WA police scanned 1.38 million faces, made 122 arrests and logged nine false alerts

Western Australia Police scanned 1,380,882 faces in a five-month live facial recognition trial, producing 122 arrests and nine incorrect alerts. The trial concluded on Saturday. Commissioner Col Blanch defended the results at a media conference on Sunday, while civil liberties advocates accused the force of misrepresenting the error rate and the state's privacy watchdog confirmed it had no role in the scheme's design.

The numbers themselves are stark. Across 100 deployments at 43 regional and metropolitan locations, the system flagged far more than the arrests that followed: officers engaged with 146 registered sex offenders and carried out 10 welfare checks on located individuals. Nine alerts were wrong. Police say that works out to an error rate of 0.00065% — a figure Civil Liberties Australia rejects as statistical framing rather than substance.

What did the trial actually involve?

The mechanics matter for anyone assessing the claims. Cameras mounted on or positioned near marked police vehicles scanned faces in designated public areas. The system matched those faces against an alert list drawn from existing police databases. Police said images of people not on the list were pixelated in real time and deleted immediately without being stored — a claim that rests on the force's own account of the safeguards, not on independent verification.

The single largest deployment illustrates the scale. Three operations at the Perth Royal Show scanned 223,749 faces. Those scans produced 15 alerts, three arrests and seven "interactions" with registered sex offenders. At that event, the ratio of alerts to scans was roughly one in 15,000. That is the pattern across the trial: large volumes of passive scanning, a thin layer of alerts, and an even thinner layer of arrests.

How did police handle the nine false alerts?

Each of the nine incorrect alerts led to a police interaction with an innocent member of the public. Blanch confirmed officers approached every person flagged, and he offered an unusually direct account of what that meant in practice.

"I want to be very clear on those false alerts," he said. "No one was arrested. No one made a complaint. Each one did result in an interaction with that person and asking them relevant questions."

The commissioner framed the error count as small but nonzero. "It's a very, very, very low number. It's not zero. We'd love it to be zero. But I think the important thing for us is to engage with our community just like we would if a human thought someone was a wanted person."

Blanch acknowledged the technology was "not perfect" and that authorities "knew this going into this trial." His defense rests on an equivalence: that an algorithmic alert deserves no more scrutiny than a tip from a member of the public. Critics would note that a human tip involves one person's suspicion of one individual, whereas the trial involved nearly 1.4 million scans of people going about their day. The absence of complaints is also the force's own reporting; nine strangers questioned by police may not all have known an algorithm had flagged them.

Why is the error rate disputed?

Bill Rowlings, chief executive of Civil Liberties Australia, said WA police were "severely gilding the lily" by claiming the incorrect alert rate was 0.00065%. His argument is arithmetic: the police figure divides nine false alerts by 1.38 million scans, counting every uneventful scan of a non-flagged person as a correct outcome.

"We will only know what the incorrect rate is when the 122 people arrested have their day in court," Rowlings said. "If half the 122 arrested … are not convicted, the 'incorrect' rate will be 50%."

He added that police had performed the almost 1.4 million scans "with no reasonable suspicion whatsoever." That phrase cuts to the core legal and ethical objection: conventional policing generally requires individualized grounds for stopping someone, while this trial inverted the sequence, scanning everyone and stopping only those the system selected.

Neither figure settles the question on its own. The police rate measures the system against the entire scanned population; Rowlings' hypothetical rate measures it against outcomes. Conviction data for the 122 arrests does not yet exist, and until it does, both claims outrun the available evidence.

What did the privacy watchdog say?

The Office of the Information Commissioner WA (OIC WA), the state's independent privacy watchdog, was not involved in implementing the trial. When the trial was announced, the OIC WA warned of "collective privacy harms" in a statement. The office was contacted for comment following the trial's conclusion; no response was included in the announcement coverage.

The watchdog's exclusion from implementation is a notable gap. An agency specifically tasked with information privacy raised harms in advance, and the force proceeded without its involvement. Blanch's Sunday statement emphasized process: the trial "focused on using modern technology responsibly to identify people who presented a risk, while allowing the vast majority of people to go about their day without any impact." Whether "no impact" describes the experience of the nine incorrectly flagged individuals — or of the 146 sex offenders engaged without any stated arrest outcome — is a question the statement does not address.

What about algorithmic bias?

The trial previously triggered backlash over privacy and algorithmic bias. Experts say the technology could disproportionately affect marginalised groups already overrepresented on police watchlists. Facial recognition algorithms have historically shown higher false-match rates for people with darker skin tones. The announced figures — nine incorrect alerts, 122 arrests — do not break down outcomes by demographic group, so the bias concern remains neither confirmed nor refuted by this trial's data.

That absence of demographic reporting is itself a finding. A trial of this scale that cannot, or does not, report whether its errors fell unevenly across population groups leaves the central bias question exactly where it started.

What happens next?

The surveillance vans have been removed from public deployment. The force is conducting an internal evaluation of how the technology performed on the ground and its community impacts, including "reviewing the safeguards and whether all requirements were met."

That evaluation is internal, not independent. The police themselves will assess whether police met their own requirements, with no stated role for the OIC WA or any external reviewer. The 122 arrests will move through the courts, and the resulting convictions or acquittals will provide the first test of Rowlings' claim that the true error rate is far higher than the advertised 0.00065%.

For jurisdictions elsewhere watching real-time facial recognition trials, WA offers both a template and a caution: mass scanning can produce arrests at scale, but the numbers police choose to publish — scans, arrests, alert rates — are only part of the record. The rest sits in court outcomes and in the nine interactions the commissioner described, and the force's internal review will determine whether the vans return.