LifePivo

Meta's Glasses Won't Train Its AI on Your Photos

Explained by Maya Chen — LifePivo fictional editorial host

A person in dark glasses, seen from behind in profile, cradles a small warm orange photograph of a shadowy figure close to their chest; from the side of the glasses a ribbon of pale blue light carries an identical picture away across a dim space to a cluster of tall black towers, where three small anonymous figures stand watching the streamed image
AI-generated editorial illustration. Not a documentary image.

Until You Ask Meta AI About One

Meta has spent a year defending its smart glasses with a single small feature: the light. Every pair has a capture LED on the front that, in the company's own words, "blinks to let people know you're capturing content," has "no off switch," and now shuts the camera off entirely if it detects the light has been blocked or destroyed. When a hacker conference, a comic convention and 1,240 New York courthouses each banned the glasses this year, Meta pointed at the light. When a July prototype was reported that would photograph a wearer's surroundings every few seconds, Meta pointed at the light.

The light answers one question well: is the person opposite you recording right now? It answers a different question not at all — where does that recording go, and who else gets to see it. That second question does not belong to the bystander. It belongs to the person wearing the glasses, and most of them have never been shown the answer.

The light is about the room, not about the cloud

We have written twice about the LED from the outside looking in — about how a working light still tells a stranger nothing across a daylit room, and about how every guide to spotting recording glasses comes with a footnote admitting the method can fail. Both are about a bystander trying to read a signal. This is the opposite vantage: you are the one wearing them, the light is doing exactly what it should, and the footage is behaving in ways the light was never designed to disclose.

Here Meta's framing is genuinely reassuring, and genuinely true as far as it goes. Photos and videos are "stored privately on your glasses," the company says; "you choose when to import them to your phone," and who can see them is "you, and only you — unless you choose to share them." A picture that sits in your gallery and goes no further is a picture Meta does not train on. On that, the company and its sharpest critics agree.

Two piles of footage, and only one of them stays yours

The distinction that matters is not between photo and video. It is between media you keep and media you hand to the assistant.

Ask the glasses to do nothing clever — snap a photo, record a clip, leave it in the gallery — and it stays in the first pile. Ask Meta AI to look at something, and it moves to the second. In 2024, when reporters put the training question directly to Meta, its policy communications manager Emil Vazquez answered plainly: "in locations where multimodal AI is available (currently US and Canada), images and videos shared with Meta AI may be used to improve it per our Privacy Policy." The same reporting confirmed the reassuring half: photos and videos you never submit to the assistant are not used for training.

So the boundary is a behaviour, not a setting. "Hey Meta, what am I looking at?" is not a neutral convenience. It is the sentence that moves an image from the pile Meta cannot touch to the pile Meta can train on. Nothing on the glasses marks that line as you cross it — the light blinks the same way for a photo you keep and a photo you are about to send to a data centre.

There is a further wrinkle in that boundary worth naming: the geography is Meta's own from 2024. The company said "US and Canada" then. Its multimodal features have spread since, and Meta has not published a country list we could open and verify, so a wearer outside North America should treat "it doesn't apply to me" as unconfirmed rather than safe.

The moment you say "Hey Meta," the picture leaves the glasses

The reason the assistant redraws the boundary is architectural, not a matter of policy. The Electronic Frontier Foundation, reading Meta's own terms, put it flatly: "You can't use any AI features locally on the glasses. So anytime you use AI features … the footage is fed to Meta." There is no on-device model doing the looking. The looking happens on Meta's servers, which means the image has to travel there first.

The default settings assume you want that. Media is "imported automatically by default into the Meta AI mobile app," the EFF notes — an app that is required to set the glasses up at all — and a "Cloud media" option, on unless you find and switch it off, sends photos and videos "to Meta's cloud for processing and temporary storage." None of this is hidden in the sense of being concealed; it is in the terms. It is hidden in the sense that a person putting on a pair of sunglasses is not reading a data-processing agreement, and the one visible indicator they have been taught to trust, the light, says nothing about any of it.

