Testing the Limits of Facial Recognition

Our Take

Credit to the creator behind Business Reform for actually testing a claim most of us just assume: that facial recognition is some infallible, all-seeing system. Turns out a few hours in a makeup chair with prosthetic cheeks, a new nose, and a heavier brow was enough to take PimEyes from 151 confident matches down to zero. That's not a hack or a software exploit — that's a person with glue and latex quietly defeating a tool that companies and, increasingly, police departments treat as gospel.

This matters because facial recognition doesn't operate in a vacuum. It's bolted onto the same surveillance infrastructure we track here — license plate readers, Flock Safety cameras, data-sharing pipelines between private vendors and law enforcement. The pitch is always the same: these systems are accurate, objective, and only catch the bad guys. One guy with Hollywood-grade prosthetics just demonstrated how shaky that pitch really is, while also showing that the technology is plenty good enough to track ordinary people who aren't wearing a disguise. Error-prone surveillance is bad whether it's failing to catch someone or falsely flagging someone else — both outcomes erode the argument that mass biometric tracking makes anyone safer.

If you want to see how this kind of automated tracking shows up in your own neighborhood, check out our camera map, and if you'd rather push back than just watch, our take-action page has concrete steps for challenging ALPR and facial recognition deployments where you live.