A researcher decompiled Meta's glasses app and found a face-recognition engine fully wired up
Inside the Stella companion app for Meta's smart glasses, a security researcher found three AI models, a local face database, and a notification channel literally named 'nametags_recognition'. The machinery is built. Whether it's switched on is the whole fight.
What the decompile actually found

The report is not a leak or a rumor — it's a static analysis. A researcher pulled apart Android build 273.0.0.21 of Stella (com.facebook.stella), the companion app for Meta's smart glasses, and traced an end-to-end face-recognition pipeline that runs entirely on the phone:
- Three bundled models (~100 MB): SCRFD for face detection, KPSAligner for alignment, SFace to generate a 2048-dimension biometric embedding.
- A local database at
person_profiles/objects.db— SQLite with a cosine-similarity vector index. - A notification channel named
nametags_recognitionthat fires text like "Person recognized."
The detail that turns this from a tech demo into a story: unmatched faces are written to a NameTagsPending/ staging folder as cropped images plus fingerprints. In plain terms, a face you haven't named yet is stored as a biometric record, ready to be labeled retroactively once an identity shows up.
The researcher is careful, and that caution matters: the pipeline "is present, it is wired together," but the user-facing card doesn't appear on standard accounts and no server-side data flow was observed on test devices. Built ≠ shipped. But built is the part you can't un-ring.
The thread that remembers 2012
Simon Willison supplied the historical mirror: back in 2012 he ran a conference startup with a database of attendee photos, and the obvious Google Glass app was "your glasses help you spot the people you're planning to meet." Google's developer terms flatly forbade it. The reversal in thirteen years isn't the technology — it's that the technology now ships in the frame.
The legal subplot ran hot. One commenter predicted Meta was "determined to make Chicago lawyers rich," pointing at Illinois's Biometric Information Privacy Act — only for another reader to note BIPA targets the collection of biometric data, not end users of a recognition service, and to ask whether Meta could simply geo-gate Illinois.
A recurring demand flipped the product on its head:
"I'd like to wear an EXACT OPPOSITE of this… if someone is using Facebook's AI-powered glasses in my vicinity, I want to get a notification so that I can avoid those persons" — a reader who wants the inverse alert
Someone pointed out such an app already exists — "Nearby Glasses" on the Play Store. And one voice cut against the pile-on with the uncomfortable nuance:
"Those glasses don't record by themselves. Somebody pressed the button and deliberately recorded it." — a commenter on where the agency sits
The honest read
The defensible version of this story is the precise one: nobody has shown Meta is running face recognition on the public today. What the decompile shows is that the cost of turning it on has dropped to a feature flag. The privacy fight over wearable cameras used to be about whether the capability could exist. That fight is over. It exists, it's local, it's 100 MB, and it's already in the app on your friend's face.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Anthropic says its engineers now merge 8x more code a day. The thread did the math
Anthropic published an essay arguing AI is starting to build AI — and dropped a number: in Q2 2026 its engineers merged 8x as much code per day as in 2024. On Hacker News, the people who review that code were unimpressed.