Home / Blog / One Photograph of a Screen: FACES and Investigative Decision Records
Case study · 2026-08-19
One Photograph of a Screen: FACES and the Records of Investigative Decisions
On June 10, 2026, the ACLU, the ACLU of Florida, and Hoguet Newman Regal & Kenney filed Dillon v. City of Jacksonville Beach in the U.S. District Court for the Middle District of Florida. The complaint concerns a wrongful arrest built on a facial-recognition lead — and, for our purposes, it is a study in investigative decisions that never became records.
What is verified
Robert Dillon, a 52-year-old commercial crabber from Fort Myers, was arrested at his home on August 26, 2024 by Lee County deputies, on a warrant arising from a November 2, 2023 incident at a Jacksonville Beach McDonald's — an attempted luring of a girl under twelve. Dillon lived more than 300 miles (a roughly five-hour drive) from the scene and was held overnight before the case against him collapsed. The lead came from FACES (Face Analysis Comparison & Examination System), a statewide system operated by the Pinellas County Sheriff's Office, which reportedly returned a "93% match." These facts are supported by the ACLU's press release, the filed complaint, and multiple independent outlets.
The complaint further alleges — and these are untested plaintiff claims — that a detective captured the FACES result by photographing the screen with a cell phone (introducing glare, low resolution, and color distortion); that automated license-plate-reader results showing no trace of Dillon's vehicle in the county were left out of the warrant application, along with witness statements pointing to a regular customer; and that photo-lineup fillers were selected to resemble Dillon. Pinellas County has publicly rejected the failure-to-train allegation as plainly wrong; the Jacksonville agencies declined comment, citing the litigation.
The records at issue
Facial-recognition vendors and agencies describe outputs like these as investigative leads, not identifications. The dispute in Dillon is about everything that happened after the lead: what was included in and excluded from the warrant application, how the lineup was constructed, who decided each step and on what authority. In most agencies those decisions leave no contemporaneous, independently verifiable record — they are reconstructed later from memory, testimony, and whatever documents happen to exist. In VAP's three-tier taxonomy of missing data, decisions that were never captured at all are Tier 1: never observed, and permanently unrecoverable by design. No forensic method recovers a record that was never created.
This case is part of what motivated the Investigative Decision Vocabulary now being drafted as a fragment of the Public Administration Profile (PAP): a standard event vocabulary for lead-generation, corroboration, exclusion, and warrant-preparation decisions, so that such steps can leave tamper-evident, attributable records where an agency chooses to adopt it. The vocabulary defines what a record would look like; it says nothing about any past case in which no record was made.
What VAP would not have done
VAP would not have improved the accuracy of the face match, and it does not prevent, block, or intercept any investigative act. It takes no position on the merits of the pending claims. And it cannot reconstruct decisions that produced no record at the time — that boundary is Tier 1, and it is a hard limit, not a feature gap.
Per the normative non-guarantee clause, conformance “…does not warrant the correctness, fairness, or safety of the underlying AI decisions — only the integrity, completeness (at anchor granularity), and attributability of their records. VAP generates evidence; competent authorities and courts evaluate it.”
Verified against: complaint in Dillon v. City of Jacksonville Beach (M.D. Fla., filed 2026-06-10); ACLU and ACLU of Florida press materials; contemporaneous reporting by CBS News, Ars Technica, Reason, and News4Jax. Allegations remain untested.