CCTV Facial Recognition: Why a 98% Match Proves Nothing

A ninety-eight percent facial recognition match sounds like certainty. But run that same ninety-eight percent threshold against a database of ten million faces, and you can still get hundreds of thousands of possible candidates. The number isn't wrong. What people believe the number means — that's the problem.

If you've ever walked into a supermarket, a train

If you've ever walked into a supermarket, a train station, or a stadium, a camera has probably captured your face. And if you've felt a knot in your stomach about that, you're in good company. According to polling in Greater London, two-thirds of people worry that a facial recognition error could get someone into trouble unfairly. Another sixty-two percent are uneasy about how their face images get stored. That worry is reasonable. But there's a gap between what this technology actually does and what most of us imagine it does. So how does a camera go from a blurry frame of video to a number that says ninety-eight percent?

Start with what a security camera actually is. A closed circuit camera records that a person was there. That's it. It's a log of an event. Facial comparison is a completely separate step. It answers a different question — is this person the same as that person? A camera is like a security logbook. Facial comparison is more like fingerprint matching. The video proves something happened. The comparison tests whether two images show the same human being. You need both. They solve different problems.

So what happens inside the comparison? The software doesn't look at your face the way you look at a friend across a room. It builds what's called a biometric template — basically a mathematical map. It measures the space between your eyes. The distance from your chin to your forehead. The angle of your jaw. The lengths of different sections of your face. Your face becomes a list of numbers. Nothing more.

Then the system compares your list of numbers to a

Then the system compares your list of numbers to a stored list of numbers. Most tools use something called Euclidean distance — a fast way of measuring how far apart two sets of measurements are. The smaller the distance, the better the match. That's genuinely useful. It's also fast and cheap to compute. But notice what just happened. The output isn't a name. It's a gap between two sets of geometry.

Now the part that changes everything. Someone has to decide, before the search runs, how small that gap has to be to count as a match. That's the threshold. Set it tight, and you'll miss real matches. Set it loose, and you'll drown in false ones. A human being picks that dial. Not the machine. For an investigator, that means the threshold is a defensible choice they have to explain. For the rest of us, it means the confidence score you see in a news story was shaped by a decision someone made in advance.

Why do so many of us hear ninety-eight percent and think "case closed"? Because percentages feel like school grades. Ninety-eight out of a hundred sounds like near-perfect. But this number isn't a grade. It's a distance between two mathematical maps. And those maps shift. Researchers point out that the hardest problem in facial biometrics is handling changes in head position, expression, and lighting. The same person, photographed twice, produces two slightly different templates. Two different people can produce templates that sit surprisingly close together.