Roseville Flock Camera False Alerts: What the 71% Audit Found and What the City Did Next

Roseville Police Department looked at 1,427 alerts its Flock Safety cameras had generated over two years — alerts flagging vehicles as stolen or connected to felonies, the kind that send armed officers after a car.

In 71% of them, the licence plate had been read wrong.

Then the department decided to keep using the cameras.

Both halves of that matter, and most of the coverage only carried the first. The number is the headline, but what a city does after it sees a number like that is the part other towns can actually learn from — and Roseville's answer was neither a cover-up nor a cancellation.

Here is what the audit found, what Flock said in response, why nobody was arrested despite a 71% error rate, what the city decided, and how other cities facing the same evidence have gone the other way.

What the Roseville Audit Actually Found

The analysis, reported by Business Insider working from Roseville Police Department records, covered 1,427 alerts across 2023 and 2024. These were not casual scans. They were the high-stakes hits: vehicles flagged as stolen, or tied to felony investigations.

71% contained a misread plate. Roughly 29% were correct.

A separate independent audit of the department's automated licence plate reader programme was carried out on 18 December 2025, before the figures reached national coverage in August 2026. That eight-month gap between measuring a problem and the public hearing about it is its own story, and a familiar one — we applied the same scepticism to a widely shared cyber threat that nobody had actually verified.

The failures were not subtle edge cases either. In April 2025, a Roseville police supervisor told Flock that 16 of the 17 plates flagged in the preceding week were wrong — including misreads where the system confused characters that do not visually resemble one another.

96% Promised, 29% Delivered

Flock does not market itself vaguely. The company states that under recommended deployment it measures more than 99% capture rate in clear weather and rain, more than 96% OCR accuracy, and more than 97% plate-state accuracy.

Roseville is not an unusual environment. It is a mid-sized Californian city running a product it was sold, on public roads, in ordinary weather.

The distance between a claimed 96% and an observed 29% is not measurement noise or an unlucky sample. It is the difference between a tool that works and one that does not — and it is exactly the pattern of an AI product whose marketed capability was established somewhere other than the place it ends up deployed. We covered the same dynamic when Apple paid $250 million to settle claims it advertised AI features it had not delivered: the legal question there was advertising, and the question here is procurement, but the underlying failure is identical.

Flock's Response — And the Licence Plate Frames

Flock attributed Roseville's results to how the city had installed the system, describing "older hardware and non-standard installation locations that resulted in the cameras being placed further from the vehicles and higher than we'd typically recommend". Roseville had also opted for rear-only plate capture, a deliberate choice made to avoid recording drivers' faces.

That explanation should be reported fairly. Camera geometry genuinely affects optical character recognition, and a rear-only configuration does remove one source of confirmation.

It also does not survive much pressure. Roseville police disputed Flock's claim that performance had significantly improved. And there is a structural problem with the defence: if accuracy collapses from 96% to 29% because cameras sit somewhat higher and further back on older hardware, then the advertised number describes a laboratory, not a street. Every real deployment is somebody's "atypical" deployment.

Then there is the suggestion that tells you more than any statistic.

When Roseville's crime centre supervisor raised the misreads, Flock suggested that drivers should remove their licence plate frames. The department rejected it — a police force cannot ask an entire city to modify its cars so a vendor's software can read them.

One Roseville resident did remove his plate frame after a police suggestion. The misreads continued anyway.

Why Nobody Was Arrested — The Detail That Matters Most

Here is the fact that almost every write-up buried, and it is arguably the most important one in the entire story.

None of the false alerts led to a traffic stop or an arrest.

Roseville requires an officer to independently confirm that a vehicle is actually stolen, or to have separate reasonable suspicion, before acting on a Flock hit. The alert is treated as a lead, not as probable cause.

That single policy absorbed a 71% error rate without producing a wrongful stop in the reviewed period. The technology failed constantly. The procedure around it held.

