Your Face Is Just 128 Numbers — And a Seal Just Proved It

A team of researchers pointed facial recognition software at four hundred and eight harbor seals. And it worked — eighty-eight percent of the time, it could tell one seal from another. But here's the part that should change how you think about this technology — the software had no idea who any of those seals were.

If a headline about facial recognition has ever

If a headline about facial recognition has ever made your stomach drop, I get it. The idea of a machine scanning your face and instantly knowing your name — that's genuinely unsettling. But the seal study reveals something most people never learn. The technology you're afraid of, and the technology that's actually in most of these systems, are two very different things. One of them knows who you are. The other one just measures. So how does a computer recognize a face without ever knowing whose face it is?

Let's start with what your face actually becomes inside one of these systems. It's not a photo. The software looks at your face and converts it into numbers — in some systems, about a hundred and twenty-eight of them. That string of numbers is called a face embedding. It's basically a mathematical fingerprint — a coded summary of what makes your face yours. The original picture? The system doesn't even need to keep it. It stores the numbers instead.

Now, two faces. Two strings of numbers. The computer does something almost boringly simple — it measures the distance between them. If those two number-sets are close together, it calls them the same face. If they're far apart, it calls them different. And the line between close and far? That's a threshold. A cutoff point somebody picked during development. Move that line, and the same two faces can flip from match to no-match. The machine isn't certain. It's measuring.

Picture sorting six months of security footage from

Picture sorting six months of security footage from your building. You've got no names, no IDs — just faces that keep showing up. Now imagine a colleague who's freakishly good at noticing when the same person appears in two different clips, even if the lighting changed or they turned away. But that colleague can't tell you their name. They can only say, "I've seen this one before." That's what facial comparison really does. With the seals, that's exactly what happened — the software noticed the same animals returning to the same beaches year after year, without ever assigning a single name.

Here's where a lot of fear comes from, and it's an honest mistake. The word "recognition" sounds like the system knows you. News stories and product ads blur two very different things together. Comparison asks — are these two faces I already have probably the same person? Recognition asks — who is this stranger, out of millions? Those are completely different jobs. The seal software only ever did the first one.

And about that eighty-eight percent — context is everything. That accuracy came from a closed pool of four hundred and eight known seals. Testing against four hundred faces is a world away from testing against ten million. The system also didn't hand back one answer. It returned a ranked list of likely matches — best guess first. A human still has to check. For an investigator, that means the algorithm gives you candidates, not verdicts.