The misread problem: What ALPR accuracy actually looks like.
Flock Safety claims 93% read accuracy. An independent police study found 37% of fixed-position ALPR hits are erroneous. The gap between vendor marketing and field performance is where wrongful stops, multi-million-dollar settlements, and class-action liability accumulate.
Three rates, not one.
Public conversation about ALPR accuracy typically asks the wrong question. "How accurate is ALPR?" implies a single number. The honest answer is that there are at least three different accuracy measurements in current use, they measure different things, and the differences matter operationally. Conflating them produces the gap between vendor marketing and field performance that drives wrongful-stop litigation.
Flock Safety, the largest US ALPR operator with cameras deployed in over 5,000 communities, markets a 93% read accuracy rate. This is the per-scan character-recognition accuracy under typical operating conditions, as documented by the vendor itself. By inversion, the vendor's own marketing acknowledges a 7% misread rate on every scan attempted.
An EFF analysis published November 2024 (Gowri Nayar) found that ALPRs misread the state of approximately 1 in 10 plates, not counting other character-level errors on top of that. Recent Flock-specific testing (Rain Intelligence, May 2026) corroborates the roughly 10% state-misidentification rate for Flock cameras specifically.
A police-run study in Vallejo California, published in The Police Chief magazine, found that up to 37% of fixed-position ALPR hits were erroneous. This is the operationally consequential number because hits, not scans, are what drive law-enforcement stops. The 37% includes both misreads and stale-database errors.
The three rates are not contradictory. They measure different stages of the same operational pipeline. The vendor's 93% is the per-scan read rate. EFF's 10% is the per-scan state-misread rate, which is one specific failure mode that compounds on top of the vendor's general number. The Vallejo 37% is the per-hit error rate after the hot-list comparison happens, which incorporates both the read errors and the database-quality errors that produce false positives.
Why the math compounds.
ALPR accuracy is not a single technical property. It is the product of multiple distinct error sources stacking on each other through the operational pipeline. Each source has different mitigation requirements and different liability exposure. Understanding which source is producing a given error is what determines whether the fix is technical, procedural, or structural.
Stacking math at deployment scale
Flock Safety reports approximately 20 billion vehicle scans per month across its deployed fleet. Applying even the vendor's own marketed accuracy rate to that scale produces error numbers that are difficult to absorb intuitively.
Not every misread produces a wrongful stop. Most misreads never match a hot list, so they fall through the pipeline without consequence. The wrongful-stop population is the small subset of misreads that happen to match an active hot-list entry. That is the population the Vallejo 37% number measures. Even at 1.4 billion misreads per month, the operational question is not the raw misread volume; it is what proportion of those misreads convert to hits, and what proportion of those hits convert to stops without independent officer verification.
The Vallejo data suggests that conversion rate is high. The documented wrongful-stop cases in Section IV suggest officer verification before stop is the exception, not the norm, in real-world deployment.
The hit-list problem.
Multiple cases below involve correct ALPR reads against incorrect databases, or against correct data interpreted incorrectly by officers. The technology itself worked; the operational layer underneath it failed. These cases form the population that vendor accuracy improvements do not address, because the cameras are not the source of the error.
Recovered vehicle still on the stolen list. Brian Hofer's rental car had been reported stolen and then recovered. The recovery was never reflected in the hot list. The car continued to flag positive for weeks after the underlying status changed. When the ALPR matched correctly, the response treated him as the driver of a stolen vehicle.
Closed case still active in the system. Multiple documented cases involve hot-list entries that were valid at one point but should have been cleared after the underlying investigation closed. The technology read correctly against a database that should have been updated and was not.
Cross-jurisdictional staleness. Hot lists shared across agencies and states age differently in different jurisdictions. A plate cleared in one state's system may persist in another's. ALPR networks that aggregate across jurisdictions inherit the staleness of the most poorly maintained input.
