WA Police Facial Recognition Trial Got 9 of Its 292 Alerts Wrong

Frank Keding did the division on a Facebook post about the trial before most of the coverage had. “That means a 3.08% false positive rate of the alerts,” he wrote under PerthNow’s report of the results, and then he gave the police figure its analogy: “That’s like boasting that a breathalyser is 99.999% accurate because it correctly ignores sober people.” Jason Makin, on the same post, put it shorter: “9 out of 292 alerts is 3%.” The WA police facial recognition trial ran from 22 June to 31 October 2026.

Police divided nine incorrect alerts by 1,380,882 faces scanned across 100 deployments at 43 locations. That is the figure on the Western Australia Police Force page that reports the trial’s conclusion, alongside 122 arrests, 146 registered sex offenders engaged by officers, and 10 people located for welfare checks. The same page gives the other denominator in the same sentence: the cameras generated 292 alerts, including the nine incorrect ones. Nine of 292 is about three in every hundred alerts. An alert is the only moment in the whole operation when anything happens to a person. A scan that produces no alert produces nothing at all, and police say the image is pixelated in real time and deleted immediately, with no record kept.

Makin called the nine “false arrests.” The police account is narrower, and it matters here because the commenters and the commissioner are describing different events. Commissioner Col Blanch said at the 1 November media conference that officers engaged with each person who was wrongly flagged, and that “No one was arrested. No one made a complaint.” Donna Green, also on the PerthNow post, asked the question that sits between the two accounts: “How many of the 292 matches were NOT arrested and how many of those were innocent people who were stopped, questioned and had their ID checked?” The published summary does not answer her. It does not say how many of the 122 arrests followed directly from an alert, and it does not say what happened to the alerts that produced neither an arrest nor an incorrect flag.

WA Police’s reporting by district runs to 11 October and splits every alert three ways: positive, where the identity matched; incorrect; and unconfirmed, defined in the sheet as “person not engaged.” Across 88 deployments at 40 locations, it records 252 alerts: 210 positive, 34 unconfirmed and 8 incorrect, with 93 arrests, 134 registered sex offenders engaged, 10 vulnerable people and 54 outcomes listed as other. The outcomes do not add up to the alerts, and the sheet does not pretend they do. A deployment in Geraldton, in the September update, produced two further arrests, a summons, six move-on notices and an infringement beyond its six alerts. What the sheet establishes is simpler: more than one alert in eight, to 11 October, was left unconfirmed. The final summary, published after 31 October, keeps the total alerts and the incorrect count and drops the split. The last 20 days added 40 alerts and one incorrect alert, with no published account of how many of the 40 were positive, incorrect or unconfirmed.

The force’s update of 22 September recorded 905,117 faces scanned, 209 alerts and eight incorrect alerts, an incorrect-alert rate of 0.0009% on the per-scan measure, with 79 arrests. Eight of 209 alerts is a higher share than nine of 292. The trial’s false-alert share of alerts fell as the trial went on, and the scan count rose fastest where the crowds were thickest. Three deployments at the Perth Royal Show scanned 223,749 faces and generated 15 alerts, three arrests and seven registered sex offender engagements. One event, out of 100 deployments, supplied 223,749 of the 1,380,882 faces in the denominator police used.

The Privacy Impact Assessment of May 2026 names NEC as the vendor and says the technology, if endorsed, would be NEC’s NeoFace Watch using the NeoFace M40 algorithm, the system used by London’s Metropolitan Police, benchmarked by the United States National Institute of Standards and Technology and the United Kingdom’s National Physical Laboratory. The assessment says alerts are reviewed by trained officers before engagement, that there are no automated decisions, and that alerts are treated as lines of enquiry. It says a person walking through the recognition zone cannot opt out of being captured. It records that funding for the project was not yet determined. On fairness, it says algorithm calibration includes monitoring data to identify any demographic basis, and that post-deployment reviews will assess detection accuracy and false positive rates. Independent audits and oversight by governance committees, it says, will provide additional scrutiny. None of those reviews, audits or demographic breakdowns appears in the published trial record. Blanch, discussing the eight interim incorrect alerts, attributed them to factors including darker skin and lighting conditions, or the quality of the reference image. A 2023 evaluation by the National Physical Laboratory found no significant demographic differences in NEC’s real-time algorithms at the recommended settings, but the Ada Lovelace Institute’s Michael Birtwisle has pointed out that the study did not specifically examine performance on First Nations people. The assessment itself answered “No” to the question whether the activity would primarily or especially affect Aboriginal people, and so no Aboriginal Information Assessment was completed.

The Office of the Information Commissioner for Western Australia said it was not invited to take part in formal consultation or to provide input into the design of the trial. Police said the office was briefed in April and invited to a demonstration. In July the office warned about collective privacy harms, about behaviour changing when people feel watched, and about what wider or covert deployment would do to the risk. It pointed to the Privacy and Responsible Information Sharing Act 2024 and its requirement that entities automating decisions assess the risks of bias, harm and discrimination. Peter Collins, legal services director at the Aboriginal Legal Service of Western Australia, said he was contacted two days before the launch and asked for 20 minutes, and called the approach “an exercise in tokenism.” Police said deployment locations were chosen on policing needs and objectives.

Grant Moulds, on the same PerthNow post where Keding did his division, wanted the vans everywhere: “Good. Roll it out everywhere. If you have nothing to hide, you have nothing to fear.” His comment answers a different question from the one the spreadsheet raises. The commenters who divided nine by 292 got the arithmetic right and the consequence wrong, calling engagements arrests. The officials who divided nine by 1,380,882 got their arithmetic right and chose the denominator that makes the system look most accurate. What would settle which denominator judges the trial is the evidence the pre-trial paperwork promised: the demographic checks, the independent audits, the reviews of false positive rates.

WA Police will evaluate the trial itself. The evaluation announced on the force’s page will examine system performance, operational effectiveness, community impact, and governance and compliance. Faces that did not match anyone on the alert list were pixelated in real time and deleted immediately, with no record kept of the data.