Facial Recognition System: WA Police Scan 1.4M Faces, 122 Arrests

Picture walking through a busy public place and never knowing a camera just checked your face against a police database. That is what happened to almost 1.4 million people in Western Australia, and the trial ended with 122 arrests. This is the story of how facial recognition works in real life, and why a flag from a camera is never the same thing as proof.

Facial recognition in the Western Australia trial scanned almost 1.4 million faces and led to 122 arrests, but a camera flag is only a lead that a human must check, never proof that someone did something wrong.

Let me answer the 11pm question first. Are you in danger? Probably not. But if a camera ever points at you by mistake, what happens next depends almost entirely on what the police do in the next ten minutes. That is the thing worth understanding.

What Western Australia's Facial Recognition Trial Actually Did

Western Australian Police finished what was described as Australia's first police trial of live facial recognition (cameras that check faces in a crowd as people walk by, in real time). According to the reporting, the cameras looked at almost 1.4 million faces. Police reported 122 arrests. The Guardian's headline rounded it to one million faces, but the larger number comes from the corroborating coverage, including MLex.

Police use facial recognition to scan one million faces in WA trial leading to 122 arrests

The Guardian, headline of the original report

Now, hold on to that percentage. It looks tiny, and it is. But it does not tell us the thing you actually care about. It does not say how many of those 122 people were the right people. It also does not say how many people were wrongly flagged and quietly let go. This article is part of a series, start with Facial Recognition False Arrest 9 Wrong Alerts 1 38m Faces.

How police facial recognition turns a face into a "possible match"

Here is the plain version. The facial recognition system takes your face and turns it into a set of measurements. It compares those measurements to a database of faces police are looking for. If two faces look alike enough, it produces a match score (a number for how alike two faces are). A high score means "these look similar." It does not mean "this is the same person." Think of a friend yelling "Hey, is that Dave?" across a parking lot. That is a guess, not an ID card.

Why a Facial Recognition Match Is a Lead, Not Proof

The research behind this story makes a point I wish more headlines made. A facial recognition scan that finds 122 people out of 1.4 million proves the facial recognition system works at scale. It does not prove it works correctly. Those are very different claims.

Leading face recognition systems do well in lab testing. The research puts their miss rate (false negatives, meaning wanted people the camera fails to spot) under 1%. Their wrong-person rate (false positives, meaning the camera says "that's him" about the wrong face) sits around 0.3%. That sounds small. Now do the math on 1.4 million faces. Even a tiny error rate, applied to that many people, can mean thousands of wrong flags. I am not saying that happened. I am saying nobody has shown us it did not.

Is facial recognition accurate enough to arrest someone?

Accurate enough to start looking, yes. Accurate enough to arrest on its own, no. Law enforcement good practice says a match is not grounds to detain anyone. Officers should do more checks first, such as comparing the two images carefully by eye, checking where the person was, and looking for other evidence. When that step gets skipped, facial recognition gets blamed for what is really a people problem.

Courts already know this. According to legal scholarship summarized in the research, no court has put facial recognition through rigorous scientific review and then accepted it as identification evidence. Law enforcement can use it to find a lead. A judge will want something stronger before anyone is convicted.

Western Australia Facial Recognition: What Is Missing From the Results

The trial numbers are real, and they are also incomplete. The researcher who dug into this found that the police have not reported false negatives. They have also not broken down how well the facial recognition system worked across different groups of people. Privacy concerns about the trial were raised publicly months earlier, in coverage by ABC News Australia, including questions about accuracy for different demographic groups.

Why does that matter? Because of what happened elsewhere. The Innocence Project reports at least seven confirmed cases of someone being wrongly identified by facial recognition, and six of those people were Black. Seven is a small number, and it is also seven real people who dealt with a mistaken accusation. Previously in this series: What Is A Biometric Passport A Chip Skimmed From 14 Inches.

Why the 122 arrests stick in our minds

Here is a quirk of the human brain. We judge how common something is by how easily an example comes to mind. Psychologists call it the availability heuristic. A headline that says "122 arrests" is vivid. A million people quietly walking past a camera is not. So the success stories feel like the whole story, and the unseen mistakes feel like they do not exist. Keep that in mind when you read any number like this.

To be fair, there is a real counterpoint. Surveys cited in the research show about 89 to 90% of people support police use of facial recognition for criminal investigations and finding missing people. Police also say unmatched faces are deleted immediately. Those are fair points. Most people are not against catching dangerous offenders. They are against being wrongly caught up in the search.

Is a Familiar Face or Voice Ever Proof? This Week's Other Warnings

The same lesson showed up twice more this week. Utah began enforcing digital age checks on adult websites, which means more people will be asked to complete identity verification with a face image or an ID. And three men were charged in a $129 million voice-cloning fraud scheme (software that copies how a real person sounds). Different stories, same trap. A face on a screen or a voice on the phone feels like certainty. It is not.

If you have ever tried to verify whether a photo, a profile, or a voice is really who it claims to be, that is the exact question this kind of technology exists to answer. Here is the one thing you can do: treat any single match as a first clue and ask what a second, independent check would show. That means a call back on a number you already know, or a careful side-by-side comparison of two face images by someone trained to do it. A match with no second check is a guess wearing a suit.

So go back to the hook. A camera can flag your face without proving you did anything wrong. Out of 1.4 million faces, 122 people were arrested, and we still cannot see the number that matters most: how many people were stopped, questioned, or scared by a match that turned out to be wrong?

Facial Recognition: Frequently Asked Questions