License Plate Misreads and ALPR Accuracy

“Accuracy” is really five different questions

Claims about license plate reader accuracy often mix up measurements that answer different questions. Keeping them apart is the first step in evaluating any number you see.

A system can do well on one of these and poorly on another. The last two aren’t camera accuracy at all. See ALPR hotlists and alerts and vehicle identification vs. driver identification.

Why alert error rates can be high when read accuracy is high

These numbers can look contradictory: a vendor reporting OCR accuracy above 96 percent, and a police department finding most of its alerts were misreads. They measure different things, and both can be true at once.

A camera reads every passing plate, but only a tiny fraction of plates are on a hotlist. In Seattle’s patrol ALPR data for 2021, the system recorded 5,691 hits out of 5,888,081 valid reads, roughly one in a thousand. When true matches are that rare, even an uncommon reading error, repeated across millions of reads, can produce wrong matches in numbers comparable to the true ones. So the share of alerts that are wrong can be large even when the share of reads that are wrong is small. This is a general property of screening rare events, not a finding about any specific system.

The practical consequence is the same either way: an alert needs to be checked before anyone acts on it.

What the published data shows

Here is what we found, with what each source actually measured and its limits. None of these is an independent, audited test of Flock’s per-read accuracy under real conditions.

Flock’s own figures (vendor claims)

• Claim: Flock states its system “achieves over 98% plate capture, over 96% OCR accuracy, and over 97% plate-state accuracy” (blog post, July 23, 2026).
• Method: none published. No sample, test conditions, date range or independent audit is given.
• Also from Flock: its LPR policy says “license plate translation may be incomplete or inaccurate” and that users “should confirm the computer translation prior to taking any action.”

Roseville, California (agency alert data, 2023–2024)

• Finding: of 1,427 alerts involving stolen or felony vehicles, 1,011 (71%) were misreads and 416 (29%) were correct matches. The department reports that none of the misreads resulted in a stop or arrest, which it credits to its verification process.
• Scope: automated alerts only, not manual searches. The department says its cameras capture the rear of vehicles. News reporting identifies the cameras as Flock’s.
• Method: the department doesn’t define “misread” or describe how alerts were classified.
• Dispute: Flock called the figure a “mischaracterization” and attributed it to “a particularly unique deployment,” including “older hardware and non-standard installation locations” with cameras “further from the vehicles and higher than we’d typically recommend.” Roseville disputes Flock’s claim that performance has since improved.

Seattle Office of Inspector General (agency audit, 2021–2022)

• Finding: 976 misreads in 2021 and 948 in 2022, which the report summarizes as about “one misread for every five unverified hits.”
• Definition: a misread is “a valid read that initially registers as a hit but is a false positive.” Examples given include mistaking a 1 for an I or an 8 for a B, and matching characters from a different state.
• Limits: eleven patrol-car ALPR units, not Flock’s fixed cameras. Misreads were logged by officers, and the report says the count is “likely an undercount because officers do not always log all misreads.”

Vallejo, California (practitioner study, 2018)

• Finding: “35 percent of all hits were misreads for the mobile readers, with a similar number (37 percent) for the fixed readers.”
• Limits: reported in a short magazine column about a randomized trial run with the nonprofit BetaGov, not a peer-reviewed paper. No sample size, vendor or definition of “misread” is given.

Older and indirect sources

• AAMVA (2012): a best-practices guide on plate design reports that “a study conducted by one tolling authority established a misread rate of 20%.” The authority isn’t named and no method is given. Toll cameras are a different setting from police ALPR.
• NHTSA (2021): in case studies of six agencies, officers reported “numerous misreads,” mostly with older units. Newer equipment was reported to be much improved, and weather can reduce accuracy. These are interview findings, not measurements.
• Peer-reviewed lab study (2024): researchers applied simulated weather and camera noise to about 20,000 plate images and tested three ALPR software systems (not Flock). “Snow and frost had the most severe effects, often reducing system accuracy to near zero.” Average accuracy on the undistorted images also varied widely between systems. This is a simulation, so it shows that conditions matter, not how any deployed camera performs.

