Western Australia Police Deploy Real-Time Facial Recognition, Sparking Privacy Concerns

The arrest unfolded in real time, behind a bank of television microphones at Perth’s busy Mirrabooka bus station. As Western Australia’s police commissioner, Col Blanch, spoke to reporters on Thursday 2 July, a camera mounted on a nearby marked police van had been quietly panning across the interchange.

It can scan hundreds of faces a minute, instantly cross-referencing their profiles against a database of about 4,000 people: those with outstanding arrest warrants, registered child sexual abuse offenders and people reported missing.

Midway through the media conference, the algorithm pinged. Officers stepped into the crowd and led away a woman wanted for failing to appear in court.

“Commissioner, can we just get your reaction? There’s just been an arrest from the cameras behind you right now,” a reporter said.

“Been happening all week,” Blanch said. “I’d love to say I planned that but I think that’s exactly what the technology is … People are being arrested because they are wanted.”

The striking live demonstration marked the public debut of an Australian-first for law enforcement: a five-month mobile trial of real-time facial recognition technology.

While Australian police have used facial recognition retrospectively on stored footage for more than a decade, this trial uses live video feeds for real-time alerts.

In its first week, the single police van scanned more than 130,000 faces in crowded public spaces, generating 33 alerts that resulted in 19 arrests. The technology also prompted two welfare checks and identified 21 registered sex offenders who were “engaged with” by officers.

While WA police and the state government hail the trial as a cutting-edge tool for public safety, it has come under criticism from experts and regulators.

The state’s independent privacy watchdog, the Office of the Information Commissioner WA (OIC WA), said it was not involved in implementing the trial, warning of “collective privacy harms” in a statement.

“The OIC WA is aware of the trial but was not invited to participate in any formal consultation process with WA Police and did not provide input into the design of the trial,” it said.

‘This is not about mass surveillance’

Blanch said the van displayed visible signage and the technology did not retain data on ordinary citizens, pixelating people not in the database.

“This is not about mass surveillance,” he said. “This is about specifying those in the community who are wanted by police.

“This is actually a way that we can increase the freedoms and the privacy of our community. This is less of an intrusive capability than a standard CCTV camera.”

The OIC WA challenged the claim that short-term processing exempted the technology from scrutiny, saying that “even where this information is only held for a very short time, it is still a collection and obligations attach to that activity”.

Prof Dali Kaafar, the executive director of the Macquarie University cybersecurity hub, labels the commissioner’s comparison to standard security cameras a “false equivalence”.

“The whole point of live facial recognition is that it identifies people, and generic CCTV doesn’t,” he says. “The underlying reality is that every face in the crowd is being captured, processed and biometrically matched against a database. That is not what CCTV does.”

The OIC WA highlighted data from the 2026 Australian Community Attitudes to Privacy Survey showing 45% of Australians now consider facial recognition to be the biggest privacy risk they face – up from 27% in 2023. Public comfort with police using biometrics in public safety settings is on a downward trend, slipping to 57%.

Accuracy and algorithmic bias

Police acknowledged two false alerts during the first week – relying on a “human in the loop” to manually verify matches before acting – and experts warn that real-world environments degrade the technology’s accuracy.

The OIC WA note that the potential for false matches resulting in “unwarranted interventions” remains a significant risk, particularly when images are captured in variable conditions.

Commercial vendors may claim laboratory accuracy rates exceeding 99% but experts say public deployments tell a different story.

“The devil is in the detail, with accuracy dropping to about 90% to 95% in uncontrolled conditions,” says Prof Katina Michael, an expert in biometric surveillance at the University of Sydney. “Factors like variable lighting, crowd density, camera angles, motion blur, and the weather … error rates increase according to the amount of ‘noise’ in the environment.”

Dr Adam Andreotta, a business ethics lecturer at Curtin University, says that even a seemingly minor 10% error rate has massive consequences when applied to large crowds.

“When thousands of people walk by, a 10% mistake rate is obviously not good, it means a significant number of innocent people could be stopped, questioned or detained for things they didn’t do,” Andreotta says.

A primary concern among academics is the well-documented issue of algorithmic bias. Independent tests by the US National Institute of Standards and Technology (NIST) have historically shown that many facial recognition algorithms suffer from significantly higher false-match rates for women, younger people and individuals with darker skin tones.

“Multiple studies show African and Asian faces are wrongly matched at rates 10 to 100 times higher than white faces on some systems,” Kaafar says.

In an Australian context, experts say this bias will disproportionately affect marginalised groups, particularly Aboriginal and Torres Strait Islander communities who are already overrepresented on police databases.

“If the watchlist is already skewed toward a group, both the correct matches and the false matches will hit that group harder,” Kaafar says.

‘Function creep’

Critics and regulators are also deeply concerned by “function creep” – the gradual expansion of a technology’s deployment beyond its original boundaries. The OIC WA warns that shifting the trial towards covert deployments or integrating it into widespread, fixed CCTV networks would “significantly increase privacy risks”.

Michael notes that Australia still lacks federal or state legislation strictly dedicated to police live biometric surveillance that could safeguard against function creep.

“Without binding limits on watchlist criteria, data retention and permitted use cases set in laws and regulations … this expansion is a foreseeable risk, not a hypothetical one,” she says.

The OIC WA says the trial will serve as an early test for the state’s Privacy and Responsible Information Sharing Act 2024 (PRIS Act). It is the first law in Australia to explicitly regulate the use of personal information in automated decision-making. Under the act, WA police are required to assess the risks of bias, harm and discrimination in their automated systems and prove they are being managed. The act also mandates that a formal privacy impact assessment (PIA) be completed for high-risk projects.

The OIC WA confirmed it was now actively “monitoring and requesting information” from police about the trial to ensure compliance and shape future biometrics policies.

Blanch has refused to rule out using the live facial recognition van at public protests, saying it would depend on intelligence regarding significant risks of harm.

“Once real-time face-scanning is normalised at major events, expanding it to protests, then to routine deployment, then to covert deployment, [it] becomes a matter of policy nudges rather than a fundamental shift,” Kaafar says.

Commercial v public surveillance

The legal landscape for facial recognition in the private sector is shifting.

The retail companies Bunnings and Kmart faced scrutiny when the national privacy commissioner ruled they had breached the privacy of customers by scanning them to curb theft and fraud. The ruling against Bunnings was overturned on appeal in February.

Unlike retail shoppers, people walking through public entertainment precincts cannot easily choose to take their business elsewhere.

“You can’t just tell people it is happening via a sign,” Andreotta says. “If the public doesn’t fully understand the long-term risks of biometric collection and function creep, they aren’t giving meaningful consent.”

The OIC WA notes that, unlike credit card numbers, compromised biometric data cannot be easily changed or replaced, making it a “higher-risk category of personal information”.