The hit-list problem is structurally different from the OCR problem because the mitigation runs through governance and data hygiene rather than through better imaging or better algorithms. It is also why "improving ALPR accuracy" can produce no operational improvement in wrongful-stop rates: if the database driving the hit-list comparison is itself stale, a perfectly accurate read still produces a wrongful stop.
The documented wrongful-stop cases.
The cases below are the published, settled or settling, named-plaintiff cases that have produced measurable financial and human cost from ALPR errors. Per recent Rain Intelligence reporting, journalists have documented at least 12 wrongful law-enforcement encounters caused specifically by Flock misreads; the cataloged set below is a floor, not a complete list.
Black mother and four children ages 6 to 17 held at gunpoint and several handcuffed face-down on pavement. The Aurora Police Department's ALPR matched her Colorado SUV plate to a stolen Montana motorcycle plate. The vehicle-type mismatch (SUV versus motorcycle) was visible to any officer; the system provided no contextual flag and the officers did not perform independent verification before the high-risk stop.
Settlement: $1.9 million, February 2024.
Pulled over at gunpoint and forced to her knees after an ALPR misread her plate 5SOW350 as 5SOW750. Dispatch confirmed the suspect plate belonged to a gray GMC truck. Officers ignored that Green was driving a dark burgundy Lexus sedan. The vehicle-type mismatch was again available and unused.
Settlement: $495,000.
The privacy advocate himself, detained at gunpoint on Thanksgiving Day after an ALPR flagged his rental car as stolen. The vehicle had been recovered prior to the stop. The hot list was never updated. The ALPR read correctly; the database was wrong.
Settlement: $49,500.
Detroit police searched the ALPR network for all Dodge Chargers within camera range of a shooting. Robinson's car had been recorded fully two miles away from the crime. Ten officers raided her home, placed her 2-year-old autistic son in a patrol car, and impounded the vehicle for three weeks. Detroit PD spokesperson subsequently confirmed officers "are authorized to use license plate readers in reverse."
Settlement: $35,000, April 2025.
Pulled over at gunpoint on his way to work and handcuffed face-down on suspicion of driving a stolen vehicle after an ALPR misread an H as an M on his plate. No independent verification was performed before the high-risk stop.
Detained at gunpoint with her 12-year-old sister placed in the back of a patrol vehicle after an ALPR mistook a 2 for a 7 on her plate. Resolved without serious injury but with the same operational pattern: high-risk stop initiated on automated alert without verification.
San Diego police arrested Parra (passenger) and Beltran (driver) on attempted-carjacking charges after a Flock ALPR captured a red Alfa Romeo five miles from a Golden Hill carjacking. Officers identified the captured vehicle as the suspect vehicle. Timestamp evidence within the Flock data itself showed the captured Alfa Romeo could not have been the suspect vehicle at the time of the crime. That metadata was not consulted before the arrest. A witness misidentification compounded the error: Parra was wearing a white hoodie when arrested; the suspect was described as wearing gray.
Parra was jailed for 30 days. Civil rights lawsuit filed June 2026 seeking $1.5 million for each plaintiff. The case is notable because the exculpatory data existed inside the ALPR system that produced the false positive.
The pattern across all seven documented cases is consistent. The ALPR alert is treated as actionable intelligence sufficient to justify a high-risk stop or arrest. Independent verification (running the plate manually, observing the vehicle type, checking the make and color against the suspect description, examining the timestamp metadata that ALPR systems themselves produce) does not occur prior to weapons being drawn or charges being filed. The technology's stated purpose is to alert officers to vehicles of interest. The operational reality is that the alert frequently functions as the entire investigation. The Parra case sharpens this point further: the exculpatory data existed within the ALPR system that produced the false positive, and was not consulted before the arrest.
The discrimination dimension.
The wrongful-stop cases cluster non-randomly. EFF's November 2024 analysis emphasized that the documented cases disproportionately affect Black motorists and other marginalized groups. The pattern has two structural drivers.
