A Face Match Is Not An Arrest Warrant – OpEd

October 21, 2026

By Burak Oktenli


Key Takeaways:

The article distinguishes facial-recognition alerts from adequate evidence for police action, citing Western Australia’s trial of more than 905,000 scans that produced 209 alerts and eight incorrect ones (0.0009 percent), alongside U.S. cases such as Kimberlee Williams’s 2021 wrongful arrest after a false match.

It notes differing legal approaches, including Western Australia’s rule that a live alert alone is not grounds for arrest, the EU AI Act’s limits on real-time remote biometric identification, and Detroit’s 2024 settlement requiring independent evidence beyond a machine lead for warrants.

The author proposes a transferable test: if the algorithmic suggestion were removed, what reliable, independently obtained evidence would still connect the person to the offense, and urges reporting of false alerts, interventions, and downstream errors rather than relying on headline accuracy rates alone.

From wrongful arrests in America to growing biometric policing across continents, the essential safeguard is evidence that stands independently of the algorithm.

A tiny percentage can obscure a consequential question. In an update to its 2026 live facial-recognition trial, Western Australia Police reported more than 905,000 faces scanned, 209 alerts and eight incorrect alerts. Its published incorrect-alert rate was 0.0009 percent of faces scanned. That is a useful measure of how rarely a scan generated a reported incorrect alert. It is not a measure of whether the evidence supporting a later police intervention was adequate.

The distinction becomes painfully concrete in the United States. Kimberlee Williams was arrested in Oklahoma in 2021 on a Maryland warrant after a facial-recognition search incorrectly linked her to bank fraud. According to her lawyers, she had never been to Maryland and spent six months in jail before the charges were dismissed. In April 2026, the American Civil Liberties Union described her as the fourteenth publicly known person wrongfully arrested in the country after police relied on erroneous facial-recognition results. That was the organization’s documented count, not a national estimate of all such cases.


These are different uses and different events. The Australian figures do not establish a wrongful arrest there, and Williams’s case does not show that every police deployment is defective. Together, they put the real question into focus: when a machine identifies a possible face, what must happen before the state acts against the person behind it?
Two different jobs, two different evidentiary tests

Police use facial recognition for at least two distinct tasks. Live systems may help locate someone whose identity is already established and for whom a lawful warrant exists. Retrospective searches may instead produce the first suggested name for an unknown person in a crime-scene image. In the first task, officers still need to verify that the individual encountered is the person sought. In the second, they must independently investigate whether the suggested person is connected to the offense.


Both tasks can be valuable. Britain’s Home Office reported that live facial recognition helped the Metropolitan Police locate people in cases associated with 962 arrests over the year ending September 2025. The same consultation reported more than 25,000 retrospective facial searches monthly on the Police National Database. These are government-reported operational figures, not proof that every match or arrest was justified. They explain why a serious policy cannot simply pretend that police have no use for the technology.

Nor is the output self-validating. A similarity score indicates that a system finds two faces alike under a particular method and threshold. It does not establish identity beyond doubt, still less criminal involvement. The National Institute of Standards and Technology’s ongoing one-to-many evaluations show that performance varies with algorithms, test conditions and demographic groups. Their findings must be interpreted for the kind of search actually being used rather than transferred indiscriminately between verification and identification tasks.

Human review can help catch false alerts, but it is not independent evidence merely because a person looks at the same photographs. If a computer selects an innocent lookalike and detectives place that photograph in a lineup, a witness choosing the lookalike does not automatically create a second, independent basis for suspicion. The computer’s first choice helped determine whom the witness was asked to consider.
A global rollout, not a single legal model

The expansion is international, but the legal approaches are not interchangeable. Western Australia’s published operational policy expressly says a live alert alone provides no grounds to arrest or exercise police powers. The force reports 79 arrests during its first three months of trial operations, alongside the eight incorrect alerts, and describes officer confirmation before intervention. Those are relevant safeguards. They do not answer every question about what happens to a disputed identification later in an investigation.


In India, the government described its June 2026 CrPI application as combining facial recognition, iris and DNA matching to aid offender identification and CCTV-based investigations. That announcement documents technological expansion; it does not establish the same kinds of wrongful-arrest cases seen in the United States. But it gives another example of why evidentiary standards matter at the point where a technical match becomes a police decision.

In Europe, Article 5 of the European Union’s AI Act narrowly limits real-time remote biometric identification in publicly accessible spaces for law enforcement, allowing specified exceptions subject to conditions and oversight. The law also bars decisions with adverse legal effects based solely on the output of an authorized real-time system. That provision does not mean every retrospective photo search is forbidden or that EU law supplies an identical arrest standard for every member state.

Britain is pursuing its own framework rather than importing the EU model. Its Home Office launched a consultation on police use of facial recognition and related biometrics, while an October 2026 Home Office letter asks its scientific ethics advisory committee to examine the implementation of facial recognition. These differing approaches point toward one portable safeguard: police must be able to distinguish an algorithmic lead from the independent grounds required for the next consequential action.
The independent-evidence rule

The most practical model comes from Detroit. Robert Williams was wrongfully arrested in 2020 following an incorrect facial-recognition lead. A 2024 federal-court settlement established more demanding police procedures. A lineup cannot be based only on the machine-generated lead, and that lead plus a resulting lineup identification cannot by themselves justify seeking an arrest warrant. Detectives must document independent, reliable evidence and obtain supervisory review.

Every jurisdiction has its own law of detention and arrest. The transferable test is simpler than a universal legal formula: If the algorithm’s suggested name were removed from the case file, what reliable evidence would still connect this person to the suspected offense? A credible alibi, conflicting physical description or evidence that places someone elsewhere should be investigated before arrest, not dismissed because a screen supplied a convenient name.

The record should also show where the name came from. Investigators and prosecutors need the original probe image, the candidate list where legally obtainable, relevant limitations and search conditions, the reviewing officer’s judgment, and each subsequent step built on the original lead. Courts and defense counsel need legally appropriate access to information that may support a challenge. A weak identification does not become stronger merely because it is copied between police databases or agencies.

Evaluation must follow the person as well as the algorithm. Police should report false alerts relative to all scans, incorrect alerts relative to all alerts, interventions that follow, and consequential errors discovered afterward. These denominators answer different questions. Where reliable data permit, agencies should assess differences across populations and review whether safeguards actually prevented wrongful stops, detentions or arrests. An impressive headline error rate is not an adequate substitute for a defensible case.


None of this requires discarding a potentially useful investigative tool. It requires preserving the work that makes an investigation legitimate. For someone subject to a valid warrant, the immediate task may be confirming identity. For someone newly suggested as a suspect, resemblance must remain a lead until evidence independent of that resemblance supports the action that follows.

A computer may help police find a face. It cannot, by itself, tell a court why that person should lose their freedom. The decisive safeguard is what remains after the suggested match is set aside.

About Burak Oktenli
Burak Oktenli holds an MBA and a Master of Professional Studies in Applied Intelligence from Georgetown University. His research addresses the governance of authority in autonomous and AI-enabled systems, and his writing has appeared at the Modern War Institute at West Point, RUSI, RealClearDefense, RealClearMarkets, and Geopolitical Monitor. He is the author of Authority Architectures for Autonomous Systems, a ten-volume series on how authority in autonomous systems is delegated, monitored and recovered, at authority-architecture.me.
View all posts by Burak Oktenli →