Meta AI Glasses Can ID Strangers: 70 Groups Demand a Stop

If someone wearing Meta's Ray-Ban smart glasses walks past you on the street, AI facial recognition can surface your name, your social profiles, and your city — before you've said a word. That AI automation capability exists in hardware available today, and in April 2026, more than 70 civil rights organizations drew a hard line against it.

The American Civil Liberties Union (ACLU), the Electronic Privacy Information Center (EPIC), and Fight for the Future are among the signatories of a joint letter demanding Meta immediately disable the facial recognition (the ability to identify a person from a photo of their face alone, without any other identifying information) feature embedded in its consumer smart glasses. The coalition is one of the largest ever assembled against a single AI product feature.

Meta Ray-Ban AI Glasses: How Facial Recognition Works

Meta's Ray-Ban smart glasses — co-designed with Oakley — include onboard cameras that capture faces and cross-reference them against publicly indexed social media data in real time. The identification happens passively. The wearer takes no deliberate action. The glasses look like ordinary sunglasses with no visible indicator that the person nearby is being scanned.

Here is how the identification pipeline works:

• Continuous capture: Onboard cameras film the environment while the glasses are worn.
• Facial embedding: An embedding model (a tool that converts a face image into a string of numbers that AI can compare against a database of known faces) processes each detected face in real time.
• Identity retrieval: The system queries publicly indexed social media profiles, returning a name and profile link within seconds of detection.
• No opt-out: There is currently no public mechanism for individuals to remove their likeness from the matching dataset.

Google and YouTube are also named in the coalition letter as platforms whose publicly indexed data feeds the identification system — extending responsibility beyond Meta to the broader social media ecosystem.

70+ Civil Rights Groups and the AI Privacy Risks They Are Fighting

The breadth of this coalition signals a shift: this is no longer a niche privacy debate among technologists. Organizations representing domestic violence survivors, immigration rights, LGBTQ+ advocacy, and civil liberties all found the same reason to oppose a single consumer product feature. That alignment is rare.

The letter names three groups at acutely elevated risk:

• Abuse survivors — People who have relocated to escape an abusive partner. A facial scan in any public space can instantly eliminate the geographic safety barrier a protective order or relocation created.
• Undocumented immigrants — For whom real-time identity confirmation in public could trigger enforcement contact, regardless of any pending legal proceedings.
• LGBTQ+ individuals not publicly out — Being linked to a social media profile that reveals identity information they have not chosen to disclose exposes them to violence, family rejection, or workplace discrimination.

The framing — "the AI smart glasses feature would endanger abuse victims, immigrants, and LGBTQ+ people" — deliberately leads with the most vulnerable populations rather than abstract privacy principles. The letter also puts Meta on a collision course with European regulators: under the GDPR (the EU's General Data Protection Regulation, which prohibits biometric data processing — including facial recognition — without explicit, freely given consent), passive identification systems like this are presumptively unlawful across the EU's 27 member states.

Anthropic and OpenAI Just Picked Opposite Sides on AI Regulation

The Meta glasses story is one front in a broader regulatory conflict playing out simultaneously in Springfield, Illinois, where a proposed state AI liability law has exposed a direct philosophical split between the two most prominent AI labs in the world.

The Illinois law would shield AI companies from tort liability (the legal framework that holds manufacturers financially responsible when their products cause injury, death, or large-scale financial harm) in cases involving catastrophic outcomes from AI systems.

• OpenAI supports the liability cap — arguing that legal exposure at catastrophic-harm scale paralyzes AI deployment and pushes development to less regulated jurisdictions.
• Anthropic opposes the same shield — taking the position that accountability is not a burden on safety but its primary mechanism. Remove consequences for harm, and you remove the structural incentive to build safer systems.

This is not a procedural disagreement. It reflects incompatible theories about how AI safety actually works in practice. OpenAI's approach mirrors internet platform immunity — companies cannot be held liable for downstream misuse. Anthropic's approach mirrors pharmaceutical law — a manufacturer carries liability for foreseeable harm even when the product works as designed. Illinois is being watched by California, Texas, and New York, all of which have active AI regulation processes. The framework that wins here will spread.

A humanoid robot just landed on AliExpress for $4,370

While legislators debate and civil rights groups sign letters, hardware is moving faster than either. Unitree's R1 humanoid robot — a full-size, bipedal machine (a two-legged robot engineered to walk and perform physical tasks the way a human does) — is now available internationally on AliExpress at an entry price of $4,370.

Context that makes this figure striking: humanoid robots cost between $150,000 and $500,000 as recently as 2023–2024, and were purchased exclusively through enterprise B2B procurement contracts — not consumer retail channels. The Unitree R1 price represents a decline of more than 97% in roughly 18 months.

The economic math has now changed for small operators:

• A full-time U.S. worker at minimum wage costs approximately $30,000–$50,000 per year in total employment cost including benefits.
• A $4,370 robot amortized (spread across its useful operational lifespan) over two years costs approximately $2,185 per year for routine physical tasks.
• The payback period for light warehouse and delivery automation is now measurable in weeks, not years.

The distribution channel is itself a signal. AliExpress — a consumer-facing marketplace — is not where enterprises buy industrial equipment. Unitree is positioning humanoid automation as a small-business purchasing decision, not a multi-year capital expenditure requiring executive sign-off. If you operate in logistics, light manufacturing, or last-mile delivery, start modeling cost-per-task comparisons now. The decision window is Q2 2026, not 2027.

Anthropic's Mythos: 15 companies in, everyone else watching

One more April 2026 development rounds out the picture: Anthropic has deployed its new Mythos AI model — but only to 15 companies in a tightly controlled initial release. The rest of the market has no access and, currently, limited visibility into what the model can do.

Mythos is generating simultaneous reactions from two communities:

• A capability wake-up call: Experts say Mythos marks a threshold at which developers who have treated security as an afterthought — deferring patches and skipping threat modeling because "nothing bad has happened yet" — will face real consequences. The model is capable enough that complacency is no longer a viable operating posture.
• A potential attack tool: Security researchers flag Mythos as a possible "hacker's superweapon" — capable of assisting with social engineering (manipulating people into revealing passwords or granting system access through deception rather than technical exploits), vulnerability discovery, and attack planning at speed and scale that existing defenses are not designed to handle.

The 15-company restriction is designed to control for unintended consequences before broader deployment. But it also creates a knowledge gap: the organizations with early access are developing an understanding of Mythos' capabilities that the rest of the market lacks. When the controlled phase ends — likely within weeks — the transition to wider availability will be rapid.

If your organization operates in cybersecurity, compliance, or any domain where AI-assisted attacks represent a material risk, apply for Anthropic's extended testing program now, before general availability forces a reactive scramble. You can track AI model releases and access programs on the Guides section as new information becomes public.

Related Content — Get Started | Guides | More News