On June 22, Western Australian police became the nation's first law enforcement agency to use live facial recognition technology to find persons of interest.
Author
- Nicola Lockhart
Lecturer and Course Coordinator - Counter Terrorism, Intelligence & Security Science, Edith Cowan University
The trial involves a clearly marked van with cameras driving around Perth and Mandurah. It scans the faces of everyone it passes, comparing them against a watchlist of about 4,000 people. It includes those with outstanding arrest warrants, reportable offenders , people subject to lawful exclusion orders, and missing persons.
The exercise is overt and widely publicised , with the dates of deployments posted in advance.
In its first week, media reports put the trial at more than 130,000 faces scanned and 33 alerts. It led to 18 arrests , along with engagements with registered sex offenders. Work that would have taken weeks of conventional investigative effort has been done in days by a single van.
But harder questions surround its governance. Can we really call it a "trial"? Who gets to authorise such deployments? And who checks the safeguards?
Can we actually call it a 'trial'?
Any trial of new technology requires success criteria defined in advance, independent evaluation, and the possibility the technology won't be deployed on a permanent basis after the trial completes.
That's not what seems to be happening in WA.
WA police both runs the deployment and compiles the results it publishes, and no independent evaluator has been named. Its own privacy impact assessment says funding is not yet determined, and lists event security among the intended uses.
Arrests are an incomplete measure of success, because everyone arrested was already wanted. These are not crimes solved, but simply a known watchlist being worked through more quickly.
And an arrest count only records the hits, not the misses. What remains undisclosed is the human cost - any innocent people the system might misidentify and send officers to stop, question or detain. That cost determines whether a coercive power's benefits justify its harms, and therefore whether it should be kept.
WA police told media it recorded just one false positive, a rate of 0.0008% . Other reports state two errors . Measured against the 33 alerts that sent officers toward a person, the same data gives a much higher error rate of 3-6%. No official breakdown of those alerts has been released, nor any comparison with conventional policing over the same period.
Facial recognition algorithms have well-documented biases . They produce higher false-match rates for women, younger people, and people with darker skin. In a policing context, this raises the risk of wrongful stops for those groups.
The system in the van is NeoFace Watch from Japanese multinational company NEC. The UK's National Physical Laboratory (NPL) found no statistically significant race or gender bias in a version of NEC's algorithm at its default settings.
But the system uses a pre-set threshold to indicate a match between a face and the watchlist. When testers lowered that bar, which operators can do to catch more matches, false alerts fell disproportionately on darker-skinned people .
WA police states the algorithm has been benchmarked by NPL, but as far as we know, no results from WA's own deployment have been independently verified.
If a tech pilot only measures successes, it's not a trial - it's a product launch.
What about safeguards?
As the system scans people , it pixellates any faces on the operator's screen that don't generate an alert. The biometric data of anyone not matched is automatically and immediately deleted, and no data is shared with third parties.
Before any van is deployed, the authorising superintendent has to record why the operation is necessary and proportionate, why that particular location was chosen, and which categories of people belong on the alert list.
Every alert the system generates is reviewed by an officer before any action is taken, and WA police stresses that an algorithmic match cannot on its own justify an arrest .
These are good measures. But there are no independent audits of these safeguards in operation and no external party verifies that data deletion happens as described.
Governance is crucial
Poor governance can put a halt to technology that works just fine in practice. In 2020, a court ruling in the UK found South Wales Police's live facial recognition use unlawful.
Australia has no national legal framework for governing biometric surveillance. But with this police trial, WA has now become a national test case for such legislation.
The state's Privacy and Responsible Information Sharing Act 2024 is the first Australian law to directly govern how personal information is used in automated decision-making.
Under the act, WA police must be able to demonstrate how the risks of bias, harm and discrimination are being managed. Meanwhile, international human rights law permits interference with privacy only where it is lawful, necessary and proportionate.
Put simply, the benefits of live facial recognition must be measured against harm to the individual.
Does WA police's use of the technology comply with this? To find out, we need independent scrutiny. Early signs are not encouraging. The police trial, arguably WA's most consequential test of its new privacy law, was designed without the state's privacy regulator in the room .
What responsible governance would look like
What would responsible governance of live facial recognition look like?
There's a list of criteria any trial should meet:
- success and failure criteria published before a trial starts
- independent evaluation of the deployed system, with false alert rates broken down by demographic group
- independent benchmarking of the system as deployed
- external audit of any data privacy measures.
Additionally, the rules for live facial recognition should be legislated by the federal government: when a deployment can be authorised, what a watchlist may contain, and who provides oversight. These safeguards should not rest on internal police policy that can change without anyone outside the force noticing.
Nicola Lockhart does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.
