The Facial Recognition Accuracy Myth
Facial recognition vendors advertise accuracy rates above 99%. These numbers come from controlled lab environments with perfect lighting, front-facing subjects, and high-resolution images. Real-world deployments tell a different story.
Lab vs. reality
NIST’s Face Recognition Vendor Test (FRVT) evaluates algorithms under controlled conditions. Top vendors achieve 99.7%+ accuracy on mugshot databases. But the conditions that produce these results almost never exist in real-world surveillance:
- Camera angle: Surveillance cameras are mounted above, capturing faces at angles the algorithms weren’t optimized for
- Lighting: Outdoor cameras deal with shadows, backlighting, glare, and darkness
- Resolution: Most CCTV operates at resolutions far below what vendors test against
- Movement: People walk, turn, and look down. Lab tests use static images
- Distance: Surveillance cameras are often 10-50 meters from subjects
The false positive problem
Even if an algorithm is 99.9% accurate, running it against a database of millions produces thousands of false matches daily. In a city of 1 million people scanned once per day:
- 99.9% accuracy = 1,000 false matches per day
- 99.99% accuracy = 100 false matches per day
Each false match is a person wrongly flagged by police systems. Several documented cases of wrongful arrest based on facial recognition have been reported in Detroit, New Orleans, and New Jersey.
Demographic bias
Multiple studies (NIST, MIT Media Lab, Georgetown) have found that facial recognition error rates are dramatically higher for:
- Black women: up to 35% error rates in some systems
- Darker skin tones: 10-100x higher false positive rates
- Women: higher error rates than men across most algorithms
- Young people and elderly: higher error rates than middle-aged adults
This isn’t a minor calibration issue. It means facial recognition systems are systematically less accurate for the populations most likely to be surveilled.
Where it’s deployed
The EFF Atlas of Surveillance documents 402 facial recognition deployments across the US. Sentinel maps these alongside other surveillance technologies so you can see which areas face the highest combined surveillance density.