Meta Said Its Facial Recognition Feature "Does Not Exist." Then Its CTO Explained It in Public.

On June 4, 2026, WIRED revealed that Meta had shipped dormant facial recognition code — internally called "NameTag" — to more than 50 million smartphones via the Meta AI app, which pairs with its Ray-Ban and Oakley smart glasses . Meta's communications VP Andy Stone responded that the feature "does not exist" and was merely "exploratory" code that was never activated . Industry analyst Matt Navarra summarized Meta's posture bluntly: "Meta says that its face-recognition system 'doesn't exist'" .

Just over five weeks later, on July 8, 2026, Meta CTO Andrew "Boz" Bosworth gave a public interview in which he described the same NameTag system in detail. He called it a "local, encrypted feature used to identify people you already know" — not a system that pulls data from a central database — and confirmed it would store biometric data on-device . The Verge reported that Bosworth "defended an unreleased smart glasses feature called NameTag" and explained how it would work .

This is not an accident. It is a deliberate two-track communications strategy.

The Denial: "This Feature Does Not Exist"

When WIRED published its investigation on June 4, 2026, it revealed that Meta had quietly incorporated NameTag code through various updates over months . The code was designed to convert faces captured by the glasses' camera into unique biometric "faceprints" and match them against an on-device database . It was never activated, but it was real — and it was on 50 million devices.

Meta's response was immediate and emphatic. Andy Stone argued that the code was "not enabled" and that Meta had "no current plans" to activate it . Stone called the reporting "intellectually dishonest" and "advocacy-driven click bait" . On June 5, just one day after the report, Meta quietly removed nearly all NameTag code from the app .

The core of Meta's defense was a semantic argument: a dormant, unactivated feature buried in code did not qualify as a real "feature" .

The Contradiction: Bosworth Publicly Describes NameTag

On July 8, 2026, Andrew Bosworth sat for an interview and described NameTag as a local, encrypted feature that would recognize people you have already met . He framed it as a solution to what he called the "cocktail party problem" — forgetting the name of someone you have been introduced to . He emphasized that the biometric data would stay on the device and that the system would not pull from a central database .

Bosworth did not deny the feature existed. Instead, he defended it.

The Timeline

Date	Event
Feb 2026	NYT reports Meta is considering adding facial recognition to smart glasses
June 4, 2026	WIRED reveals NameTag code already shipped to 50M+ devices
June 5, 2026	Meta quietly removes nearly all NameTag code from the app
June 5+	Andy Stone argues the feature "does not exist" because it was never activated
July 8–10, 2026	CTO Andrew Bosworth publicly describes and defends NameTag's design and purpose

How to Explain the Contradiction

Three factors explain the gap between what Meta's PR team said and what its CTO described five weeks later.

1. Legal and regulatory risk management. Meta faces extraordinary legal exposure on biometric privacy. In 2021, it shut down its Facebook facial recognition system and deleted more than one billion faceprints after years of lawsuits . In 2024, it paid $1.4 billion to Texas alone to settle biometric privacy claims . That came on top of a $650 million settlement in Illinois in 2021 . The Electronic Frontier Foundation calculated that Meta has paid nearly $7 billion in combined settlements for facial recognition violations . When Meta says "this does not exist," it insulates itself from immediate legal liability — even if the code is already on millions of phones.

2. Product development continues internally. The code shipped to 50 million phones was not a prototype. Security researchers who reviewed it found that NameTag used three AI models working in sequence: one detected faces, a second converted each face into a 2,048-number "faceprint," and a third matched it against a stored database . The code had two deployment variants . Bosworth's July comments confirm that Meta fully intends to build and ship this feature. The denial was about the feature's activation status, not its existence .

3. Semantic gamesmanship. As multiple analysts observed, Meta's communications team leaned on a narrow definition: a feature "does not exist" if it is not yet user-facing . This allows the company to plausibly deny and simultaneously advance the same feature in the same quarter. WIRED itself summarized the episode on July 16: "Does a software feature exist if its code has been deployed to the devices of millions of people but they can't use it yet? Not if you work at Meta" .

What This Reveals About Meta's Pattern

The NameTag episode fits a consistent corporate pattern:

Repeated biometric privacy settlements. Meta has paid nearly $7 billion in total for facial recognition violations . The EFF noted: "Meta should already know the privacy risks of face recognition technology, after abandoning related technology and paying nearly $7 billion in settlements" .

Ship first, disclose later — or not at all. The NameTag code was pushed to 50 million phones over months of updates before anyone outside Meta knew it existed . Researchers at the EFF discovered it only by examining the code . Internal Meta documents published by the New York Times in February 2026 showed the company had planned to roll out the feature during a "dynamic political environment," when Meta believed its biggest critics would be preoccupied .

Public denial, quiet retreat, then reintroduction. On June 5, Meta stripped the code after exposure . But Bosworth's July comments — describing the feature in detail and defending its privacy design — signal that the company has not abandoned the plan. It is waiting for the controversy to subside.

Minimizing readiness while deploying infrastructure. The pattern mirrors earlier Meta AI rollouts: the company ships the supporting code broadly, denies the feature is imminent when caught, then flips the switch later. NameTag's code was engineered, tested, and distributed. Only the activation flag was missing .