How Meta and Google held back their tech to recognize unknown people's faces due to privacy worries, opening the door for startups like Clearview AI and PimEyes
Engineers at the tech giants built tools years ago that could put a name to any face but, for once, Silicon Valley did not want to move fast and break things. X: @sociogeeks_ , @bronwynwilliams , and @kashhill . LinkedIn: Kashmir Hill X: @sociogeeks_ : The absurd hat-phone, a particularly uncool version of the future, contained a secret tool known only to a small group of employees. What it could do was remarkable. @kashhill https://www.nytimes.com/... Bronwyn Williams / @bronwynwilliams : We share a planet. “It didn't frighten me, though I knew it should. It was clear people who own a tool like this will inevitably have power over those who don't. But there was a certain thrill in seeing it work, like a magic trick successfully performed” https://www.nytimes.com/... @kashhill : In early 2017, a Facebook engineer sat in a conference room in Menlo Park with a smartphone standing on the brim of his baseball cap, attached w/rubber bands. Everyone was laughing at how absurd it looked. Then the phone started speaking their names aloud. https://www.nytimes.com/... LinkedIn: Kashmir Hill : In early 2017, a Facebook engineer sat in a conference room in Menlo Park with a smartphone standing on the brim of his baseball cap, attached with rubber bands. …
Context & Ripple Effects
Meta and Google reportedly developed the capacity to identify unfamiliar faces but chose not to release it, making privacy restraint—not technical feasibility—the key competitive distinction. That left room for companies built around web-scale face search.
Clearview had already claimed to have scraped billions of online images and been used by hundreds of law-enforcement agencies in its earlier expansion into police-facing face search. Its CEO later described a collection exceeding 10 billion images and tools intended to extend police identification capabilities as its image corpus and investigative features grew.
First-order effects
- Meta and Google avoid directly distributing a highly sensitive consumer identification capability, while Clearview AI and PimEyes can occupy demand for identifying people from images.
- Individuals whose photos are publicly available face a more fragmented privacy landscape: major platforms’ restraint does not prevent specialist services from offering similar search functions.
Second-order effects
- Specialist facial-search providers gain a clearer opening to compete on data collection and identification workflows rather than on the underlying recognition model alone.
- The contrast increases pressure on platforms and policymakers to define whether publicly accessible images can be repurposed for identity search—a central issue raised by Clearview's web-image scraping model.
Third-order effects
- If large consumer platforms continue to self-limit while smaller vendors proceed, sensitive AI capabilities may migrate from general-purpose platforms to narrower, less visible providers.
- The enduring fault line is likely to be the public-data permission boundary: technical capability can spread faster than a common standard for consent, access, and accountability.
The trend: This is one instance of sensitive AI functions shifting toward specialist vendors when dominant platforms judge the privacy and reputational costs of broad release too high.