Facial Recognition Bias: Racial Bias in 905,117 Police Scans

Picture a normal day out in Western Australia. You walk past a camera, and it scans your face. Nobody tells you. That happened roughly 905,117 times in three months, and it puts facial recognition bias in front of ordinary people, not just suspects.

Facial recognition bias and the 900,000 people nobody warned

Let's start with the part that should make you sit up. Yahoo News Australia reported that Western Australia Police ran a three-month trial across 36 locations. The cameras scanned 905,117 faces. Almost all of those people were shoppers, commuters and families who were never suspected of anything.

Hidden detail in police facial recognition trial after 900,000 Aussies unknowingly caught up

Headline of the original report, Yahoo News Australia

Did it work? In some ways, yes. The trial led to 79 arrests. Police also say it flagged 114 registered sex offenders, who were then approached, and led to eight welfare checks (police stopping to see if someone is okay). Nobody should wave that away.

But here is the thing I would text you at 11pm. A system can be useful and still be unfair. Both can be true at once.

Why a 0.0009 per cent error rate can mislead

Police described the eight wrong alerts as a 0.0009 per cent error rate. That sounds like nothing. But look at the bottom of that fraction. It is every face that walked past, not just the faces the system flagged. Divide a handful of mistakes by almost a million passers-by and nearly any system looks perfect. This article is part of a series, start with Biometric Login Uae Banks Drop Text Codes Scammers Pounce.

Lauren Perry, a researcher at the University of Technology Sydney, warns that figures like this can give a false impression of how well the facial recognition technology tells wanted people from everyone else. She is right to. A small number tells you nothing about who got the mistakes.

And here is what police themselves acknowledged. The eight wrong alerts were linked to darker skin tones, poor lighting and low-quality reference photos (the saved picture of the person police were looking for). That is facial recognition bias showing up inside a trial that was being called a success, and darker skin tones are exactly where racial bias is usually measured.

What facial recognition technology bias looks like in the numbers

Is this just a one-off? No. The US government's testing lab, NIST (the National Institute of Standards and Technology), has studied this for years. Its research found that a majority of the recognition algorithms it tested made different kinds of mistakes for different demographic groups, which is a form of demographic bias.

The bigger problem is the false positive (the system says "match" when it is not the same person). Across groups, false positive rates could differ by a factor of ten, or even a hundred. The Federation of American Scientists summarised it this way: as of March 2025, the top performing system produced 358 times as many false positives for West African women over 65 as for Eastern European men aged 35 to 50. Their write-up on face recognition bias lays out the details.

Read that again. It was the best system, and the gap was 358 to 1. That does not prove the Western Australia system behaves the same way. It does show why "the overall error rate is tiny" is the wrong thing to be reassured by.

Who pays when the match is wrong?

When a person is wrongly flagged, the cost is not a rounding error. A stranger in uniform walks up to you in public and says the computer thinks you are someone else. You are embarrassed, delayed, maybe scared. If you are in a group the system gets wrong more often, that happens to you more often. Same camera, same shopping centre, a very different day.

Australia police facial recognition: what was missing before launch?

The detail that stings most is not technical. Western Australia Police did not consult the state's privacy watchdog before the trial began. In plain terms, nobody independent checked the plan before 900,000 faces went through the system. Whatever you think of the technology, that order of events is backwards for law enforcement. Previously in this series: Age Verification Florida Chatbots Next Then Every Phone.

The Conversation asked whether the trial was an ethical success even though the operational results were strong. That is the right question. A trial can catch people and still fail the public on notice, consent and fairness. For anyone following Australia police facial recognition, those are the tests that count.

Is facial recognition biased against some skin tones?

Often, yes. A NIST study found that many systems make more mistakes for some demographic groups than others. Police in this trial linked their wrong alerts to darker skin tones, but a handful of cases cannot show how big the gap is. That is why experts want results broken out by group, not one tidy overall number.

Now the part that is about you. If you have ever wondered whether a face in a photo or a profile is really the person it claims to be, that is exactly the question this kind of technology exists to answer. The useful habit is simple. Before you trust any face-matching result, ask for the error rate for different groups of people, plus how the photos were checked. A tool that cannot answer is guessing with a nice interface.

So here is where I land. The tiny error rate is the least interesting number in this story. The interesting one is zero: the number of public warnings those 905,117 people got. If the cameras are as fair as police say, a breakdown by group would prove it. So why wasn't one published? Up next: Facial Recognition Bias What 905 117 Police Scans Hide.

facial recognition bias: Frequently Asked Questions