS Intelligence, that learns your goals and routines from the apps you connect. The more the glasses do, the more of an ordinary day passes through a camera and a microphone connected to someone's cloud. If you want the technical side of overlaying digital content on the world, our guide to augmented reality programming explains how AR apps track and render.

Where the photos, voice and video go

The privacy questions are different for each thing the glasses do, so it helps to follow the data. Meta sets it out in its supplemental privacy policy (the version in effect since June 29, 2026), its voice privacy notice and its privacy page for the glasses.

The AI query is the case people miss. When you ask the assistant about a menu, a plant or a landmark, the camera switches on briefly and the image goes to Meta to produce the answer. Meta's policy says images and voice recordings from Meta AI may be used to improve its products, including the models behind Meta AI, and that a prompt or photo can be shared with third parties described in its AI terms.

Voice has its own rules. Meta's voice notice says voice interactions include anything said when the assistant is triggered, even by mistake, plus background sound, which can mean the people around you. Mistaken activations are deleted within 90 days of being detected; other stored recordings are kept for up to a year unless you delete them sooner. You can turn storage off in the Meta AI app, but the June 2026 policy is explicit that Meta AI on the glasses stores transcripts and recordings by default.

Snap says its SPECS do not continuously record, and that personal content from accounts connected to SPECS Intelligence is not used to train or fine-tune its models or to target ads. Google says it has been gathering privacy feedback on its prototypes from trusted testers; its announcements so far do not describe how its glasses will signal recording.

What the capture LED shows, and what it misses

Meta's AI glasses with cameras each have a white capture LED on the front of the frame. It blinks briefly for a photo and keeps blinking for as long as a video records, and it has no off switch. On current models the camera is disabled when the glasses detect that the light is covered, and in July 2026 Meta said it was updating the glasses to disable the camera when the LED has been physically tampered with or destroyed. It also said it removes ads and posts for LED tampering services and takes legal action against people who sell them.

The weakness is attention, not engineering. In The Verge's testing, reported in October 2024, the light on Ray-Ban Meta glasses was hard to notice outdoors in bright light, and people often did not notice they were being filmed, especially in crowded places. Meta says it tested the brightness so the light is visible during the day. Both can be true: a light only protects people who know to look for it.

Two more limits matter. Meta presents the light as the signal for photos and video; it does not describe it as a signal for voice commands, calls or live captions, which use the microphones. And the glasses look like any other frames, which is what makes them easy to wear and hard to spot. Snap puts a "prominent" LED in the centre of the SPECS frame, a design difference worth comparing when you choose a model.

How the I-XRAY demo identified strangers

In October 2024, two Harvard students, AnhPhu Nguyen and Caine Ardayfio, published a demonstration they called I-XRAY. Wearing Ray-Ban Meta glasses, they looked up the names, addresses and relatives of classmates in real time, and approached strangers on public transit as if they knew them. According to The Verge and the students' own write-up, the chain worked like this, in principle:

1. The glasses livestreamed video, as they are designed to.
2. A program watching the stream picked out faces.
3. A reverse face search engine found other photos of the same face online, with the pages they appeared on.
4. A large language model read those pages and settled on the person's name.
5. A people-search site turned the name into a home address, phone number and relatives, and the results appeared in a phone app.

The students did not release the tool. Their write-up says they built it to show what is already possible, deliberately withholds the specific technologies in the pipeline, and ends with a guide to removing yourself from the reverse face search engines and people-search sites that make it work. The new ingredient, they argued, was the language model, which connects pages about the same person automatically, with no one doing the searching by hand.

The lesson for everyone else is that nothing in the glasses was hacked. The camera was only the input; the identification came from face search engines built on photos already online and from people-search sites that publish addresses. The same profile is also raw material for a convincing approach by a stranger, the kind of social engineering covered in our post on how attackers use AI.

Can smart glasses recognize faces?

Not as a feature Meta or Snap offers, as of September 2026. Meta says faces and licence plates are blurred automatically in the images it uses to answer AI questions, and Snap says its SPECS do not use facial recognition to identify people. The line is an old one: Canada's privacy commissioner noted in 2014 that Google had committed, at the time, not to deploy facial recognition apps on Google Glass after privacy authorities questioned it.

It helps to separate two uses. Canada's privacy commissioner describes verification as a one-to-one comparison, such as unlocking a phone with your face, and identification as one-to-many: comparing a face with a database of templates to find out who someone is. Identification needs a large database of named faces, and that is what the controversy is about.

Regulators have been blunt about building such databases from the open web. In February 2021, the privacy commissioners of Canada, Quebec, British Columbia and Alberta found that Clearview AI's scraping of over three billion images to build a facial recognition tool amounted to mass identification and surveillance, and that photos posted online and reused for an unrelated purpose are not exempt as "publicly available". Since February 2, 2025, the EU AI Act has prohibited AI systems that create or expand facial recognition databases through untargeted scraping of facial images from the internet or CCTV footage.

Is it legal to record people with smart glasses?

It depends on three things: whether you capture a conversation or only images, where you are, and what you do with the recording afterwards. This is general information, not legal advice.

The practical reading: recording a conversation you are part of is generally not interception in Canada or under US federal law, but recording a private conversation between other people, without either one's consent, can be. In California, even a participant needs everyone's consent when the conversation is confidential, meaning the circumstances suggest the parties want it kept among themselves.