Deployment density. Fixed-position ALPRs are deployed more heavily in areas where municipalities have higher policing budgets, which correlates in many US contexts with majority-Black and majority-Latino communities being subject to higher per-resident ALPR exposure. The same vehicle traveling the same distance is read more times in some communities than in others; the per-vehicle error opportunity is correspondingly higher.
Action-conversion bias. Even where the raw error rates are comparable, the conversion from automated alert to high-risk stop appears to depend on officer-level discretion. The cases documented above are not random across the demographic distribution; they cluster in patterns that suggest the same automated alert produces different operational responses depending on who is being stopped.
The structural problem this creates is that "improving ALPR accuracy" addresses one input but not the other. Even if the per-scan misread rate dropped to zero, the action-conversion bias would continue to produce disparate outcomes from the same false-positive rate.
From accuracy to liability.
The accuracy question has moved out of the technical-evaluation frame and into the litigation frame. Three developments in 2025 to 2026 have changed the institutional-buyer calculation.
The California class-action wave
Following a landmark California appellate ruling in February 2026, four class-action suits were filed within six weeks against businesses (shopping malls, medical centers, retailers, commercial campuses) that deployed Flock Safety cameras without compliant California-statute privacy disclosures. At least eight additional investigations were actively recruiting plaintiffs as of recent reporting.
The litigation theory: under California's ALPR privacy law, any business that scanned license plates without posting a compliant privacy policy owes every person whose plate was captured at least $2,500 in statutory damages, with no proof of data misuse, breach, or monetary loss required. Simon Property Group (the largest US shopping mall operator) updated its privacy policy in February 2026 to add ALPR language, but the Leonard complaint alleges the disclosure was buried behind three hyperlinks at the bottom of the policy. The complaint calls this "the opposite of conspicuous." The litigation is the first sustained test of whether statutory-damages-based class actions can apply ALPR-specific privacy law to deployers, not just the vendor.
The settlement precedents
The seven cases catalogued in Section IV illustrate the developing financial precedent. Documented settlements total approximately $2.5 million across Gilliam ($1.9 million), Green ($495,000), Hofer ($49,500), and Robinson ($35,000). The Parra and Beltran civil rights lawsuit filed June 2026 seeks $1.5 million per plaintiff ($3 million total demanded, not yet settled), and would push the case-precedent ceiling above the Gilliam high mark if successful. Gilliam alone at $1.9 million already establishes that single-incident liability for ALPR-driven wrongful stops can exceed seven figures. The settlement pattern is the precedent that subsequent plaintiffs and their attorneys will cite when negotiating future cases.
The norfolk Fourth Amendment challenge
The Institute for Justice filed suit against Norfolk Police Department in October 2024 challenging Flock surveillance on Fourth Amendment grounds. The case, Schmidt v. Norfolk, is on appeal to the Fourth Circuit as of early 2026. A ruling that ALPR surveillance constitutes a Fourth Amendment search would produce a structural shift in deployment liability across multiple jurisdictions. The case is being watched as the leading current constitutional test of mass ALPR deployment.
Reading the hub through the accuracy lens.
The data in this supplemental file produces additional observations when applied across the main-series documents. Each cross-reference below identifies where the misread and hit-list error patterns add operational texture to arguments already developed elsewhere in the hub.
Document 02 (Two Stacks) traces the technical and commercial convergence between the Motorola Solutions and Flock Safety stacks. The accuracy data adds a liability dimension to that convergence picture. The California class-action wave is currently Flock-specific because Flock's commercial deployment model brought it into the regulated space (private businesses deploying public-facing surveillance) faster than Motorola's primarily law-enforcement deployment model. The accuracy gap and the liability surface are not yet symmetric across the two stacks, and that asymmetry is worth tracking forward.
