Flock’s 71% Problem: What the License Plate Accuracy Numbers Actually Mean

Did Flock misread license plates in 71% of police alerts?

Business Insider reported that Flock Safety’s cameras misread license plates in 71% of the alerts sent to police in Roseville, California, in 2023 and 2024. The figure comes from the police department’s own analysis of 1,427 alerts and appears to be accurate. It does not mean the cameras are wrong 71% of the time. It measures the share of alerts that were wrong, which is a different and much rarer category of event. A detection system can read plates correctly the overwhelming majority of the time and still generate mostly false alerts, because genuinely wanted vehicles are so rare. That distinction is missing from nearly all of the coverage, and the vacuum where an explanation should be is a communications problem as much as a technical one.

License plate recognition has become one of the most contested technologies in American policing. Flock Safety, valued at $8.4 billion, says it operates cameras in thousands of communities and processes roughly 20 billion vehicle reads a month. It has also become a lightning rod, with communities from Oak Park, Illinois to Los Angeles ending contracts.

The Roseville story is the sharpest data point yet. It deserves a careful reading, because the honest version is more interesting than the headline.

What Roseville police records show about Flock camera errors

Roseville has used Flock cameras since 2021. Over 2023 and 2024, Flock sent the department 1,427 alerts flagging vehicles as stolen or connected to a felony. A department analysis found that in 71% of those alerts, the software had read the license plate incorrectly.

The records describe specific character confusions. A 9 read as an 8. A 3 read as a 2. A 1 read as a 4. In one case a resident’s plate was misread at least six times, to the point that a dispatch supervisor wrote that it was easier because they had it memorized. In April 2025, a real-time crime center supervisor told Flock that 16 of the 17 plates flagged the previous week were wrong.

Two facts from the same reporting get far less attention. None of the incorrect alerts in Roseville led to a traffic stop or an arrest, because the department requires officers to independently verify a plate before acting. And Flock says Roseville’s deployment is unusual, using older hardware with cameras mounted higher and farther from vehicles than the company recommends, configured to capture only the rear of each car.

Both of those points are relevant, even though neither of them makes a 71% alert error rate acceptable.

Why Flock can be 96% accurate and still wrong in 71% of alerts

Flock publishes three figures: better than 98% plate capture, better than 96% OCR accuracy, and better than 97% plate-state accuracy. Coverage of the Roseville story routinely sets the 96% against the 71% as though one disproves the other. However, they measure different things.

OCR accuracy is a per-character measure. Getting 96% of characters right is not the same as getting plates right. On a seven-character plate, assuming errors are independent, 0.96 to the seventh power is about 0.75. Roughly one plate in four would contain at least one wrong character. Layer in the capture and state figures and the share of reads that are fully correct on both string and state falls to about 71%.

The second effect is larger and less intuitive. It is a base rate problem.

Vehicles that are genuinely on a hotlist are rare. When true positives are rare enough, even a small error rate produces alerts that are mostly false.

Maryland’s fusion center found that about 0.2% of reads produced a hit of any kind, and the ACLU’s analysis found that the overwhelming majority of those were registration or emissions issues rather than serious crime. Minnesota State Patrol data showed 0.05% of reads leading to a citation or arrest.

Work through what that implies. Suppose one vehicle in a thousand passing a camera is genuinely wanted, and suppose that for every 400 reads of a vehicle that is not wanted, the misread string happens to land on a hotlist entry. Across a million reads you would get roughly 1,000 true alerts and roughly 2,500 false ones. That is 71% of alerts wrong, produced by an error affecting a quarter of one percent of vehicles. The arithmetic is illustrative rather than measured, but the mechanism is real and well established outside this field as the false positive paradox.

Misreads also are not random. They cluster on visually similar characters, 0 and O, 8 and B, 2 and 7, 1 and 4. That makes an erroneous read far more likely to collide with a real hotlist entry than random chance would suggest.

The conclusion is uncomfortable for both sides of the argument. A high false alert rate is a predictable property of rare event detection, and every vendor in this category faces it. It is also exactly why an alert cannot be treated as probable cause.