A figure that is not an accuracy rate

The California State Auditor reported that 99.9 percent of the 320 million images stored by the Los Angeles Police Department were of vehicles that weren’t on a hotlist when photographed. That figure describes how much data is collected about vehicles of no investigative interest. It says nothing about whether plates were read correctly, and it shouldn’t be cited as a misread or false-positive rate.

What affects whether a plate is read correctly

Sources consistently point to the same kinds of factors. Flock lists “plate visibility, lighting, camera placement, vehicle speed, weather, plate condition.” Other sources add:

• Similar characters: 1 and I, 8 and B (Seattle OIG). In Green v. San Francisco, a plate ending in 350 was read as 750.
• Darkness and image quality: in Green, it was “late and dark outside, which rendered the ALPR photograph blurry and illegible.”
• Plate design and condition: AAMVA’s guide lists factors such as headlight glare, plate reflectivity, varied plate designs, and plate frames or covers.
• Camera placement: angle, height and distance. Flock cited these in disputing Roseville’s results.
• Weather: reported by NHTSA’s interviewees and shown in the 2024 lab study.
• State and plate type: identical characters can be issued in different states and on different vehicle types.

Human verification: where errors are supposed to be caught

Because reads can be wrong, vendors and agencies tell officers to confirm an alert before acting. Flock’s own policy says so, and Roseville credits its verification process for keeping misreads from leading to stops. When verification is skipped, the consequences can be serious.

A misread: Green v. City and County of San Francisco (9th Cir. 2014)

An ALPR misread Denise Green’s plate (5SOW350 as 5SOW750), a plate associated with a stolen vehicle. The plate the system matched belonged to a gray GMC truck; Green was driving a burgundy Lexus. The court found the sergeant never read her full plate, “even though nothing obscured his ability to do so.” Green was held at gunpoint and handcuffed. The court stated that “an unconfirmed hit on the ALPR does not, alone, form the reasonable suspicion necessary” for the stop, a point the defendants didn’t contest. It found there were triable issues on her civil rights claims and sent the case back to the trial court. Note that this was a ruling on summary judgment, not a final verdict.

A correct read, wrong match: Aurora, Colorado (2020)

In 2020, Aurora police held Brittney Gilliam and four girls at gunpoint after a camera flagged her SUV as stolen. According to the Associated Press, the vehicle that had actually been stolen was a motorcycle in Montana with the same plate number. Gilliam’s car had Colorado plates. Prosecutors called the incident “unacceptable and preventable,” and the city agreed to a $1.9 million settlement in 2024. The characters were apparently read correctly here; the failure was in matching across states and in what happened next.

These are civil cases about police conduct, not criminal cases about evidence. They show where the process can break down: the image, the plate text, the state, the hotlist entry and the human check.

Raising a misread in a criminal case

Whether a misread matters depends on how the ALPR data was used. Questions a defense attorney may look into include:

• Does the plate in the original image match the plate text the system recorded, including the state?
• What were the conditions: time of day, weather, camera angle and distance?
• Did officers visually confirm the plate and verify the alert before acting, and is that documented?
• Does the agency keep misread logs or accuracy data for its cameras, as Seattle and Roseville did?
• If the stop was based on an unverified alert, was it lawful? See traffic stops based on an ALPR alert and motions to suppress.
• Even if the plate read was right, what connects the driver to the vehicle?

The original images and alert records may be deleted on short schedules. See what data ALPR cameras collect for what typically exists.

Limits of the available evidence

• We found no independent, audited field test of Flock’s plate-reading accuracy, and no federal (NIST or FHWA) ALPR accuracy test.
• The agency figures above measure alerts, not reads. They come from different systems, settings and years, and they define “misread” differently or not at all. They shouldn’t be averaged or compared directly.
• Vendor figures are unaudited and come without a published method.
• Accuracy depends heavily on deployment. A result from one city’s cameras doesn’t establish how a camera performed in your case. That’s a question for the specific records.

We’ll update this page as better data is published. See our corrections log for changes.