Omdia counted 4.2 million AI glasses shipped worldwide in the first half of 2026, up 127% from a year earlier. Meta shipped 3.4 million of them, 81% of the market, almost all camera-equipped Ray-Ban and Oakley frames. The United States took 53.7% of global shipments. So in six months, more than two million cameras went onto American faces, built to look like ordinary sunglasses.
The same week, Liquid AI and Qualcomm announced Liquid Context, a layer that learns a user’s routines, preferences and needs from signals on the device. It keeps the resulting profile on Qualcomm’s Hexagon NPU and passes pieces of it to AI agents when the user allows. The agent runs on a 2.6-billion-parameter model. A 3-billion-parameter vision model is part of the stack. None of it needs a cloud round trip.
Put the two releases together and you have a problem for anyone who does collection or counter-collection. A camera nobody notices, paired with a model that doesn’t need a server, is a sensor that leaves no network trail.
Security agencies have already drawn the line
Institutions have moved faster than lawmakers. The US Air Force banned Meta AI glasses in uniform on January 9, citing operational security, and the 104th Fighter Wing extended the ban to civilians and to personnel out of uniform. The Army, Navy, Marines and Space Force left it to commanders. The Army is even testing the glasses for vehicle maintenance training. The NYPD’s counterterrorism unit circulated a memo in January. Fusion centers in New York, Maine, Virginia, Tennessee and San Diego have followed with bulletins about covert recording inside detention facilities: cell layouts, camera placements, patrol routines. They pointed to a July 2025 TikTok video filmed inside a South Carolina detention center. DEF CON 34 banned camera glasses outright in August.
The operational record is already long. Shamsud-Din Jabbar wore Meta glasses on a scouting trip through the French Quarter before the New Orleans truck attack on January 1, 2025. In October 2024, two Harvard students paired Ray-Ban Metas with PimEyes and public records and identified strangers on the street. From August 2025, Border Patrol agents were documented wearing the glasses on immigration raids in at least six states, recording members of the public.
That last case shows what the state actually objects to. ICE acting director David Venturella signed a memo on August 20 banning Meta glasses for all ICE employees at work. His reason was that they could “capture, record, or transmit sensitive information.” At the same time, DHS is seeking $7.5 million to build its own AI glasses with biometric facial identification, tied into Mobile Fortify, the face-matching app ICE agents already carry. So the objection is to where the footage lands. Identification from a pair of glasses is something DHS wants for itself.
On-device inference removes the choke point
So far, the main safeguard on camera glasses has been the vendor. Meta’s unreleased face-recognition feature, internally called NameTag, was found as dormant code in the Meta AI companion app, which runs on more than 50 million devices. EPIC asked the FTC and state attorneys general to block it. Senator Markey wrote to the company. Meta’s line at Connect this month was that NameTag “has not been deployed to consumers.” That promise means something only because identification would run through Meta’s cloud. There, it can be governed, logged, rate-limited and subpoenaed.
On-device models take that choke point away. Meta’s own VR Glasses, due in spring 2027, run Qualcomm’s Snapdragon Reality Elite at 48 TOPS and are marketed as able to run large vision models “without a cloud round-trip.” Liquid AI’s pitch is the same thing sold as privacy: “the personal context stays on the device.” For the wearer, that’s a real gain. For everyone the wearer looks at, the audit trail disappears. A face-matching model with a downloaded gallery can run on the frame. No request crosses the network, so there’s nothing to log, nothing to block at the vendor’s policy layer, and no provider to serve with legal process. Once the model is local, the rules are whatever the person loading it decides.
The recording LED isn’t much of a defence either. Pennsylvania has proposed making a visible capture indicator mandatory. Some owners already pay to have it disabled, and the frames look like any other sunglasses.
The same devices hold a profile of the wearer
The personal context layer has a second side that gets less attention. Liquid Context exists to build and maintain a model of the user: routines, relationships, places, habits. Its examples are rescheduling a meeting when a child needs an early school pickup, or changing a car’s cabin temperature after a workout. That’s a pattern-of-life file, kept on the device and updated all the time. A seized or compromised pair of glasses, or the phone it pairs with, would give up a lot more than photos. It would give up the device’s own summary of who the wearer is and where they go. Security teams used to worry about glasses as a camera pointed out at the world. Now they also have to treat them as a dossier on the person wearing them.
For open-source researchers, the same numbers point the other way. First-person footage from glasses is becoming a large and growing share of what gets posted from raids, protests, accidents and conflict zones. It’s shot from a known height with a fixed field of view, and it usually runs longer than phone clips because nobody has to hold anything up. That’s good for geolocation and chronolocation. Agents wore glasses to film protesters, and nothing stops protesters from filming agents the same way. Either way, the footage ends up online.
The market may be sending its own signal. The number two vendor by shipments, Even Realities, sells glasses with no camera at all, 102,000 units in the half. Meta’s newest frame, the Ray-Ban Meta Audio, ships on October 13 without a camera. Even Omdia’s research director, Jason Low, says vendors “must define exactly what always-on cameras capture.” When the market leader offers a camera-free version, it’s telling you which part of the product causes the trouble.
For now, the pairing still works as a detection point. The glasses announce themselves over Bluetooth to reach a phone, and a scanner can pick that up. Once the model runs on the frame and the frame no longer needs the phone, even that signal goes away.