One million, three hundred eighty thousand, eight hundred eighty-two people walked past a police van in Western Australia between 22 June and 31 October and had their face read by a computer without being asked, without being suspected of anything, and without being told that nine of them would be treated as the cost of doing business. That is the number WA Police put in its own closing summary, and it is the only number in the whole exercise that was not chosen to flatter the force.
The trial ran 100 deployments across 43 sites. It generated 292 alerts. It produced 122 arrests, led officers to 146 registered sex offenders and to ten people who needed a welfare check. And it produced nine incorrect alerts, which the force has converted into a headline statistic: an error rate of 0.00065%, as the Guardian reported and Nine News repeated.
That figure is correct arithmetic and terrible accounting. Divide nine by 1.38 million and you get a number small enough to vanish inside a press release. Divide nine by the 292 times the system told an officer to act, and the shape changes. Divide nine by the nine people stopped on a pavement because a model misread a face they have owned all their lives, and there is no percentage — just a person answering questions nobody should have asked.
Col Blanch — The Error Rate Is Real, and It Is the Wrong Number
Commissioner Col Blanch met the press on Sunday and did the thing competent institutions do when they have decided to keep a tool: he conceded the flaw early and framed it as honesty. The technology is not perfect, he said. The force knew that going in. Each of the nine false alerts was followed up by officers asking questions, exactly as they would if a member of the public had pointed at a stranger and said that one. Nobody was arrested on a false alert. Nobody complained.
Read that twice, because it is the whole defence. Nobody complained. In a country where being stopped by police is a routine indignity in some communities and a frightening event in others, the absence of formal complaints measures Australian patience, not system accuracy. The nine were pulled aside, questioned, and released into a day that no longer belonged to them. They did not agree to be the quality-assurance sample.
The force also says non-matching faces were pixelated in real time and never stored. That is a real design choice, and a promise from the operator about software nobody outside can inspect. Trust is not a safeguard. It is a mood.
Bill Rowlings — Judge the System by the Courtroom, Not by the Scan Count
Bill Rowlings, chief executive of Civil Liberties Australia, called the presentation exactly what it is: a number dressed up for an audience. His argument is precise and it has not been answered. The honest error metric is not scans per false alert. It is outcomes per arrest.
The 122 arrests are, for now, accusations. If a large share never produce a conviction, the system did not deliver 122 results — it delivered fewer results and dragged more people into the criminal process on the strength of a camera. Rowlings did the arithmetic out loud: if half of those arrested are never convicted, the real error rate is half.
And beneath the arithmetic sits what the force never has to defend in a media conference: 1.38 million people scanned with no reasonable suspicion whatsoever. That is the design. Suspicion is what the system manufactures after the fact.
An arrest is an accusation with handcuffs on it. It is not a result, and it is not a conviction, and the difference between those three things is where civil liberties go to die quietly.
Sarah Schwartz — 1.38 Million Faces Is Not a Trial, It Is Infrastructure
Sarah Schwartz, legal director at the Human Rights Law Centre, refuses the framing that a low false-alert count settles anything. Every false match is a real person who may be stopped, questioned or investigated because a model got it wrong, she said — and the consequences land on that person, not the error rate.
Her larger objection is scale. Biometric surveillance of more than a million people is not a pilot program; it is a policy wearing a lanyard. She has also flagged the deployment of these vans around public demonstrations, which should end any claim that this is only about catching wanted offenders. A crowd at a protest is a crowd of faces, and a system that reads crowds does not care why they gathered.
Her sharpest point is governance. WA Police have published remarkably little about the safeguards, and what exists reads like it was borrowed from jurisdictions that made other choices under other laws. Misuse, bias, accountability: unanswered, and none of it is in the media release.
The Perth Royal Show — 223,749 Faces on a Family Outing
Forget the statewide totals and look at three operations at the Perth Royal Show. Showbags, dagwood dogs, children on their father’s shoulders, and 223,749 faces scanned. That produced 15 alerts, three arrests and seven interactions with registered sex offenders.
Three arrests for a quarter of a million scanned faces is defensible if you believe the alternative was zero. It is a poor advertisement for mass screening if what you are measuring is whether it works. And it tells you what the capability is for: not investigations, which begin with a crime, but attendance. The van does not need a reason. It needs a crowd.
The Regulator — ‘Collective Privacy Harms’ Was the Warning, and Nobody Had to Answer It
Western Australia’s information commissioner said in July, before the vans became a fixture, that her office was not involved in implementing the trial and warned about collective privacy harms — the kind that do not show up in any individual’s complaint because they are distributed across an entire population that never consented to being measured. The statement is still on the state government’s own website.
Read that again. The state’s independent privacy regulator was outside the loop of a mass biometric deployment, said so in writing, and it proceeded. That is not an oversight failure inside a police force. It is an oversight architecture that was never built, and a regulator documenting its own irrelevance in a PDF nobody had to read.
The vans are withdrawn while the force conducts an internal evaluation of performance, community impact, safeguards and compliance. Internal. The operator is marking its own homework, and the public will get a summary of the marking.
The Nine — ‘No One Was Arrested’ Is a Defence That Concedes the Point
The nine incorrect alerts are filed as a technical footnote. They are the story at human scale. Nine times in five months someone going about an ordinary day was approached by police because a model produced a match that was not there. Each was innocent of the thing they were stopped for, and each is now part of the statistic the force uses to prove the system works.
Those errors do not distribute randomly, and the research is not new. Facial recognition has a documented history of higher false-match rates for people with darker skin tones, and watchlists come from databases that record decades of over-policing in particular neighbourhoods. Errors concentrate where policing already concentrates. That is not a hypothetical about bias; it is what happens when old policing data goes into a pattern matcher and the output is called neutral.
When the trial was announced in June, with cameras on marked vans at sites announced in advance, the force sold it as crime prevention with safeguards. Five months on, the evidence for the first claim is 122 arrests, and the evidence against the second is a privacy regulator nobody asked in.
What Happens Next — The Vans Come Back, Because Capability Does Not Get Un-Built
There is no version of this story where Western Australia decides the cameras are too dangerous and melts them. The evaluation will land, the force will report strong performance, the arrests and sex-offender contacts will be listed first, and the nine false alerts will sit in a paragraph near the bottom with the word minimal attached. A minister will call it measured and evidence-based. Then it stops being a trial, which is what trials are for.
That is the pattern wherever this is tried, and it is why the denominator matters. A system judged by scans gets to call its mistakes microscopic. A system judged by the people it stopped — nine of them, all innocent, all questioned — has to defend a harder proposition: that 1.38 million Australians should be read by a camera so 122 can be arrested, and that the nine should be grateful nobody complained.
The vans are parked. The capability is not. It never is.
Sources: The Guardian, Nine News, Office of the Information Commissioner WA.