Florida Face-Recognition Arrest Exposes Critical Gaps in Police Matching Protocols

Wrongful Arrest Reveals Fundamental Misuse of Face-Recognition Scores

A 52-year-old commercial crabber from Fort Myers, Florida, was arrested and jailed based on a facial-recognition system match that police appear to have fundamentally misunderstood. According to Wired AI, Robert Dillon was taken into custody after FACES—operated by Pinellas County Sheriff’s Office—returned a “93 percent match” against a photograph from a November 2023 child-luring incident in Jacksonville Beach, more than 300 miles away. The critical flaw: Dillon had never visited Jacksonville Beach, and investigators possessed evidence—vehicle license-plate reader data—showing his registered cars were nowhere near the crime scene on the date in question.

The American Civil Liberties Union (ACLU), which filed suit on Dillon’s behalf, frames this not as an isolated algorithmic error but as a systemic misreading of what confidence scores actually represent. The 93 percent figure indicates how visually similar the images are to the algorithm, not the statistical likelihood that both photos depict the same person. According to Wired AI, police routinely conflate these concepts when seeking warrants, a distinction that the judge who authorized Dillon’s arrest apparently never evaluated.

How Evidence Was Excluded from the Warrant

The complaint alleges a cascade of investigatory failures beyond the face-recognition match itself. A Jacksonville Beach officer sent cell-phone photographs from McDonald’s surveillance footage to surrounding agencies in November 2023. A sergeant with the Jacksonville Sheriff’s Office (JSO) ran those images through FACES and flagged Dillon. However, according to Wired AI, when the investigating officer subsequently ran license-plate readers on two vehicles registered to Dillon—covering the dates around the incident—neither vehicle appeared anywhere in Jacksonville Beach. These exculpatory results were omitted from the warrant application submitted six months later, in July 2024.

The complaint also notes that a McDonald’s manager told investigators the suspect was a “regular customer” she had seen multiple times—a detail inconsistent with Dillon’s claim that he had never visited the city.

The Aftermath and FACES’s Scale

Dillon was arrested at his home in front of his wife, held overnight in an unlit van, and forced to pledge his truck’s title to secure bail. The arrest came during peak stone crab season, causing him to fall behind on rent and nearly lose his home. His mugshot remained on the county website for nearly a year until a television reporter prompted its removal. According to Wired AI, FACES holds tens of millions of Florida mugshots and driver’s-license photos, making it one of the longest-running police face-recognition databases in the United States.

Why This Matters

This case exposes a foundational problem in how law enforcement deploys facial-recognition systems: the conflation of algorithmic confidence with evidentiary certainty. If a 93 percent similarity score can justify a warrant despite contradicting exculpatory evidence—vehicle location data, witness descriptions, geographic implausibility—then the system is not failing at recognition; it is failing at warrant standards. The omission of license-plate reader results from the warrant application suggests either investigative negligence or deliberate suppression. Police departments and prosecutors must clarify whether confidence scores trigger judicial review of all available evidence or merely satisfy probable cause on their face. Until that distinction is codified in policy and training, facial-recognition matches will remain a shortcut to arrest rather than an investigative starting point—particularly for defendants without resources to contest them until after detention.