Do duplicate license plates cause false license plate reader alerts?

One explanation circulating is that some states issue the same plate number to more than one vehicle, so the camera reads correctly and simply matches the wrong car.

That issue does exist. NBC 10’s I-Team documented that Rhode Island and Massachusetts issue identical character strings across different plate types, including passenger, trailer, commercial, farm and taxi plates, generating toll bills for drivers who were never there. AAMVA’s License Plate Standard advises against the practice, stating that a plate number should not be reused “regardless of the license plate type.” It matters for policing too. The Illinois State Police LEADS manual confirms that when a plate is run against NCIC, the plate type field “has no bearing on the hot file searching process,” so a trailer plate can trip an alert entered for a passenger car.

However, this still does not explain Roseville. California partitions its plate formats by vehicle class, which makes cross-type collisions structurally unlikely. Most importantly, it is the opposite failure mode. A duplicate plate produces a correct read matched to the wrong vehicle. Roseville documented incorrect reads.

What the LAPD license plate reader audit revealed about stale hotlist data

In July 2026, an LAPD Inspector General audit found that 161 of 498 hotlist alerts were false, about one in three. But this wasn’t Flock’s fault. The cameras had read the plates correctly, but the databases were wrong. Recovered vehicles never cleared from hotlists, theft reports were outdated, and data entry errors were never corrected.

That reframes the whole debate around Flock. The Institute for Justice has catalogued at least 27 cases of innocent drivers stopped at gunpoint or jailed after license plate reader errors, and found that machine error accounts for only about a third of them. The rest trace to stale government data and to officers acting on an alert without verifying it.

There are three failure layers, and coverage tends to collapse them into one:

Roseville is the proof of the third point. A 71% alert error rate produced zero stops and zero arrests, because the department requires verification. Montana, Virginia, Washington and Kentucky require it by statute. California does not.

Where Flock Safety’s crisis communications fell short

None of the above absolves the vendor, and this is where the story becomes a case study in how a technical problem turns into a reputational one.

Flock’s accuracy figures are self-reported, and the company has not published a methodology, a sample size, or test conditions beyond the phrase “optimal deployment conditions.” In separate litigation, Flock has stated that its cameras accurately capture 93 of every 100 plates, a plate-level figure it has not reconciled publicly with the 96% character-level number. Two accuracy claims that appear to conflict, with no published method to explain the difference, invites the least generous interpretation.

The independent testing question compounds it. IPVM has stated publicly that it repeatedly offered to pay full price for Flock hardware and was blocked from purchasing, and it documented a Flock spokesperson making an inaccurate claim about IPVM’s testing to a reporter before retracting it in writing. The ACLU has argued that communities should reject surveillance products whose makers will not permit independent evaluation.

Set aside whether the underlying technology is good. A company that declines independent verification has no credible way to defend its numbers when a customer’s own analysis contradicts them. The 71% headline landed as hard as it did partly because there was nothing authoritative to set against it.

Four things would have changed the trajectory:

1. Publish the methodology before you need it. An accuracy claim without a documented protocol is marketing. Published test conditions, sample composition, and plate-level as well as character-level results give reporters something to weigh.
2. Report the metric that matters to the user. Customers care about alert precision, not character accuracy. Publishing the harder number first denies anyone else the chance to reveal it.
3. Let independent evaluators buy the product. Refusing testing reads as concealment whether or not anything is being concealed, and it removes the third-party validation that is most valuable in a crisis.
4. Resolve a four-year customer complaint or say why you cannot. Roseville raised issues for years and still meets with the company twice a month. Documented, unresolved customer frustration is the raw material of an investigative story, and public records requests make it available to any reporter who asks.

The broader lesson applies well beyond this company. Any organization whose product makes a measurable claim should assume that claim will eventually be audited by a customer, a regulator, or a journalist with a records request. The time to define how performance is measured is before someone else defines it for you.

Crisis communications for public safety technology companies

Red Banyan advises technology companies, including public safety technology providers, facing scrutiny over performance claims, data practices, and contested reporting. These situations turn on getting the technical explanation right and delivering it before someone else frames it. Contact our team for a confidential conversation.