Inside Flock's AI-Powered Police Search Tool: How It Works and Why Experts Say Its Safeguards Fall Short

Flock Safety has expanded its surveillance capabilities far beyond license plate recognition. The company's latest tools allow police to search for people on camera using written descriptions—a feature now caught up in dozens of abuse cases and allegations of stalking by officers using the system to track exes and romantic interests.

WIRED obtained and examined Flock's software interface, revealing how officers interact with the system and where the company's newly announced safeguards fit into the architecture. What they found raises concerns: while Flock promises to add auditing and require case codes for all searches, experts say the protections may deter casual misuse without actually stopping determined officers.

How the Search Tools Function#

Flock's suite includes multiple ways to hunt for people and vehicles. The core capability converts written descriptions into digital representations and compares them against stored surveillance footage, ranking results by similarity. An officer can search using plain language—for example, "person wearing scrubs"—or filter vehicle searches by specific characteristics like color, make, and model.

One distinctive feature is "Smart Sort," which allows officers to vote on results. When an officer approves or rejects images, the system reshuffles rankings, pulling footage that resembles approved images toward the top. As Flock's tooltip suggests: "Too many results? Vote on a few results to quickly refine."

The most aggressive tool is an AI-powered watchlist. An officer draws a geographic perimeter on a map, and the system continuously scans every camera inside it for anyone matching a written description. Alerts can be adjusted for frequency, and a separate "People Detection Alert" notifies officers simply when anyone appears in a specified area.

The Moderation Problem#

After facing backlash over misuse, Flock announced reforms in August, including shorter data retention periods, mandatory case codes for searches, and automated auditing. However, WIRED's analysis of the underlying code shows that while these guardrails may discourage some abuse and create a record of it, they won't actually block a determined officer from conducting a problematic search.

The company's AI model evaluates officer queries against eight categories of sensitive content—including race, religion, nationality, and political expression—before the search runs. The model returns one of three verdicts: allow, block, or warn. Race, ethnicity, religion, nationality, and "subjective or biased terms" trigger blocks. But critically, political, social, and cultural expression only produces a warning that officers can override by clicking through and leaving a comment.