This is the actual lesson, and it cuts against the easy reading in both directions. The story is not "AI cameras are dangerous" and it is not "the errors did not matter". It is that an unreliable system becomes dangerous exactly at the point where a human stops verifying it. The same principle shows up wherever AI is given authority without a check on it, from a model breaking out of its sandbox to an AI agent compromising a major platform. In the documented cases from other jurisdictions — a family stopped at gunpoint in Sherwood, Arkansas over one misread character, a San Diego man jailed for nearly a month — the difference was not worse cameras. It was the absence of Roseville's verification step. Those cases and the wider pattern are covered in our full breakdown of Flock Safety AI errors.

What Roseville Decided to Do

The department announced it would continue using the licence plate readers despite its own figures.

Stated plainly that sounds indefensible. The reasoning is more coherent than the headline allows: if alerts are treated strictly as investigative leads requiring independent confirmation, a wrong alert costs an officer a few seconds of checking, while a correct one can recover a stolen vehicle. Under that framing, 29% accuracy is not a safety failure — it is a low-precision tool used with appropriate scepticism.

The counter-argument is equally reasonable, and residents have made it. A system that is wrong seven times out of ten is generating a continuous stream of surveillance records on innocent drivers to produce a small number of usable leads. A local DeFlock Roseville petition has pushed for removal outright. And the safeguard is a policy, not a technical property — policies change with a new chief, a new contract, or a bad week.

Both positions follow from the same data. That is what makes this a genuine civic argument rather than a scandal with an obvious answer.

How Other Cities Responded to the Same Evidence

Roseville's choice was not the only one available, and 2026 has produced a natural experiment.

Grass Valley, California halted its use of Flock licence plate cameras amid privacy concerns.

Denver saw city council members move to block a Flock contract, citing concerns about ethics, transparency and credibility — a procurement fight rather than a technical one.

Troy, New York went the audit route, with officials describing an audit conversation as progress while maintaining that more still needs to be done.

Halt, block, audit, or continue with safeguards. Four cities, four answers, one vendor. The variable is not the technology — it is whether anyone measured it before deciding.

Regulation is arriving from the courts as much as from councils, which is the pattern European courts have set with AI image generation and the open-weights policy fight in Washington both illustrate: the rules are being written case by case, after deployment rather than before it.

What to Do If Your Town Has Flock Cameras

The Roseville numbers only exist because somebody inside the department went looking. Most places have not.

Ask for the alert accuracy rate, not the scan volume. Vendors and departments both prefer to report how many plates were read. The number that matters is what share of alerts were correct. If nobody has measured it, that itself is the finding.

Ask what the verification policy is. Specifically: can an officer initiate a stop on a Flock alert alone? Roseville's answer is no, and that is why its story has no victims. A department that answers yes is running the same unreliable system without the thing that made it survivable.

Ask how long the data is retained and who it is shared with. Accuracy is only one axis. A misread plate that is wrong and deleted in thirty days is a different problem from a misread plate that is wrong and shared across a national network indefinitely.

Use the public records route. Alert logs, audit reports and vendor contracts are generally public records in the United States. Roseville's figures came from exactly that kind of material.

If the broader question of what gets collected about you is what brought you here, our piece on what your everyday activity already reveals covers the commercial side of the same problem, and Salt Typhoon is the cautionary case for what happens when surveillance infrastructure built for lawful use is reached by somebody else.

The Short Version

Roseville Police Department measured its own Flock alerts and found 71% of 1,427 stolen-vehicle and felony hits had misread the plate, against a vendor claim of more than 96% accuracy. Flock blamed the city's camera placement, suggested drivers remove their plate frames, and was rebuffed on both counts.

Nobody was wrongly stopped or arrested, because officers there must independently verify a hit before acting. The department is continuing the programme; residents are contesting it; other cities looking at the same evidence have halted, blocked or audited instead.

More of our reporting on this sits in our security coverage. The transferable finding is not about Flock. It is that the accuracy of an AI system and the safety of an AI system are different questions, and the second one is decided by the humans standing next to it.

Frequently Asked Questions