Dell Cameron and Dhruv Mehrotra collected the JavaScript that Flock Safety sends to a police officer's browser before the officer logs in, and rebuilt the interface from it. That is a clever piece of reporting. It is also the only way anyone outside the company could learn what the software does, which is the part worth sitting with.

The tool does what the marketing says. An officer types a plain language description of a person ("person wearing scrubs" is the example Flock itself offers), draws a shape on a city map, and the cameras inside it run a continuous automated search for anyone matching. A second feature, the People Detection Alert, takes no description at all. The officer boxes a doorway or a stretch of pavement and gets a notification when a person is inside it.

Before the search runs, an AI model on Flock's servers scores the officer's description against eight categories of sensitive content and returns one of three verdicts. Allow. Block. Warn. The model reaches the verdict using a confidence score. The department never sees that score. It gets a canned outcome.

I have spent a decade in rooms where somebody proposes to deploy a model against patients, and the first question is always the same. What are its operating characteristics, in whom, measured how. Flock has shipped a classifier into thousands of police departments and answered none of it.

Here is where the clinical training is useful, because the mistake in play is the oldest one in screening.

A cordoned watch list is a screening test. It runs continuously against an unselected population, almost none of whom are the person being sought. The performance that matters is not how often the model is right. It is how often an alert is right, and that depends on how rare the target is.

Take a downtown cordon with 20,000 people passing through in a day, exactly one of whom is the person in the description. Give the matcher a sensitivity of 95 percent, which is generous for moving video in variable light. Then vary the specificity.

A 99 percent specific classifier sounds excellent in a demo and generates 200 wrong people a day in this scenario. Getting to a coin flip requires four nines, which no deployed vision system has demonstrated on live street video against a free text description. Kate Ruane's observation that no content moderation done at scale is necessarily accurate, and that moving video goes wrong more often, is the same statement arrived at from the other direction.

Every one of those wrong people is a stop, a run, a name in a report, occasionally a use of force. In a clinic a false positive gets a confirmatory test. On a sidewalk it gets a police contact, and the confirmatory test is the officer's judgment under adrenaline.

The screening literature settled this argument decades ago. We do not deploy a test into a low prevalence population without knowing its operating characteristics, because the false positives will outnumber the true ones and the harms of the workup land on healthy people. Flock has built a screening program, sold it to police departments, and skipped the part where anyone computes the number.

None of this is exotic. It is the checklist any hospital AI committee would run before a model touched a patient, and it maps onto this product without much translation.

This reporting does not stand alone, and the two weeks after it ran matter for how you read it.

The Washington Post has identified at least 50 officers charged with or accused of misusing license plate readers, 46 of those cases involving Flock, and in more than half the alleged target was a wife, a girlfriend, an ex partner, an ex partner's new boyfriend, or a woman the officer wanted to meet. A Texas deputy searched more than 83,000 cameras for a woman who had an abortion, first described publicly as a welfare check. The Illinois Secretary of State audited Flock and found it had let federal immigration agents reach state camera data in violation of state law. Senator Wyden wrote to the company on 17 September. Two House members had already written to CEO Garrett Langley.

And on 18 September, WIRED reported that Flock is offering voluntary severance to its roughly 1,500 employees as cities cancel contracts. The market is doing the governance the product did not.