CaraComp
CaraComp
Forensic-Grade AI Face Recognition for:
Get Started7-day refund guarantee**
Podcast

ID Scan: How a 93% Face Match Put the Wrong Man in Cuffs

ID Scan: How a 93% Face Match Put the Wrong Man in Cuffs

Quick answer

What does a 93% face match score actually mean?

A 93% face match score measures how similar two images look. It does not mean there is a 93% chance the person is the one sought. Across a gallery of millions of faces, a high score can still return many lookalikes, so the score is a lead to check with other evidence.

ID Scan: How a 93% Face Match Put the Wrong Man in Cuffs

0:00-0:00

This episode is based on our article:

Read the full article →

ID Scan: How a 93% Face Match Put the Wrong Man in Cuffs

Full Episode Transcript


A man in Fort Myers, Florida, was arrested for a crime in Jacksonville Beach, about three hundred miles away. The reason was a face scan that said he was a ninety-three percent match. That number sounds like near certainty. It isn't.


His name is Robert Dillon, he's fifty-two, and he's

His name is Robert Dillon, he's fifty-two, and he's one of the people at the center of this story. According to the research behind this episode, he's the fifteenth person known to be wrongfully arrested after a facial recognition match. Most of those fifteen were Black.

If you've ever been photographed in public, or posted a picture online, this touches you. And if that feels unsettling, it should. But the fear gets smaller once you see where the mistake actually happens. It's not where most people think. So what does a ninety-three percent match really mean?

Start with what the system does. It doesn't confirm who someone is. It compares one photo against a huge gallery of faces, then hands back a ranked list of people who look similar.

Here's the part that surprises people. According to the research, the system returns that list of candidates whether or not the real person is in the database at all. It always hands back its best guesses. If the true suspect was never enrolled, somebody still lands at the top.


The misconception

Now the misconception. Most of us hear ninety-three percent and think there's a ninety-three percent chance this is the guilty person. That's a natural reading. It's how percentages work on tests, on weather forecasts, everywhere else.

But the score only measures how similar two images are. It says nothing about the odds that the person did it.

Then comes the math of scale. Galleries can hold hundreds of millions of faces. Even a system that's right nearly every time will, at that size, produce an enormous number of lookalikes. A high score applied to millions of faces still returns hundreds of thousands of false candidates. The score means very little without knowing how big the pool was.

For investigators, that means a score isn't evidence. For the rest of us, it means a lookalike can be a stranger who simply shares your cheekbones.


Trusted by Investigators Worldwide
Run Forensic-Grade Comparisons in Seconds
Detailed facial comparison reports. Results in seconds.
Get Started
7-day refund guarantee**

It also isn't equally fair to everyone

It also isn't equally fair to everyone. According to N.I.S.T., the federal standards agency that tests these systems, missed matches vary only a little between demographic groups. False matches vary much more. Women, older people, young people, and people of color all see higher rates of wrong matches.

That's why the pattern in those fifteen arrests matters.

So how does a wrong guess become a handcuffed person? Through a second step, and this one is human.

According to the A.C.L.U., police often take the system's top candidate and build a photo lineup around it. That's one suspect photo and five filler photos. A witness is asked to pick. But the candidate was chosen because he looks like the suspect. So of course he stands out.


Research on this is striking

Research on this is striking. People paired with a hint of guilt, like a system recommendation, were most often picked as the perpetrator. The machine's guess quietly steers the witness. Then the witness's pick gets treated as independent confirmation. It isn't independent at all.

Jalil Richardson spent over fifty days in jail after an eighty-five percent match. That's more than seven weeks, from a number that sounded like proof.

Think about a sketch artist. A sketch says, this person resembles the suspect. A fingerprint says, this is the suspect. Facial recognition is the sketch. But it's been treated like the fingerprint.

One more piece explains why. Every system has a cutoff for what counts as a match. Tighten it, and you get fewer false matches but miss real ones. Loosen it, and you catch more real ones but accuse more innocent people. Police tend to set it to catch suspects, not to avoid wrongful arrests.


Here's what it comes down to

So here's what it comes down to. The software isn't the real danger. The danger is treating a list of lookalikes as a confirmed identity.

In security, where you reject doubtful matches, that works. In a criminal case, doubtful matches get investigated as if they were solid.

Facial Recognition False Arrest Fargo Cops Face 10m Suit doesn't say who someone is. It says who looks similar. A ninety-three percent score measures resemblance, not guilt.

The scan took seconds. Clearing a man's name took months.


The Bottom Line

Whether you carry a badge or just carry a phone, knowing that a match is only a starting point is how you stay informed instead of afraid.

The full story's in the description if you want the deep dive.

Ready for forensic-grade facial comparison?

Full forensic reports with detailed similarity scoring. Results in seconds.

Run My First Search