I still replay the audio: officers barking orders while a family froze under the night sky after an automated plate alert. You can almost see how a machine’s mistake turns into a human crisis. That was not an isolated glitch — it was a signal.
I’m going to take you through what happened in Roseville, why Flock Safety’s numbers don’t land where it matters, and what that mismatch means for any city considering these cameras. You’ll get the facts, the fingerprints of responsibility, and the policy question every mayor should answer.
A Roseville patrol car got 1,427 alerts over two years. The camera network sent warnings that were wrong more often than not.
Between 2023 and 2024, Roseville Police Department received 1,427 alerts from Flock Safety cameras. Those alerts were supposed to flag vehicles tied to felonies or reported stolen property. But an internal analysis reported by Business Insider found that in alerts related to stolen cars or felony-connected vehicles, 71% of the plates were misread.
That 71% figure doesn’t describe every scan the cameras make; it targets the moments when the system said, in effect, “this car is connected to a crime.” Those are the moments that put officers on the road with authority and adrenaline — the moments where a false read becomes a false arrest or a gun pointed at an innocent family.
How reliable are Flock’s license plate readers?
Flock points to internal tests showing a >99% capture rate in clear weather, >96% OCR accuracy, and >97% plate-state accuracy — metrics companies use to sell LPR (automatic license plate reader) systems. But those aggregate numbers include every single scan, most of which never trigger an alert. A camera that can log your plate into a database accurately is not the same thing as a system that never falsely accuses someone of a crime.
A Roseville officer flagged bad reads to Flock dozens of times. The city says they repeatedly warned the company.
Roseville officials reported problems repeatedly over roughly four years. According to Business Insider, the department notified Flock dozens of times about bad reads and other issues. The police department says it did not arrest anyone based solely on a Flock alert — a relief, but also a sign of how fragile trust in this tech has become.
Flock’s response to Gizmodo leaned on deployment differences: older hardware, non-standard mounting height, and a request from Roseville to capture only rear plates to avoid faces. Flock also said most misreads occurred in 2025 or earlier and that camera performance has improved — a claim Roseville disputes.
Have false plate reads led to wrongful arrests?
Yes. The Institute for Justice has documented at least 27 cases where drivers were stopped, detained, or arrested after LPR errors tied to Flock alerts. In Arkansas, a family was held at gunpoint after a plate was misread; in San Diego, a man spent nearly a month jailed because a camera linked his vehicle to a violent crime he wasn’t near. These are not hypothetical harms — they are documented civil-rights traumas.
Roseville paid roughly $450,000 for the system. The city paid for alerts that turned out to be wrong.
Roseville’s tab for deploying the cameras is reported at about $450,000 (€414,000). That’s half a million dollars in public funds to buy a surveillance network that, according to the city’s own reporting, produced a majority of false positives on its most consequential alerts.
That $450,000 (€414,000) didn’t just buy hardware; it bought a workflow where police time, public trust, and the risk of rights violations were outsourced to pattern-recognizing software and optical character recognition (OCR) routines.
A police SUV pulled over a vehicle after an alert. The systems feeding officers are simply not neutral.
When a camera issues an alert, it triggers human judgment under pressure. Your officer has a plate, a description, and an instruction that someone behind that wheel might be tied to a crime. Bad data in means risky decisions on the street. The system acts as a broken compass, pointing officers toward a target that may not exist.
That’s why civil liberties groups — the ACLU and the Institute for Justice among them — have been tracking the downstream effects. They document how cross-jurisdictional plate databases allow searches across state lines and how errors can migrate from one report to another. Flock’s database access model raises the stakes: an incorrect tag in one town can influence policing decisions hundreds of miles away.
Can cities stop using Flock cameras if they want to?
Yes. You can choose not to install the cameras, or you can limit how alerts are used. Some jurisdictions have paused deployments or imposed strict rules on retention, querying, and officer reliance. The policy choices are simple: accept the risk and outsource policing decisions to algorithms, or keep those decisions inside accountable human institutions.
A prosecutor received a case based on a camera alert. The legal system is still catching up.
Prosecutors and courts are wrestling with evidence that emerges from LPR systems and the OCR layers that translate images into plate numbers. Defense attorneys are increasingly pointing to misreads and systemic error rates when contesting stops and prosecutions. In many legal battles, the machine’s certainty is treated as evidence of fact — even when that certainty is based on imperfect sensors and proprietary models.
Alerts can behave like a rumor mill, amplifying weak or incorrect signals until they seem indisputable. That amplification makes it harder to unwind mistakes once they enter police workflows and case files.
You’re entitled to ask whether a private company operating a nationwide network of tens of thousands of cameras should carry this much responsibility. Flock Safety says it serves over 120,000 cameras and more than 12,000 customers across 49 states; independent reporting shows the consequences when those cameras misread plates and officers act on the results.
So what do you do about it? Cities can ban LPRs, require stricter accuracy audits tied to deployments, forbid cross-jurisdictional searches, or demand transparency and public dashboards that track false-positive rates. Tech firms can publish real-world error rates, let independent auditors test systems, and build protocols that prevent alerts from automatically triggering escalated responses.
I’m not arguing for technophobia — LPRs can help recover stolen vehicles and connect evidence. But when the technology can, with regularity, misidentify the innocent, you have to ask which institutions will take responsibility: the vendor, the police department, or the public that paid for it?
Do you want law enforcement to rely on automated eyes that make accusatory mistakes nearly three-quarters of the time on the scariest alerts?