What comes back when an officer types a description is a ranking rather than an identification. Every image a camera captures is stored as a set of numbers describing what it shows, the words in the search box are converted the same way, and a model puts the nearest footage at the top [19]. The filters, thresholds and extra instructions that shape that order stay on Flock's servers [20], and WIRED reports that much of what governs any given search runs there rather than in the client code it was able to read [9].
The August package shortens default retention, requires a case code on every search, and adds automated auditing [8]. All three act on a search that has already run [21], which is the same thing WIRED concluded from the code: the guardrails may discourage some abusive uses, but they log those searches as they happen rather than stopping them [10]. Flock says the package becomes mandatory only by the end of the year [8], so a department going live this quarter is running the expanded capability with the controls still optional [22].
The screen sitting in front of a search is itself a judgment made somewhere else. Flock's system scores the officer's wording against sensitive categories including race and religion, and political, social and cultural expression is the only one that will not stop a search [14]. Tom Bowman, policy counsel with the Center for Democracy and Technology's security surveillance project, told WIRED that this is one of the most protected areas under the First Amendment and "seems like that actually has the least concern from a policing perspective" [15]. Experts told WIRED the model decides, out of the department's sight, which descriptions an officer may search for, and that no screen operating at that scale is accurate [11]. WIRED sent 16 questions and got a statement saying queries are evaluated against content policies, with nothing on how the model was built or what separates one verdict from another [16][17].
The buyer for the watchlist is a detective with a witness description and no plate [2]. The person answering for it is a chief in front of a council that approved plate cameras and has since watched other councils vote them back out [6]. A grid worth drawing before renewal: one axis is whether a control refuses a search or only records it, the other is whether your department can inspect the control or only Flock can. Case codes and audit logs you actually receive sit in the quadrant that is worth paying for. The sensitive-content screen sits in the other one, recording and vendor-side, and its error rate is measurable only inside the company [12]. The Texas deputy who ran more than 83,000 cameras for a woman who had an abortion [5] is the case that decides which quadrant you care about, because a record of that search shows it happened, it does nothing to have kept it from happening.
