A face search can do its job perfectly and still put the wrong person in a cell.
That is the uncomfortable lesson in two cases out of Jacksonville, Florida, collected by Biometric Update on 1 October. In the first, a convicted offender lifted photos from another man’s Facebook account and used them on a dating app to contact a minor. Investigators ran those photos through a face search. The search returned the man in the photos, which was correct. He was arrested, jailed and separated from his children. He had done nothing. The photos were his. The crime was not.
In the second, a Charlotte man named Jalil Richardson was matched from surveillance footage against a fake Georgia ID that someone else had used to sell a stolen vehicle. A warrant followed. Nobody first checked whether he had been in Florida at all.
Neither case is really about algorithm accuracy. In both, the software found a face that resembled the input. The failure came one step later, when a resemblance was treated as an identity.
There is a model of this technology spreading through the industry that starts with a face and works outward. Run the face, pull every image that looks similar, crawl the open web for names, addresses, employers and associates, and assemble a profile. The face becomes the selector, and everything attached to it is presumed to belong to the person in front of the camera.
That model has a name problem as well as a privacy problem. It is increasingly sold as “identity intelligence.” We use that phrase too, and we mean almost the opposite.
For us, identity intelligence starts with a record that already exists, created by a law-enforcement agency through due process: a booking, a felony warrant, a conviction, a sex-offender registration, a missing-person report. The question is never “who could this face be?” The question is narrower and easier to defend. Does this person match a specific, verified record that is relevant to this facility, today?
Start from verified records only. UMbRA, the law-enforcement-sourced identity database, holds 58.4M+ records and draws on nothing else. No social media. No news. No credit files. No scraped photos. If a record did not come from a law-enforcement source, it is not in the comparison set. A stolen Facebook photo cannot become a suspect because Facebook was never searched.
Compare to a record, not to the world. When a comparison is needed, it is one-to-one against the record in question, the way QAPLA one-to-one face comparison works, rather than one-to-many against everything on the internet. A non-match is discarded immediately. No biometric template is kept.
Treat every result as a lead. Nothing we produce is an identification on its own. There is no predictive score and no automated decision about anyone’s liberty.
Between the comparison and any alert, a trained analyst looks at the record and the capture side by side and asks the questions the Jacksonville investigators skipped. Is the record current? Is the warrant still active? Does the context make sense? Is there a reason to doubt this is the same person? If the answer is unclear, the alert does not go out.
That review is not a formality. It is the part of the system that knows a resemblance is not a fact.
Detroit learned this the expensive way. After Robert Williams was wrongly arrested on a face match, the city settled in June 2024 and agreed that police would no longer make arrests based solely on facial recognition results, or on photo lineups that follow straight from a face search. Civil-rights groups called it the strongest such policy in the country. It is also simply good practice, and it matches the facial recognition use policy we recommend to our own law-enforcement partners.
If you are evaluating identity technology for a police department, a casino floor, a hospital or a bank branch, three questions sort the field quickly.
A vendor that cannot answer all three in plain language is selling a face search with a profile attached. The Jacksonville cases show where that leads.
For more on why the record matters more than the comparison, see A match is only as good as the record behind it. For how consent and biometric rules differ by jurisdiction, see our state privacy policy pages.
Sources
Featured photo by Ilya Semenov on Unsplash. Inline photo by MJ Duford on Unsplash.
In the space of three days this month, two local governments walked away from the...
Read articleA stop or an arrest that starts from an identity match fails in one of...
Read articleBy a Biometrica staffer Drug related offenses accounted for almost one-third of all cases reported...
Read article
