Science1 publisher2 min readPublished
One alert in 26 from Western Australia's facial recognition trial was a wrong match
Since June 2026 the Western Australian Police Force scanned more than 900,000 faces, logged 209 alerts, made 79 arrests and recorded eight incorrect matches. What those eight mean depends on which total you divide them by.
The Scientist · Science desk
What happened
- The Western Australian Police Force has been running Australia's first live facial recognition trial since June 2026, with cameras in public metropolitan and regional locations including major crowd events.
- More than 900,000 faces were scanned over the course of the trial, and 209 of them were flagged as matches to someone on a police watch list.
- The trial produced 79 arrests.
- Eight alerts were incorrect, and Police Commissioner Col Blanch attributed those errors to factors including darker skin, lighting conditions and the quality of the reference image.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- constraint Eight errors is too thin a sample to test whether misidentification falls unevenly across groups, so the bias question stays open however the trial's totals are characterised.
- exposure A wrong match costs the person an officer approaches in public, and the phys.org analysis argues that burden is what the released totals leave unaddressed.
- decision States and territories weighing the same cameras now have a reporting template, and each has to decide whether to demand per-group error rates and watch-list criteria before signing off.
- precedent Australian policing has already settled into arguing about AI enforcement tools on effectiveness totals, and this release invites the same argument for live face matching.
Eight errors is a small count, and its size depends on what you divide it by. Against the more than 900,000 faces the cameras processed, it works out at one incorrect alert per 112,500 scans [2]. Against the 209 alerts that reached officers, it is 3.8 percent, about one in 26 [1]. The first figure describes the software. The second describes the chance that a police approach in this trial began with a machine error.
Both denominators are defensible, for different questions. Nearly everyone in the 900,000 was never on a watch list, and WA police said images of those people were pixelated in real time and not saved [7]. The scan total is the right base for a privacy question. For an error question the alert is the event that counts: an alert is what sends an officer toward a person.
The yield was one alert per 4,306 faces processed [5]. Of the 209 alerts, 79 ended in an arrest, or 37.8 percent [3], and eight were logged as wrong, leaving 122 that the release does not tie to either outcome [4]. Police said the technology identified and engaged 114 registered sex offenders [4], a count that may include people among the 79 arrests.
A miss rate is harder to come by. Counting false negatives means knowing how many people on the watch list walked past a camera without triggering an alert, and that needs a source of truth outside the system; the released results report neither false negatives nor performance by demographic group [8]. Experts cited in the phys.org account have cautioned that the technology may perform differently across demographic groups [11].
Commissioner Blanch's explanation of the failures sits awkwardly against that. Eight events cannot test whether darker skin raises the error rate: split across groups the counts fall to ones and twos, and any comparison of rates also needs the number of faces scanned in each group [6].
Australian police have deployed AI enforcement tools before. Road safety cameras identify drivers using mobile phones and not wearing seat belts, police have promoted them as life-saving and presented evidence of their effectiveness [13], and WA authorities reported a significant drop in mobile phone and seat belt infringement cases after they went in [14]. The phys.org analysis of the face trial argues that the question worth asking is what happens if the technology fails and who bears the burden when it does [10].
What to watch
- Whether WA publishes alert and error counts by demographic group together with the number of faces scanned in each group.
- Whether future WA reporting states a disposition for every alert rather than only arrests.
- Which other state or territory cites the WA totals when announcing a live trial of its own.