And the review at the far end is not always a machine. Some of the footage shared with the AI, the EFF says, "goes through human review." Engadget, reporting the same month, found the same clause: Meta's terms require the wearer to agree that "either humans or automated systems may review sensitive data." The person who says "Hey Meta, read me this letter" has, in the same breath, agreed that a contractor might read it too.

Your voice is kept by default, and the off switch is a delete key

Video and photos are the loud part of this story. The quiet part is audio, and it is arguably the more surprising.

As of an update pushed on 29 April 2025, TechCrunch reported, AI features were switched on by default and the glasses began storing the speech a wearer says after the "Hey Meta" wake word — kept, per Meta's privacy notice as reported, up to a year "to help improve Meta's products." There is no toggle that stops the storing. A wearer who does not want their voice used to train the assistant, TechCrunch found, "will have to manually delete each recording from the Ray-Ban Meta companion app," one at a time. The EFF's summary of the same arrangement is blunter: the audio "is also saved by default, and if you don't like that, tough luck, unless you go in and manually delete them every time."

An opt-out that requires a daily housekeeping chore is, for most people, not an opt-out. It is a default with a manual undo, and defaults are what almost everyone lives inside.

What a lawsuit and Meta's own answer agree on

Earlier this year the human-review clause stopped being abstract. On 4 March 2026, two Ray-Ban Meta owners, Gina Bartone and Mateo Canu, sued Meta Platforms and Luxottica of America in the Northern District of California (Case No. 3:2026cv01897). Their complaint alleges Meta failed to disclose that videos captured by the glasses are, as the filing puts it, "transmitted to servers and then to a Kenyan subcontractor for manual labeling." The specifics of what reviewers encountered come from a Svenska Dagbladet investigation relayed through the coverage, and they are unproven allegations; the point that survives verification is narrower and, in a way, more telling.

It survives because Meta does not deny the mechanism — only its adequacy. Asked about the reviewers, the company's response was not "that does not happen." It was that "when people share content with Meta AI … we sometimes use contractors to review this data to improve people's experience with the glasses, as stated in our privacy policy," adding that "this data is first filtered to protect people's privacy." To Engadget, Meta said that "when live AI is being used, we process that media according to the Meta AI Terms of Service and Privacy Policy." Strip the dispute over how well the filtering works, and both sides describe the same pipeline: use the assistant, and a human you will never meet may look at what you showed it.

Where this leaves the person wearing them

None of this makes the glasses a covert recorder, and it is worth being precise about that, because the loudest version of this story treats every pair as a spy. The gallery really is private. The light really does blink. A wearer who never invokes the assistant keeps their footage on the device, and Meta's account of that case holds up.

The exposure lives entirely in the feature that makes the glasses worth wearing. The whole pitch is that you can ask them what you are looking at — read this menu, translate that sign, identify this plant. For some wearers that pitch is not a novelty but the point of the device: the same assistant that reads a blind user their post and finds a dropped bag is the one that ships the image to Meta to do it. Every one of those requests is the border crossing. The practical boundary, the one worth keeping in mind, is not between wearing the glasses and not wearing them; it is between looking through them and talking to them. The first is a camera with an honest light. The second is a standing arrangement in which your images and your voice can train a model and be read by a stranger, on by default, switched off only by hand.

Meta keeps answering the privacy question by pointing at the light, and the light keeps being a real answer — to the one question it was built for. The question the wearer actually has to weigh is not lit up at all.

Related coverage

What Does the Meta Glasses Recording Light Actually Cover? Read Meta's Own Wording — the matching gap facing the other way: what the light promises the room, and what it does not mention

Sources and verification

What we could not confirm and left out: the graphic specifics attributed to reviewers (secondhand via Svenska Dagbladet); that the lawsuit's allegations are proven (they are disputed); and that the training arrangement now extends beyond the US and Canada, which is where Meta itself last placed it.