When police treat a facial-recognition “hit” as evidence rather than a lead, weak images and suggestive procedures can harden into probable cause, sweeping an innocent person into jail; the Dillon lawsuit shows how that transformation happens step by step, and why policy guardrails—not just better algorithms—determine whether the technology serves justice or subverts it.
The Short Version
- The Dillon case centers on a 93% facial-similarity score from low-quality footage that cascaded into arrest and jail time; charges were later dropped.
- Law enforcement says face-recognition results are not matches and must be corroborated—yet in practice they often anchor identifications and warrant decisions.
- Across the U.S., a growing set of documented wrongful arrests trace to the same pattern: algorithmic suggestions amplified by flawed lineups and thin corroboration.
- The durable fix is procedural: transparency, source documentation, non-suggestive lineups, disclosure to defense, and a bright-line ban on arrests based solely on facial-recognition returns.
What the Dillon lawsuit alleges and what the public record shows
Robert Dillon’s complaint describes a familiar arc. Investigators obtained low-quality surveillance imagery from a Jacksonville Beach McDonald’s investigation and asked a regional facial-recognition system to propose candidates. The software returned Dillon with a reported 93% similarity score; that number then featured in downstream decisions—structuring a photo lineup and informing probable cause—culminating in Dillon’s arrest and months in jail before prosecutors dropped the case. The American Civil Liberties Union filed suit on his behalf, arguing officers overstated what the score meant and concealed exculpatory information; the complaint frames the 93% as “similarity,” not a quantified likelihood of guilt.
Law enforcement’s public stance is more cautious in theory: the sheriff’s office involved has said facial-recognition results are never “matches” and require independent investigation to establish probable cause. Reports quoting officials also note that a witness later picked Dillon from a photo array, implying the arrest did not rest solely on the algorithm. Yet that is precisely the pitfall civil-rights litigators flag: once an algorithm proposes a face, subsequent procedures can become subtly suggestive, and “independent” corroboration may in fact be downstream of the initial lead. In Dillon’s case, media accounts indicate the 93% score—generated from screen-captured stills—became the key evidentiary thread that defeated his alibi until charges were dropped.
Mechanism: how a similarity score turns into probable cause
Facial-recognition systems used in policing perform one-to-many searches: they take a probe image and return a ranked list of candidates with similarity scores. A 93% figure of this kind is not a Bayesian posterior of guilt; it is a measure of how closely the pixels in the probe align with an enrolled photo under the system’s feature space. On pristine, frontal photographs, modern systems can be highly accurate; on smeared, off-angle surveillance frames or screen grabs, similarity scores become volatile. If an investigator treats “93% similarity” as if it meant “93% likely,” a lead quietly morphs into evidence. When that number appears in a warrant application or briefing, it can shape a lineup—subtly cueing an administrator or a witness—and can overshadow contradictory facts, especially when time pressure is high.
This is not merely a matter of algorithm quality; it is institutional behavior. Good procedure fences off the machine result from the human identification: document the source image lineage, require a blind administrator for any lineup, avoid including the algorithm’s top candidate unless selection criteria are pre-registered, and never state or imply numeric “confidence” to witnesses or magistrates. Absent those controls, the initial suggestion contaminates the rest of the investigation, creating a veneer of corroboration where all threads trace back to the same seed.
Pattern and precedent: this is not an isolated failure
The Dillon lawsuit sits inside a broader pattern documented by legal scholars, journalists, and civil-liberties groups: face-recognition suggestions entering case files as quasi-forensic proof, despite agency assurances to the contrary. The Georgetown Law Center cataloged instances where face searches served as probable cause in practice, and where defendants were denied a fair chance to challenge the technology’s role. Investigations have shown that, across jurisdictions, agencies frequently use these tools in ways system vendors and independent experts warn against—on poor-quality frames, without rigorous disclosure, and with lineups that reflect, rather than test, the algorithm’s pick.
Tallying the number of known wrongful arrests can distract from the mechanism that produces them, but the directional evidence is consistent: multiple documented arrests involved misidentification catalyzed by facial recognition, with charges later dropped or cases dismissed. The core through-line is procedural: weak images plus algorithmic suggestion, followed by suggestive photo arrays and thin corroboration. That is exactly what Dillon alleges happened to him; the public reporting, while crediting a lineup identification as well, places the 93% figure at the center of the probable-cause narrative.
Where disagreement actually lies
There is no real dispute that a face-recognition search returned Dillon as a high-similarity candidate, that he was arrested, and that charges were dropped. The live disagreements are narrower and more technical: what the 93% score represented, whether officers adequately separated the algorithmic lead from the eyewitness identification, and whether exculpatory information was sidelined. Law enforcement asserts policy says face-recognition is only a lead; Dillon’s complaint argues practice diverged from policy. The weight of independent research suggests such divergences are common unless agencies impose and audit specific procedural constraints rather than rely on general admonitions.
Put bluntly, the gulf is between aspiration and implementation. Saying “never arrest on a face-recognition hit alone” is easy. Preventing the hit from steering every subsequent step—lineup construction, briefing language, warrant affidavits, even the tone of witness interactions—requires design, training, documentation, and oversight. The Dillon record, as assembled through the complaint and reportage, is consistent with that diagnosis.
What durable safeguards look like
Four reforms, repeatedly endorsed by experts, would address the failure modes exposed here. First, bright-line rules: prohibit arrests, searches, or warrants based solely on face-recognition returns; require independent, source-agnostic corroboration that can stand on its own without reference to the algorithm. Second, procedural firewalls: any eyewitness identification must be double-blind and non-suggestive, with the face-recognition candidate not privileged in lineup construction; administrators and witnesses must never be told about scores or “confidence.” Third, disclosure and traceability: defendants, prosecutors, and courts should receive the full chain of custody for probe images, all candidates returned, system settings, and human filtering steps, so downstream actors can test whether corroboration is truly independent. Fourth, governance: internal audits and external reporting should measure real-world adherence to policy, not just the existence of policy on paper.
The bigger picture: accuracy improves, but process decides outcomes
Vendors will continue to tout gains in benchmark accuracy, and some evaluations do show strong performance on controlled datasets. That matters—but it does not dissolve the core risk illustrated by Dillon’s case. Most police probes are not passport photos; they are dim, off-angle, and compressed. Even a highly ranked candidate can be wrong, and once that candidate becomes the gravitational center of an investigation, confirmation pressure does the rest. The decisive variable is not the marketing metric; it is whether institutions treat algorithmic output as a hypothesis to test or a result to confirm. The former is science. The latter is how an innocent person ends up behind bars.
Sources:
foxnews.com, aclu.org, abcnews.com, theguardian.com, wusf.org, cbsnews.com, jaxtoday.org, cltampa.com
