A reconstructed interface shows map-based person searches, image-ranking feedback, and moderation controls that can block or warn on sensitive queries.
By RuntimeWire Staff · Published
Primary source: WIRED
Why it matters #
WIRED's reconstruction shows where Flock's person-search governance lives: server-side models evaluate what police may search, while the interface blocks some sensitive queries and warns on others. Cities buying the system must evaluate those controls alongside camera coverage, price and investigative utility.
Garrett Langley's Flock Safety lets police draw a boundary on a city map and create an AI watchlist for anyone matching a written description, according to WIRED's September 3 interface reconstruction. Cameras inside the selected area can then run a continuous automated search for a potential match.
The reconstruction adds the officer-facing controls that were missing from RuntimeWire's August 19 coverage of OS Investigate. That earlier report established that Flock's developing investigative software could start with descriptions or behavior instead of a known plate. The new reporting shows how an officer configures a person watchlist, refines ranked images with Smart Sort, and encounters moderation decisions that can block a query or issue a warning.
WIRED built its mockup from files Flock sends to a user's browser before login. Those files show available searches, network calls, expected responses, controls, and warnings. They do not include camera feeds, search results, department settings, or the models and instructions running on Flock's servers.
What the officer sees
A person search accepts a written description such as "person wearing scrubs," along with an area and a period of time. Flock converts the text and camera images into numerical representations, then ranks footage according to how closely the images fit the description, WIRED's reconstruction found. Unlike Flock's vehicle search, which provides menus for color, body type, make, model, and features, the person-search interface relies on free text.
Smart Sort lets an officer approve or reject returned images. Those votes reorder the results, moving footage resembling approved images higher in the ranking. Additional filters, thresholds, and model instructions remain on Flock's servers and outside the interface available to police departments, according to WIRED.
Flock warns officers that text-to-image results may be inaccurate or incomplete and should not serve as the sole basis for police action. The interface places the risk of inaccuracies on the officer conducting the search, WIRED reported. The practical burden also falls on administrators responsible for reviewing flagged activity when the system detects abnormal use.
The watchlist functions sit alongside OS Investigate, a broader product still being tested with a small group of law-enforcement partners. WIRED's analysis of OS Investigate found 69 prepared prompts and code describing 45 underlying tools that could work across camera metadata, case files, dispatch logs, arrest records, ballistics results, and commercial identity databases. Flock has said the product remains in development and its capabilities may change before a broader release.
That background matters here mainly because the reconstructed interface shows how one part of the system reaches an officer's screen. The watchlist does not require a plate or name at the outset. A description, map boundary, and time window can start the search.
Some searches are blocked; others receive warnings
WIRED's interface report says Flock allows some searches, blocks inappropriate or sensitive searches, and warns on searches that may implicate First Amendment rights. WIRED reported that the system evaluates descriptions against several categories of sensitive content, including race and religion, and that political, social, and cultural expression is the category that does not stop a search.
The interaction also sends information back to Flock. When the software flags a search, WIRED reported, it transmits the officer's language, the category and confidence score assigned by the model, and whether the officer continued, dismissed, or disputed the warning. Flock's statement to WIRED did not address how long those records are retained, who inside Flock Safety can access them, or whether they are used to improve the moderation system.
Experts cited by WIRED said outsiders cannot measure the screening model's error rate or determine whether it tends to permit or reject particular kinds of descriptions.
Flock's August policy response addresses some access and auditing problems. In an August 13 safeguards announcement, Langley said Flock would make Audit Assistance mandatory for law-enforcement customers by the end of 2026, automatically suspend accounts showing abnormal activity, and require case codes for searches. Emergency searches can bypass the case-code requirement but will be flagged for review.
Langley also reduced Flock's recommended default retention period for new law-enforcement customers from 30 days to seven days. Existing customers retain their approved settings. Flock said more than one-third of its customers had voluntarily enabled Audit Assistance by August; the interface code reviewed by WIRED showed that departments could still leave the feature off at the time of the analysis.
Misuse gives the interface higher stakes
Those controls are arriving after documented misuse of license-plate-reader databases. The Washington Post identified at least 50 US officers recently charged with or accused of misusing the systems, including 46 cases involving Flock Safety.
The size of Flock's sharing network can extend the reach of each search. RuntimeWire reported on August 29 that several dozen cameras in Alpharetta, Georgia, were connected through sharing relationships involving more than 2,000 organizations. A camera bought by one city can consequently supply evidence to agencies far outside the jurisdiction that approved it.
Flock says its technology is trusted by more than 12,000 communities nationwide.
Flock's combination of owned camera hardware, natural-language search, and cross-source investigative software places it among a wider group of public-safety vendors expanding beyond basic plate lookups. Axon Fusus offers AI search across connected video systems, while Motorola Solutions' Vigilant VehicleManager provides vehicle location, convoy, and associate analysis. Rekor Scout is designed to work with existing IP, traffic, and security cameras, and Genetec AutoVu focuses on forensic vehicle analytics. The WIRED reconstruction distinguishes Flock by showing how person descriptions, continuous watchlists, and officer feedback operate inside the same camera platform; the reporting does not establish equivalent accuracy or governance controls across competitors.
Langley's plate-reader company now watches descriptions
Langley, Matt Feury, and Paige Todd founded the Atlanta-based Flock Safety in 2017 and joined Y Combinator's summer batch that year. Y Combinator's profile still lists Langley as founder and CEO and Feury as founder and CTO, while other public employment materials describe Feury as a co-founder and former CTO who is now a distinguished engineer. Georgia Tech identifies Todd as co-founder and chief people officer. Y Combinator lists Flock Safety with a team of about 1,000.
Flock Safety grew from burglaries in Langley's Atlanta neighborhood. Police told him that grainy security footage offered little investigative value without a readable license plate. Langley, an electrical engineering graduate of Georgia Tech, and Feury assembled a lower-cost camera from off-the-shelf components and machine-learning software. Their prototype helped police make an arrest within 60 days, according to a Georgia Tech account of Flock Safety's founding.
Before Flock Safety, Langley co-founded the mobile live-events company Experience and oversaw engineering, design, product, data science, and customer support. Cox Enterprises acquired Experience for $200 million in 2014, according to his Georgia Tech alumni profile. He later helped launch the car-subscription service Clutch and served as president of Atlanta startup Moment.
Flock Safety has substantial financial backing for its expansion. In March 2025, Flock Safety announced a $275 million financing led by Andreessen Horowitz at a $7.5 billion valuation. Greenoaks, Bedrock, Meritech, Matrix Partners, Sands Capital, Founders Fund, Kleiner Perkins, Tiger Global, and Y Combinator also participated. Flock Safety said annual recurring revenue had exceeded $300 million and was growing 70 percent year over year at the time.
Political resistance has grown with the network. Ars Technica reported in August that a tracker maintained by the anti-surveillance group Secure Justice counted 90 cities and counties ending Flock Safety relationships during that month.
Langley has described OS Investigate as a crime analyst that is "always awake, always watching out for you," according to WIRED's earlier code analysis. The newly reconstructed interface gives that phrase a concrete implementation: an officer can define a place and description, let cameras continue looking, and refine the system's rankings with each approved image.
Langley's original product turned a readable plate into evidence police could act on. The person-watchlist interface begins with the model's interpretation of ordinary language. For city buyers, the procurement question now reaches beyond camera price and plate-recognition performance. They need to decide which person searches should be impossible, when warnings are sufficient, and who reviews search activity before an automated watchlist causes police action.