Document 03 (Monetization Pressure) identifies vendors operating under structural capital pressure to expand monetization. The accuracy lens raises a forward-looking question for each of those candidates: as their products move from law-enforcement-only deployment to broader commercial deployment, do their accuracy profiles hold up under regulatory scrutiny? The BusPatrol case (Section V of Doc 03) is particularly relevant here. The 40,000+ school-bus camera fleet was designed for stop-arm violation enforcement, not general ALPR, and the accuracy characteristics of the repurposed deployment have not been independently audited. The accuracy question becomes a procurement-due-diligence question for any agency consuming that pivoted data.
Document 04 (Feasibility) develops the threat model from the adversary perspective. The accuracy data here is the legitimate-side counterweight: the same technology that an adversary can deploy against a target also fails against legitimate users. Both dimensions matter for institutional buyers evaluating their full exposure picture.
Document 08 (Defensive Doctrine) emphasizes legal-side defensive moves available to institutions and individuals. The wrongful-stop case patterns documented here connect directly to that framework. Individuals who find themselves on the wrong end of an ALPR misread have specific legal remedies (DPPA private right of action where applicable, civil rights claims under §1983, state-specific ALPR statutes in California and elsewhere) that complement the doctrine's preventive emphasis.
What this means.
For institutional buyers deploying ALPR. The 93% vendor-marketed accuracy rate is not the operational liability rate. The Vallejo 37% hit-error rate is closer to the operational reality, and the California class-action wave establishes that statutory damages at $2,500 per captured plate can attach to deployments that fail compliance disclosure requirements. The accuracy question is no longer separate from the procurement question; due diligence requires both. Vendors offering ALPR deployment should be evaluated on their accuracy auditing practices, their hot-list hygiene processes, and the legal compliance posture of their deployment templates.
For institutional buyers consuming ALPR data. The broker-layer products that consume ALPR feeds inherit the underlying accuracy problems. Plate-to-identity records that drive automated downstream actions (collection workflows, insurance underwriting, fraud detection) are operating on input data with structurally documented error rates. Risk-management frameworks that treat ALPR-sourced data as authoritative are absorbing the misread liability without pricing it.
For individuals. The wrongful-stop cases catalogued in Section IV represent identified victims who pursued and won settlements. The undocumented set is presumably much larger. Individuals whose threat models include wrongful-stop risk should be aware that ALPR alerts in the field are not currently treated as preliminary information requiring verification; they are typically treated as actionable intelligence. The defensive doctrine in Document 08 names the legal mechanisms available to individuals affected by ALPR misreads.
For attorneys. The California class-action wave represents the first sustained statutory-damages-based ALPR litigation in the United States. The theory of liability is portable to other jurisdictions with similar ALPR-specific privacy statutes. The Norfolk Fourth Amendment case may produce a constitutional ruling that opens a broader liability surface. The settlement precedents from Gilliam (1.9 million) through Robinson (35,000) establish the value range for individual wrongful-stop cases.
For policy researchers. The structural finding here is that accuracy improvements at the vendor level do not address the hit-list and reverse-ALPR error sources, which together account for a substantial fraction of documented wrongful stops. Policy responses focused on technology mandates (mandatory accuracy thresholds, audit requirements) will produce limited operational improvement without parallel hot-list governance requirements and verification-before-stop procedural standards. The technology is not the only thing that needs fixing.
This supplemental file applies the accuracy lens to the ecosystem mapped across the main series. The hub now describes the ALPR ecosystem (Atlas), its convergence patterns (Two Stacks), its forward trajectory (Monetization Pressure), its operational accessibility (Feasibility), its consumer-tier edge (Web Tier), its regulatory frame (Legality and NDAA), its defensive doctrine (Defensive Doctrine), the upstream data source (DMV Layer), and the technology's accuracy and liability surface (Accuracy). What is missing now is the empirical question of where the next material development will surface. The ecosystem moves; the hub will move with it.