Meta’s AI Glasses Companion App Hides A Complete Face Recognition System

Reverse engineering of Meta’s AI glasses companion app, internally called Stella, has revealed a complete face recognition system that detects faces, generates 2048-dimensional biometric fingerprints, and stores unknown faces in a persistent SQLite vector database. The hidden infrastructure includes three machine learning models for detection, alignment, and embedding generation, with a notification system that announces recognized individuals—all without user-facing controls or disclosure. This discovery raises concerns about consent and transparency as Meta ships dormant surveillance capabilities in consumer smart glasses while regulators scrutinize biometric data collection.

Meta’s AI glasses companion app carries a dirty little secret: a complete, functional face recognition system that can identify people and store biometric data https://gdpr-info.eu/art-4-gdpr/ on your device. The discovery, revealed through reverse engineering https://www.buchodi.com/meta-glasses-facial-recognition/ of the Meta AI app https://play.google.com/store/apps/details?id=com.facebook.stella internally called Stella , shows three specialized machine learning models working together to detect faces, align them, and generate 2048-dimensional biometric fingerprints —all while users remain completely unaware. The hidden infrastructure goes far beyond simple camera autofocus detection. This isn’t your typical camera https://www.gadgetreview.com/best-cameras-you-can-buy-for-every-budget autofocus detection. The hidden pipeline includes: - An SQLite vector database designed for similarity matching Persistent storage for unknown faces- A notification system that announces “Person recognized” when it finds a match During controlled testing by researchers at Buchodi , the system successfully identified a portrait of Michel Foucault https://news.ycombinator.com/item?id=48403588 after pre-loading his biometric data, proving the entire recognition chain works end-to-end. Meta ships three sophisticated models that rival dedicated surveillance systems. The technical sophistication rivals dedicated surveillance app https://www.gadgetreview.com/us-operatives-built-a-surveillance-app-to-target-alberta-separatists systems. Meta ships three ExecuTorch models : for face detection SCRFD KPSAligner for positioning- A scaled-up SFace variant for embedding generation Unknown faces get stored as cropped images plus binary embedding files in a private directory that survives device reboots. Think of it as building a “faces pending identification” database without asking permission first. User-facing controls remain conspicuously absent despite complete backend infrastructure. Yet the user interface remains carefully hidden. The companion app contains hardcoded strings for a “Connections” feature that promises to “remember the people you met,” but this never appears for regular users. Recognition notifications deep-link to profile screens that don’t exist in the current build. Meta has essentially shipped the surveillance infrastructure while keeping the front door locked. This discovery lands differently than Meta’s previous face recognition controversies https://www.eff.org/deeplinks/2026/02/seven-billion-reasons-facebook-abandon-its-face-recognition-plans . While the system operates on your device, Meta’s own transparency disclosures https://transparency.meta.com/policies/other-policies/pre-disclosure/meta-ai-glasses/ indicate that AI glasses https://www.meta.com/help/ai-glasses/268269592726432/ data may still be stored on both local devices and remote servers. With regulators already scrutinizing biometric data collection, shipping dormant face recognition capabilities in consumer smart glasses https://www.gadgetreview.com/apple-cooks-up-custom-silicon-smart-glasses-and-ai-chips-signal-techs-next-evolution raises uncomfortable questions about consent, transparency, and the gap between what companies build versus what they disclose. Your AI glasses just became a lot more interesting—and concerning.