Flock's license plate cameras were wrong 71% of the time in Roseville – and the setup made it worse

An internal review found 1,427 alerts, 1,000+ false positives, and a police supervisor who knew the plate mistake by heart – the reason points to a privacy tradeoff.

Source material: techspot.com

Surveillance License Plate Readers AI Accuracy Privacy Law Enforcement Flock

What did Roseville's review actually find?

An internal review uncovered 1,427 incidents over two years where Roseville's Flock cameras flagged a vehicle as stolen or used in a felony. The cameras misread the plate in 71% of those cases – more than 1,000 false positives. The review also found a police supervisor joked about memorizing one driver's license plate because the system misread it at least six times, mistaking a 9 for an 8 each time. Roseville, a suburb of Sacramento, uses Flock's AI-powered cameras, which photograph passing vehicles and analyze license plates, color, make, and model via machine learning. The false positives never led to a traffic stop or arrest, because officers were required to verify each alert – a step not mandated by California law. Flock attributes the errors to Roseville's atypical camera setup, but Business Insider reports the results could expose flaws in an already controversial technology.

Did any of the false alerts result in traffic stops or arrests?

No. Roseville officers are required to independently verify each Flock alert before acting, and that verification step kept every false positive from becoming a stop. California law does not mandate this verification, but Roseville's policy did. The review found that none of the false positives led to traffic stops, much less arrests. This is not a universal safeguard: because the state does not require independent confirmation, other departments could act on an unverified alert, with potentially serious consequences. The verification process also imposes a cost—officers must spend time checking each hit, many of which turn out to be wrong. So while Roseville avoided the worst outcomes, the system still generates a significant burden of false leads that someone has to sift through, and the risk of error is ever-present without that extra step.

Have other police departments had accuracy problems with Flock?

Yes—the LAPD's experience is a case in point. An internal audit there found that 161 innocent people were pulled over in a two-month span before the department let its Flock contract expire, and the problem was blamed on outdated or inaccurate data from other jurisdictions. That highlights how a shared network can amplify errors: a single misread plate in one town becomes a false alert in another. The scale of the system makes this inevitable, but the key lesson is that without mandatory verification, mistakes are likely to cause real harm. Roseville, by contrast, avoided stops because its officers verified each alert—a safeguard the LAPD lacked. Departments have discretion in handling alerts: they can require independent verification before stopping a vehicle, as Roseville did, which prevented any of its 1,427 false positives from leading to a stop or arrest, or they can treat an alert as probable cause, as the LAPD did, with 161 innocent people pulled over in two months. California does not mandate verification, so the difference in outcomes comes down to departmental policy.

How do Flock cameras work, and what data do they collect?

The cameras photograph every passing vehicle and use machine learning to analyze the license plate, color, make, model, and other details. Images and metadata are uploaded to a cloud database that law enforcement can search using descriptions—for example, 'silver pickup truck'—to track a vehicle's movement across time and location. The system is not limited to vehicles; it can also analyze and track people on foot, though that is not its primary function. Police have searched the database for suspects based on extremely vague descriptions, which can lead to broad and imprecise hits. For anyone whose vehicle or appearance matches a query, that means their movements could become part of an investigation, even if they have done nothing wrong. This is why the system's accuracy and privacy implications matter so much.

What are the biggest privacy concerns with Flock cameras?

The always-on cameras photograph every passerby, and the data is stored in a searchable cloud database that law enforcement can query at will. Because license plates are tied to registered owners, a series of plate reads effectively maps a person's daily routes, errands, and visits. That kind of location tracking raises serious privacy flags, and it has already led to misuse: there are documented cases of officers using the system to stalk ex-partners. Additionally, the persistent surveillance has prompted some communities to deactivate their cameras over the past year, despite the potential for solving crimes. The lack of clear limits on who can search and what the data can be used for makes this a particularly invasive tool, especially when vague queries can pull up innocent people.

What did Flock say about Roseville's high error rate?

Flock blamed Roseville's atypical camera setup, arguing that its technology is less accurate when deployed unconventionally. But the evidence undercuts that excuse: the same vehicle was repeatedly misidentified, and the sheer volume of errors—more than 1,000 false positives over two years—suggests a systematic flaw in the model rather than a one-off glitch. The company's framing shifts responsibility to the city, yet the underlying machine-learning algorithm is Flock's own design. Roseville's decision to capture only rear plates—to avoid recording faces—may have made the task harder, but it also prevented the cameras from capturing other identifying information. That trade-off is worth noting: the city prioritized privacy over accuracy, and Flock responded by disowning the result. The unresolved question is whether the technology would perform better under a standard setup, or whether the error rate is inherent to the system.

How many communities use Flock, and what's the scale of the system?

Flock supplies cameras to approximately 6,000 communities across the US, and its cloud database ingests about 20 billion vehicle scans per month. That massive scale means errors are not isolated—a license plate misread in one town is stored and shared, so a police department on the other side of the country could act on the same bad data. The system's interconnectedness means a single unreliable input can ripple through the network, creating false hits in places far from the original error. With so many jurisdictions feeding the same database, it becomes nearly impossible to trace where an error originated or to ensure that every alert is current. For anyone pulled over based on a Flock alert, the practical implication is that the alert itself proves nothing; it is only a lead that should be confirmed against independent records.

What can Flock cameras do beyond reading plates?

Flock's capabilities extend well beyond license plates. The AI cameras process roughly 20 billion vehicle scans per month, but they also analyze and track people on foot. That means authorities can search the cloud-based database for a person based on vague descriptions—clothing, height, gait—and follow their movements over time. The privacy concerns are not hypothetical: numerous officers have been caught using the system to stalk ex-partners, and police have admitted to running searches for individuals based on extremely broad criteria. These features, combined with the sheer scale of the network, have fueled the fierce debate over state surveillance. Communities have responded by deactivating their cameras since last year, citing worries about a surveillance state. While the company and advocates point to crime-solving benefits, the ability to monitor individuals without a warrant raises questions that accuracy alone cannot answer.

Where this came from. This breakdown is based on source material published at techspot.com. Images above are used with the credits shown beneath each one.