WA Police Scanned 1.3 Million Faces, Made 122 Arrests — and 9 Wrong Matches
A five-month real-time facial recognition trial in Western Australia produced arrests and welfare checks, but critics say the real error rate is still unknown.
What happened: Western Australia police finished a five-month trial of real-time facial recognition, scanning 1,380,882 faces across 100 deployments in 43 locations, including three runs at the Perth Royal Show that alone scanned 223,749 people. The system triggered alerts leading to 122 arrests, 146 interactions with registered sex offenders, 10 welfare checks, and nine confirmed false matches.
Why it matters: Police called the error rate vanishingly small, about 0.00065 percent, but civil liberties groups say that math is misleading. Civil Liberties Australia argues the honest error rate can only be judged once the 122 arrested people go through court; if half aren't convicted, the real mistake rate looks very different. The Human Rights Law Centre says each false match is a real person stopped and questioned because an algorithm got it wrong, and questions whether scanning over a million people, including at public protests, is proportionate at all.
How it works, plainly: Cameras mounted on or near marked police vehicles scanned faces in designated public spots and checked them instantly against a police alert list of wanted or flagged people. Anyone not on the list was pixelated on screen in real time and, police say, deleted immediately without being stored. Only matches against the watchlist triggered an officer follow-up.
The rollout: The surveillance vans have now been pulled from public use while police run an internal evaluation of how the system performed and whether safeguards held up. WA's independent privacy watchdog says it had no role in designing the trial and had already warned of collective privacy harms. Legal advocates say the public still doesn't know what rules govern misuse, bias testing, or accountability, calling the published policies a thin copy of overseas frameworks.
