Facial Recognition False Arrest: 9 Wrong Alerts, 1.38M Faces

Western Australia pointed cameras at one point three eight million faces over five months. Police say it produced one hundred twenty-two arrests. It also produced nine alerts that pointed at the wrong person.

If you've ever walked through a shopping area or a

If you've ever walked through a shopping area or a train station, your face could be part of a scan like this. Most of us never get asked. That's an uneasy thing to sit with, and it's reasonable to feel that way.

So here's what happened. Western Australia Police ran a live facial recognition trial at forty-three locations. That included Perth and regional areas. The cameras compared passing faces against a watch list. When the system flagged someone, officers stepped in to check. So the real question is this. Is a tiny error rate the same thing as a trustworthy system?

Start with the nine wrong alerts. Police Commissioner Col Blanch says none of those people were arrested. He says officers spoke with each of them, and nobody complained. He also said, quote, "We'd love it to be zero." Think about being one of those nine. You're walking along, and an officer stops you because a computer said you looked like someone else. You get sent on your way. But for a few minutes, a machine's guess decided how your day went.

The math. Police say the error rate is tiny, a

Now the math. Police say the error rate is tiny, a small fraction of one percent. They get that by comparing nine wrong alerts against more than a million faces scanned. Civil liberties critics say that's the wrong yardstick. Nearly every face scanned was never going to match anyone. Those faces make the error rate look small. The critics ask a different question. Of the people the system flagged, how many were flagged correctly? And how many of those arrests ended in a conviction? Police haven't released that last number.

That gap matters outside Australia too. A small error rate across a huge crowd can still mean real people get stopped. And a headline number can hide what happened to them.

There's a second data point. Researchers at the University of Essex studied live facial recognition in real-world use. They found that fourteen of twenty-two matches that led to stops were wrong. That's well over half. Lab tests often look better than the street does. Lighting changes. People move. Cameras sit at bad angles. So a system that shines in a lab can stumble on a busy sidewalk.

Errors aren't spread evenly

And errors aren't spread evenly. Research on bias in these systems shows that false match rates can differ by ten times, or even a hundred times, between demographic groups. So one overall accuracy figure can hide that some groups get wrongly flagged far more often. If you're in one of those groups, the average doesn't describe your risk.

For investigators, the lesson is practical. A match is a lead. It isn't proof of identity. Mistakes can come from the software. They can also come from the human reviewer who pairs the search photo with the wrong candidate. So a careful investigator checks every candidate and tries to rule people out before ruling anyone in. For the rest of us, it means a computer saying "match" should be the start of a question, not the end of one.

Here's the twist. The nine wrong alerts weren't stopped by the technology. They were stopped by people. Officers checked, talked, and let those people go. The safeguard that worked was human judgment, not the algorithm.