A trial that scaled before its rules did
Western Australia Police began a live facial recognition trial in Perth and Fremantle on 22 June 2026. Cameras near a marked police van feed NEC's Neoface m40 algorithm, which compares passers-by against a police alert list of about 4,000 images. In the first week the system scanned more than 130,000 faces, generated 33 alerts and contributed to 19 arrests, according to AI Weekly's summary of the reporting. Biometric Update reports that one false match was documented in week one and that later data has not been published.
The policing case deserves a fair hearing. The alert list is described as covering people wanted for serious offences, reportable offenders, missing persons and people subject to lawful restrictions. Police say images of people not on the list are pixelated in real time and not stored. If those claims hold, live matching can be less intrusive than the alternatives: a human officer scanning a crowd from memory is slower and no less biased, and nineteen arrests of listed people is not nothing. A trial is also the right instrument for a new tool, because it generates the evidence on which a permanent decision should rest.
The problem is not that WA is trialling. It is that the trial's governance looks thinner than its technology.
What the critics have documented
The reported criticisms are specific and mostly about process, which is why they are hard to wave away.
- Consultation. Aboriginal Legal Service director Peter Collins says the service received 20 minutes' notice before launch and called the approach "an exercise in tokenism". WA's Information Commissioner said it was not invited into meaningful consultation, per Biometric Update.
- The impact assessment. Police's privacy impact assessment reportedly concluded the trial would not "primarily or especially affect Aboriginal people", which meant no separate Aboriginal Information Assessment was required. Collins argues that deploying in areas with significant Aboriginal populations "can only have been by design, not chance". Police say locations reflect "policing needs and policing objectives". Former Privacy Commissioner Malcolm Crompton and other experts have criticised the assessment as thin.
- Accuracy evidence. Algorithms have historically performed worse on women and people with darker skin. A 2023 UK study found no significant demographic differentials for NEC's technology, but the Ada Lovelace Institute's Michael Birtwisle notes it did not examine First Nations faces specifically.
- The watchlist. How names enter the list, who reviews it and how long they stay on it has not been made public.
Each of these is fixable without abandoning the technology. That is the point.
The legal vacuum around the trial
Australia has no dedicated facial recognition statute. The closest thing to a standard is the privacy regulator's guidance, and it is aimed elsewhere. On 29 July 2026 the OAIC updated its guidance on facial recognition in retail and other publicly accessible spaces. It follows the Administrative Review Tribunal's affirmation of the Privacy Commissioner's determination against Bunnings, which confirmed, in the OAIC's words, that "there is a high bar for using facial recognition technology in Australia" and that a precautionary approach is required. That guidance addresses organisations covered by the federal Privacy Act. It says nothing, in the page we reviewed, about state police.
State police fall to state law. WA's Privacy and Responsible Information Sharing Act 2024 does cover the Police Force of Western Australia, but most of its obligations only commenced on 1 July 2026, and several protections apply only to information collected on or after that date. The trial began nine days earlier. The new Office of the Information Commissioner says it was never invited into the trial's design.
The result is an awkward asymmetry. A hardware retailer faced a multi-year regulatory process and a tribunal appeal for scanning customers' faces in shops. A police force can scan 130,000 faces in a week on a trial basis with an internal assessment and no external sign-off. Whatever one thinks of either decision, the standards should not differ this sharply by who holds the camera, since the state's powers over the person scanned are far greater.
Proportionate safeguards, not a ban
Our view is pro-innovation: biometric tools that find missing persons or dangerous offenders can deliver real public value, and blanket bans tend to freeze the technology in the hands of the least accountable users. But legitimacy is a precondition for deployment, not a courtesy after it. Five measures would cost the trial little and the public trust a great deal:
- Independent oversight. Bring the Information Commissioner in before any extension or permanent rollout, with the power to pause the trial.
- Publish the criteria. The alert list's eligibility rules, review process and maximum retention period should be public, even if individual names are not.
- Test on the people affected. Commission accuracy testing that includes First Nations faces, and release the results with the false-match rate per scan, not just arrest counts.
- Report fully and regularly. Alerts, human-review overturns, false matches and arrests by location, published on a fixed schedule rather than by media request.
- A sunset clause. Make continuation conditional on a published independent evaluation and, if the technology is to stay, on a statutory basis in Parliament.
The wider lesson
The cost of getting this wrong is not only to the people who are misidentified. Trials that proceed over the objections of the communities most affected tend to produce backlash that closes off the legitimate uses too. A statutory framework with clear purposes, thresholds and oversight would let Australian police use the technology where it works and give citizens a way to contest it where it does not. Without one, WA is learning on 130,000 faces what a law could have taught in advance.