Roseville Review Finds High Flock Plate Alert Error Rate

A Flock spokesperson told Gizmodo that Roseville had a "particularly unique deployment" using older hardware and non-standard installation locations. The spokesperson attributed the lower performance to cameras being farther from vehicles and mounted higher than Flock typically recommends.

The spokesperson also told Gizmodo that Flock had confirmed with Roseville that camera performance had improved substantially after the period covered by the data. Gizmodo reported that Roseville Police Department disputed that characterization.

TechSpot reported that Roseville configured the cameras to capture vehicles from the rear in an effort to avoid recording drivers' faces. According to that report, Flock said this configuration reduces accuracy. The available reporting does not establish a quantified contribution from camera angle, mounting height, image quality, plate covers, or any other individual factor to Roseville's 71% error rate.

Flock's system captures vehicle imagery and applies machine learning to identify plates and vehicle attributes such as color, make, and model, according to Business Insider. It then generates alerts when a read matches records associated with stolen vehicles or crimes. TechSpot reported that Flock processes about 20 billion vehicle scans per month and supplies cameras to roughly 6,000 US communities.

For public-safety AI systems, this case illustrates why evaluation must measure the operational unit that triggers human action, not only aggregate component accuracy. A character-recognition rate can remain high while a comparatively small number of ambiguous reads produce consequential false matches at scale. Comparable computer-vision deployments commonly require local validation across camera placement, lighting, motion blur, plate formats, and alert-review workflows before users treat automated matches as reliable investigative leads.

The reporting also leaves important audit questions unresolved, including how Roseville defined a misread, whether officers reviewed images before contacting vehicle owners, and how the false-alert rate varied by camera or environment. Those details determine whether an error pattern is localized to deployment conditions or reflects broader model and system-performance limits.