What Facial Recognition Steals from Us
Facial recognition technology—built on photos we shared online without anticipating this use—can collapse the anonymity we naturally have in public spaces. The core argument is that **obscurity** (personal information being hard to find) is a fundamental human protection, and facial recognition is the most powerful tool ever invented for destroying it. ---
Key Concepts
| Concept | Definition |
|---|---|
| Facial recognition pipeline | digital cameras (eyes) → deep learning algorithms (processing) → a database of known faces (memory); all three components must be present to identify someone |
| Obscurity | the idea that personal information is safer when it is difficult to obtain, understand, or connect — philosopher Evan Selinger's framework for what facial recognition threatens |
| Context collapse | the bleeding together of separate social domains (work, friendship, intimacy) when too much personal information becomes universally accessible |
| Transaction costs of information | the natural human limits on perceiving, storing, and sharing information that technologies like facial recognition reduce to near zero |
| The face-to-name reversal | traditional search (name → face) vs. facial recognition search (face → name) — the latter enables identification of strangers in the physical world |
Notes
How Facial Recognition Works
- Cameras capture an image; algorithms locate faces within it
- Facial landmarks are mapped and normalized (correcting for rotation, etc.)
- 100+ measurements are taken — exact features measured are determined by the deep learning algorithm, not explicitly programmed
- Training uses **triplets**: anchor photo + matching photo + non-matching photo; algorithm minimizes distance between matches, maximizes distance between non-matches
- Refined over millions of examples but performs unequally across demographics and image conditions
- The system cannot identify anyone until they are in a **database of known faces** — this is separate from training data
Where the Training Data Came From
- Early 2000s: digital cameras + social internet (Facebook, Flickr, YouTube) created a massive, labeled photo repository
- Facebook users tagged ~100 million photos per day, voluntarily labeling faces with names
- Professional photography went online via news sites and stock libraries
- Google image search aggregated it all — researchers then used this corpus to train models
Who Holds the Databases
- **Individuals**: e.g., Face ID — one-person database, user-consented, stored on device
- **Companies**: Facebook, Google maintain user databases
- **Governments**: largest databases of name-face pairs, collected for other purposes (IDs, passports), now repurposed for facial recognition without consent
- **Private sector**: retailers, banks, stadiums build or buy watch lists (shoplifters, VIPs, persons of interest)
- **Scrapers**: bots can harvest names and faces from LinkedIn, Twitter, Facebook — against terms of service but technically feasible
The Yandex / Russian Social Media Demonstration
- Yandex reverse image search finds the **same face**, not visually similar images
- Journalist Eric Toller (Bellingcat) demonstrated using VK (Russian social network): uploaded a low-resolution (~200×100px) frame of soccer hooligans and found individual profiles for nearly every person
- Matched people across photos, identified clothing, found family/social connections through tagged photos
- The video didn't require clear or high-resolution images to work
Real-World Uses — Both Investigative and Harmful
- **Investigative**: Bellingcat used facial recognition to identify individuals linked to the MH17 attack in eastern Ukraine
- **Accountability**: identifying police officers accused of brutality; identifying Putin protest attendees
- **Harmful/harmful potential**: doxxing, targeting sex workers and anonymous performers, stalking strangers
The Philosophical Stakes (Evan Selinger)
- Privacy in public is not a contradiction — obscurity functions as a *de facto* privacy layer
- Human cognitive limits naturally constrain surveillance; technology eliminates those limits
- Losing obscurity threatens:
- **Individuality**: experimentation in life requires some protection from total observation
- **Intimacy**: sharing different parts of yourself with different people requires contextual separation
- **Authenticity**: behaving differently across social contexts is not dishonesty — it is richness of human life
- The underlying problem: photos shared for one purpose (social connection, official IDs) have been repurposed into an identification infrastructure — a **"massive bait and switch"**
Actionable Takeaways
- Be deliberate about which photos you make publicly linkable to your real name — once indexed, they can become identification vectors
- Understand that obscurity (not just "private settings") is a meaningful layer of protection worth considering when sharing images
- When evaluating facial recognition use cases, ask whose power it expands and whether the subjects consented to being in the relevant database
- Support or follow municipal/legislative efforts to restrict government use of facial recognition, where that aligns with your values
Quotes Worth Keeping
Facial recognition is probably the most obscurity-obliterating technology ever invented.
Are we allowed to want to share and connect with other people online and still be able to expect not to be recognized when we're offline in our regular lives? I would say absolutely.
Leading a rich life requires us to be able to express ourselves in these diverse ways.
The photos we took to share with friends or document history or simply get a government ID have been used to build and operate a technology that strips away the protections that obscurity has always provided us. It's nothing less than a massive bait and switch.