In the EU, the GDPR does not apply to purely personal or household activity, and the European Data Protection Board reads that exemption narrowly. A tourist who films a holiday and shows the footage to friends and family is covered; footage published online to an indefinite number of people is not. Once you post a clip of strangers, the GDPR can apply to what you do with it.

Meta's own terms put the burden on the wearer. Its supplemental terms of service say you are responsible for complying with all applicable laws, including giving notice or getting consent where recording, biometric or privacy laws require it.

Facial recognition and biometric privacy law

The GDPR, BIPA and Canada's privacy commissioner all treat an ordinary photo as something other than biometric data. The line is crossed when software measures a face to identify the person, which is what face search engines and tools like I-XRAY do.

Three details are worth knowing. The GDPR's recital 51 says photographs are not automatically special-category data; they become biometric data only when processed to identify or authenticate someone. The European Data Protection Board says a private organization that runs facial recognition for its own purposes, such as marketing or security, will in most cases need explicit consent from everyone captured. And BIPA, which applies to individuals as well as companies, gives anyone affected a private right of action: liquidated damages of $1,000 per negligent violation and $5,000 per intentional or reckless one, or actual damages if higher. An amendment in force since August 2, 2024 (Public Act 103-769) counts repeated collection from the same person by the same method as a single violation.

If you wear AI glasses

The wearer is the one the law and the maker's terms hold responsible. A few habits cover most of it:

1. Tell people, and stop when asked. Meta's own advice is to show people how the capture LED works and to stop recording if anyone says they would rather you did not.
2. Switch the glasses off in private places. Meta lists doctors' offices, locker rooms and schools; add washrooms, clinics and anywhere a phone camera would be out of place.
3. Do not record conversations you are not part of, and in states that require every party's consent, ask before recording your own.
4. Check the settings in the Meta AI app. Decide whether voice recordings are stored, whether photos go to the cloud for processing, and delete stored voice interactions you do not want kept.
5. Think before asking the assistant about people, screens or documents. The image goes to Meta, and a photo of a colleague's screen or a client's paperwork is their data, not yours.
6. Ask before you post a clip with other people in it. Sharing with family is one thing; publishing to everyone takes you outside the GDPR's household exemption.
7. Never run strangers' faces through a face search engine. Beyond the ethics, it is exactly the kind of identification that biometric and privacy laws in Illinois, the EU and Canada restrict.

If someone near you is wearing them

You cannot control another person's device, but you can make yourself harder to identify and easier to protect:

• Look for the light. On Meta's glasses it is a white LED on the front of the frame, blinking during photos and video. In bright sun it is easy to miss.
• Ask. Meta's guidance tells its own wearers to stop recording when asked, so the request has the maker's backing.
• Opt out of reverse face search and people-search sites. The I-XRAY students listed the free removal forms for PimEyes and FaceCheck.ID and the opt-outs of several people-search sites, noting that removal is what makes tools like theirs ineffective. As The Verge points out, you can only make your information less available, not delete it everywhere.
• Limit public, tagged photos of your face, since they are what a face search engine matches against. Our guide to using social media safely shows the tagging and visibility settings on each platform.
• Be skeptical of a stranger who knows too much. What the I-XRAY pipeline returned was enough for the students to chat with strangers on public transit as if they already knew them.

Writing a smart glasses policy for your business

Glasses bring a camera, a microphone and an AI assistant into places where phones usually stay in pockets: meetings, open offices with client files on screens, and counters where customers show ID. The risks are specific. A photo or a question to the assistant can carry client data into a personal AI account. Live captions may process what colleagues say in a meeting on the maker's servers. In the US, Meta Ray-Ban Display can now join meetings from the calendar and show who is attending. And work is not personal use: the GDPR's household exemption covers only activity with no connection to a professional or commercial activity, and in Canada the privacy commissioner warned as early as 2014 that employers could contravene PIPEDA's consent and transparency requirements if staff wearing such devices captured customers' personal information too broadly.

Write the rules by place and by capability, not by brand. Google's "audio glasses" take photos, so a label is not a guarantee that a device has no camera.

For customer areas where the business itself records, the European Data Protection Board's guidance on video devices is a useful model: a sign at eye level before people enter the recorded area, stating the purpose, who is responsible and where to find the full notice, including anything that might surprise them, such as footage going to a third party.

A workable policy covers these points, in this order:

1. Scope by capability: camera, microphone, display and AI assistant, whatever the brand.
2. Zones and signage, as in the table above.
3. Meetings: announce glasses the way you would a phone that records, and get everyone's agreement before recording or captioning.
4. AI use: no client or personal data in personal AI accounts, the same rule you apply to consumer chatbots.
5. Accessibility: allow the glasses as eyewear with the camera off. Many frames take prescription lenses, and Meta's glasses work with Be My Eyes, a service that connects blind and low-vision people with sighted volunteers, so a blanket ban can lock people out of their own glasses.
6. Company-issued glasses: read the maker's terms first (Meta's supplemental terms limit use to personal, non-commercial purposes unless Meta permits otherwise), then record the devices in your asset register and include them in offboarding.
7. Training and review: a short briefing when the policy starts, and a review whenever a new capability ships.

The AI rule is usually the gap. If staff already paste client data into personal AI accounts, glasses add one more way in. Our AI and automation service starts with a review of one real workflow and builds guardrails such as removing personal data before a model reads it. Company-issued glasses belong in the same records as laptops and phones: our IT support team keeps an asset and licence register and runs offboarding in one documented sequence, so devices and access end together.