Flock's License Plate Readers Misread 71% of Alerts in Roseville, Records Show

How Roseville Police Found a 71% Misread Rate in Flock’s Alerts

Police in Roseville, California, have documented that Flock Safety’s automated license plate readers misidentified plates in 71% of the alerts it sent the department over two years for stolen vehicles or those linked to felonies. Out of 1,427 alerts in 2023 and 2024, only about 414 were correct. The high error rate, found in internal department analysis and revealed through public records requests by Business Insider, raises fresh concerns about the reliability of widely deployed AI surveillance tools.

The startup, valued at $8.4 billion and operating in more than 6,000 communities, claims that in optimal conditions its cameras read over 96% of license plate characters accurately. But Roseville’s atypical setup — cameras mounted high and far, capturing only rear plates to respect privacy — contributed to the system confusing similar-looking characters like 9 and 8, 3 and 2, or E and F. Flock acknowledged the challenging deployment but said Roseville’s configuration included older hardware and placements outside its typical recommendations.

The real-world harm of such misreads has been severe elsewhere. In a dozen reported instances across the U.S., misreads or lack of officer verification led to innocent people being stopped at gunpoint, detained, or, in one case, mauled by a police dog. In Roseville, no incorrect alert resulted in a stop or arrest because the department required officers to visually confirm a plate was actually stolen before acting.

Both Flock and the Roseville police continue to meet twice a month to address the problems, and the city says it hasn’t analyzed recent data to verify improvement claims. Meanwhile, California Assembly Member Josh Lowenthal called the errors “disturbing” and said they require immediate legislative review, potentially adding momentum to calls for statewide rules on license plate reader reliability.

Behind the Flock Misreads: Deployment, Software Limits, and Eroding Trust

How Flock’s Business Pitch Runs into Reality

Flock Safety markets its platform as a powerful tool to help police catch criminals quickly. But the Roseville data — showing that most felony-related alerts are false — undercuts the company’s proposition that its machine‑learning system can reliably automate suspect identification. For a company that reportedly captures 20 billion vehicle reads per month, a 71% misread rate in alerts from one city implies that the total volume of false positives across its network could be enormous, potentially eroding trust among the law enforcement agencies that are its core customers.

Deployment Trade‑offs: Privacy vs. Performance

Roseville deliberately configured its Flock cameras to capture only the backs of vehicles to avoid collecting faces, a privacy‑conscious choice that the police say the company accommodated. But the resulting higher camera angles and greater distances degraded image quality and increased misreads. This highlights a fundamental tension: maximizing accuracy often requires collecting more data, including potentially identifying information, while privacy guardrails can reduce performance. For agencies weighing both concerns, the Roseville experience suggests that privacy‑minded deployments may need complementary verification processes — something Flock’s own machine learning chief indirectly acknowledged when suggesting officers ask drivers to remove license plate frames.

Regulatory Pressure Is Mounting

Legislators in Montana, Virginia, Washington, and Kentucky have already passed laws mandating that officers independently verify plate hits before stopping a vehicle. California currently has no such requirement, but the Roseville revelations are strengthening calls for action. Assemblymember Lowenthal said he would push for a statewide review of plate‑reader reliability, and Senator Alex Padilla labeled the inaccuracies “deeply alarming.” For Flock, additional state‑level regulation could mean mandated accuracy thresholds or auditing, increasing costs and potentially limiting sales. The company’s past decision to block IPVM, an independent testing firm, from buying its cameras for testing may also invite more scrutiny from legislatures and procurement officials concerned about vendor transparency.

Competitive Landscape and Long‑Term Contract Risk

The errors haven’t yet prompted Roseville to cancel its contract — the city says the cameras still help solve crimes, and the department’s budget for policing is substantial. But the repeated back‑and‑forth over fixes, and the dispute over whether performance has improved, create a fragile relationship. If other cities begin releasing similar internal accuracy data, Flock could face contract losses to competitors like Motorola Solutions’ Vigilant or Genetec’s AutoVu, who could position themselves as more transparent or accurate. Moreover, the settlement paid to a mauled driver in Toledo ($35,000) may be a fraction of potential liability if class‑action or civil rights lawsuits multiply. For a startup aiming for an IPO, unresolved accuracy disputes could weigh on valuation.

What Police Agencies and City Buyers Should Take from Roseville’s Experience

For police departments and cities evaluating or using automatic license plate readers, the Roseville findings offer several concrete lessons:

• Demand deployment‑specific accuracy data. Flock’s 96%‑plus character‑accuracy claim applies to optimal conditions; Roseville’s customized setup yielded a far lower rate. Before signing a contract, ask the vendor to test accuracy using your planned camera angles, heights, and traffic patterns — and to commit to regular field‑verification reports.
• Build independent verification into policy. Roseville avoided stops from false alerts because it required officer confirmation. Agencies without such a rule — especially in California, where no state law mandates it — should adopt a written policy that no traffic stop occurs solely on a plate‑reader hit. That single step can prevent the kind of violent encounters reported elsewhere.
• Audit your own alerts. Roseville only discovered the 71% misread rate by manually reviewing each alert. Make it a standard practice to sample and analyze false‑positive rates quarterly, and share anonymized results with the vendor and city council. The data creates leverage to demand fixes or, if performance doesn’t improve, to renegotiate or exit the contract.
• Watch for legislative shifts. Assemblymember Lowenthal’s call for a statewide reliability review could lead to mandatory accuracy benchmarks or third‑party audits in California. Other states may follow. Procurement contracts should include clauses allowing termination or price adjustments if new regulations impose compliance costs, or if accuracy falls below a negotiated threshold.