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

Somewhere off the coast of New England, a harbor seal hauls itself onto a rocky ledge, blinks at a researcher's camera, and — without knowing it — teaches us the most important thing most people get wrong about facial recognition.

The seal has no name. No ID number. No entry in any database. And yet, the software knows it's been here before.

Not because it "recognized" the seal. Because it measured it.

The Seal Study That Changes Everything

A research team published a study in Ecology and Evolution (Wiley Online Library) describing a system called SealNet. The goal was to figure out which harbor seals kept returning to the same haul-out sites (the rocky beaches where seals rest in groups) and which ones were spotted together, hinting at social bonds.

The problem? Seals don't wear tags well. Coat patterns shift during molting season. Counting heads is hard when 200 slippery animals are piled on top of each other. Manual photo comparison by humans? Exhausting, error-prone, and genuinely difficult when the same seal looks different in October than it did in May.

So the team fed 1,752 photographs of 408 individual harbor seals into a facial comparison system. And it worked — hitting 88% accuracy at what researchers call "rank-1" identification, meaning the correct seal was the system's very first guess 88% of the time. Expand to the top five guesses (rank-5), and accuracy climbed to 96%.

But here's the part worth sitting with: the system never once "knew" who any of these seals were. It had no names. No prior database to check against. It simply looked at a new photo and answered one question — have we measured a face like this before?

What "Measuring a Face" Actually Means

When a facial comparison system looks at a photo, it doesn't see what you see. It doesn't notice the whiskers or the big brown eyes. What it does is run the image through a neural network (a type of software loosely modeled on how brains process information) that converts the face into a list of numbers. In some common systems, that list is 128 numbers long. Think of it as coordinates — like GPS points, but for facial geometry. This article is part of a series — start with Why Fake Faces Look More Real Than Genuine Photos.

That list of numbers is called a feature vector (basically, a face translated into math). Every photo of every face becomes its own feature vector. Two photos of the same individual should produce feature vectors that are very close together mathematically. Two photos of different individuals should produce vectors that are farther apart.

The system then calculates the distance between two feature vectors — a calculation called Euclidean distance (think of it as measuring a straight line between two points on a map, except the "map" has 128 dimensions instead of two). If that distance falls below a preset threshold — say, 0.5 — the system flags the faces as a probable match. Above the threshold? Probably different individuals.

That threshold is the secret lever nobody talks about. It's set by humans during development, and changing it changes everything. Lower the threshold and the system becomes pickier, missing some real matches but generating fewer false alarms. Raise it and the system is more generous — catches more true matches, but also groups some wrong faces together. There is no perfect setting. Every threshold is a tradeoff.

"Facial recognition technology for ecological studies emerged precisely because manual monitoring of coastal species is time-consuming and invasive — researchers needed to identify repeat individuals without tracking tags or names." — Birenbaum et al., PubMed Central / NIH

Why the Results Are Candidates, Not Answers

Remember that rank-1 versus rank-5 accuracy distinction? It's not a quirk of the seal study. It's how almost every facial comparison system actually reports its results.

The system doesn't give you a single definitive answer. It gives you a ranked list. "Here are the five faces from your dataset that are most mathematically similar to this new face, starting with the closest." You still have to verify. The algorithm sorted the pile — it didn't solve the case.

This is a genuinely important detail that gets lost in news coverage. When a headline says "AI identified the suspect," what usually happened is that AI ranked a suspect as the closest mathematical match in a set of images a human had already assembled. Then a human confirmed it. The system didn't point a finger. It handed someone a much shorter list to check.

The SealNet research makes this visible in a way that's almost charming. A seal returned to the same beach two years in a row. The algorithm said "this face vector is 0.18 units from a face vector we measured last October — that's below our threshold, probable match." A researcher looked at both photos and agreed. The seal got counted as a repeat visitor. Two steps. Machine first, human second.

The Big Misconception (And Why It's So Easy to Make)

Most people picture facial recognition like this: you walk past a camera, your face gets scanned against millions of records, and somewhere a computer goes ding — found you. Like a supercharged Google reverse image search for human beings. Previously in this series: The Most Real Face Youll See Today Was Never Born.

It's a reasonable assumption. That's basically how it's shown in every thriller, every news segment, every breathless tech announcement. And honestly? That version of the technology does exist, in some specific applications, deployed by some specific organizations with access to very large databases.

But it's not what facial comparison software is doing most of the time — and it's definitely not what it was doing with those seals.

The confusion comes from one word: recognition. That word implies the system knows who someone is. It suggests a name at the end of the process. And that makes the whole thing feel like magic — or surveillance — depending on your mood.

What's actually happening in most cases is much more like this: imagine you have 200 security photos from a parking garage over six months. No names attached to any of them. You want to know which faces appear more than once — maybe the same person has been there three different times. Doing that yourself, staring at 200 photos, is genuinely hard. People get tired. Angles change. Hats. Different lighting. You'll miss things.

Facial comparison hands that problem to an algorithm that doesn't get tired. It measures every face against every other face in your existing set, finds the mathematically close pairs, and groups them. It does not reach outside your collection. It doesn't look anything up. It has no idea what anyone's name is. It just spotted the repeat.

That's what the SealNet researchers needed. Not "which seal is this?" — they didn't have a registry of named seals to check against. They needed: "Is this the same individual we photographed last season?" Comparison, not identification. Pattern-matching within the dataset they already had.

According to the Colgate University team behind SealNet, the entire approach was built around non-invasive photography precisely because it avoided the need for tagging, tracking, or any external identification system. The faces were the only record. The math did the rest.

The Part That Should Actually Reassure You

Here's something the SealNet numbers quietly prove: you don't need a massive database for facial comparison to be useful. The system worked across 408 seals with as few as five photos per individual. Not millions of training images. Not a government archive. Just a modest collection of field photographs, carefully gathered over two years. Up next: The Most Real Face Youll See Today Was Never Born.

Why does that matter to you? Because it clarifies when face technology is genuinely powerful versus when someone is overstating what it can do.

Face comparison is strongest when it's working within a defined set — comparing images you already have, looking for patterns inside that collection. It's doing careful, tireless, mathematical work that humans struggle to do at scale. That's real and useful.

What it's not doing, in most everyday deployments, is plucking your identity out of thin air from a single image. Claiming a system can do that — without a database to compare against, without multiple reference images, without human review of the ranked results — is claiming something well beyond what the math actually delivers.

Knowing that distinction makes you a smarter reader of every facial recognition headline you'll ever see. Is the technology comparing images within an existing set? Or is it claiming to identify someone from scratch? Those are two very different claims. One is a measurement process with documented accuracy rates. The other requires a whole separate system, a separate database, and a whole lot more skepticism.

At CaraComp, that distinction between comparison and identification is where almost every real conversation about face technology should start. It's the baseline. Without it, the headlines don't make sense — and neither do the fears.

So the next time you see a story about facial recognition doing something impressive, ask yourself: was the system comparing faces it already had? Or was it claiming to pull a name out of nowhere? A harbor seal can't tell you its own name. But the math can tell you it's been to this beach before.

Those are not the same thing. And now you know why.