AI's Flawed Enforcement: The Roseville License Plate Reader Debacle





Unveiling the Hidden Flaws in AI-Driven Public Safety
Persistent Errors Plague Roseville's AI License Plate Readers
In the summer of 2024, an officer from Roseville, California, contacted a resident regarding recurring issues with his vehicle's license plate. The city's artificial intelligence-powered license plate reader system, developed by Flock Safety, had repeatedly misidentified his plate. On at least six occasions, the Flock system alerted the Roseville Police Department, erroneously categorizing his vehicle as stolen or implicated in a felony. Police observations confirmed that the AI software consistently confused the '9' on his license plate with an '8'. Despite the car owner's attempt to remove his license plate cover, the misidentification persisted.
Discrepancy Between Stated Accuracy and Real-World Performance
While Flock asserts that its cameras achieve over 96% accuracy in reading license plate characters under optimal conditions, records from the Roseville Police Department paint a different picture. Data from 2023 and 2024 reveal that Flock generated 1,427 alerts for Roseville police, identifying vehicles as stolen or involved in felonies after passing a city camera. However, an internal police analysis found that in 71% of these alerts, the machine learning software incorrectly interpreted the license plates. Public records, including emails, data, and contracts, detail extensive discussions between Roseville and Flock concerning these persistent errors. The cameras frequently failed to detect vehicles, produced indistinct images, misread license plate characters and state origins, and delivered delayed alerts to law enforcement about potentially suspicious vehicles.
Challenges in Roseville's Unique Camera Deployment
Roseville's specific camera setup, described by Flock as "particularly unique," contributed to the misread issues. The city configured its cameras to only capture the rear of vehicles, aiming to protect personal identifiable information, such as faces. Flock also noted that Roseville's deployment involved older equipment and cameras positioned at greater heights and distances from vehicles than typically advised. Despite these challenges, Flock stated its commitment to supporting customers in maintaining community safety, even under difficult deployment conditions. However, Max Isaacs, Director of Technology Law and Policy at NYU's Policing Project, characterized Roseville's 71% misread rate as "stunning" and beyond acceptable limits. He emphasized that without concrete regulations to address such misreads, this technology poses a significant risk to communities.
The Grave Consequences of AI Misidentifications
The ramifications of an erroneous Flock camera alert can be severe. Previous reports have documented multiple instances across various jurisdictions where misreads, or insufficient verification by officers, led to innocent individuals being stopped at gunpoint or unlawfully detained. In one case in Toledo, Ohio, Brandon Upchurch was attacked by a police dog after a Flock camera misread the '7' on his license plate as a '2', resulting in injuries that caused him to lose his job and home. He later settled a lawsuit for $35,000. Fortunately, in Roseville, none of the incorrect Flock alerts resulted in traffic stops or arrests, as the department mandates officers to independently verify a license plate's status or establish reasonable suspicion before initiating a stop. States like Montana, Virginia, Washington, and Kentucky have implemented laws requiring independent verification of license plate reader alerts, a safeguard currently absent in California.
Legislative Calls for Accountability and Oversight
California Assembly Member Josh Lowenthal expressed deep concern over the Flock errors in Roseville, advocating for immediate legislative action. He announced plans to initiate a statewide review of license plate reader reliability, asserting that flawed data compromises the integrity of the justice system. Lowenthal emphasized that license plate readers should enhance, not jeopardize, the safety of Californians. While Flock claims significant improvements in Roseville's camera performance, the police department disputes this, stating that their data for 2025 and 2026 is incomplete and has not been analyzed to confirm any improvements. The company refused to disclose whether it conducts independent accuracy tests of its cameras. Roseville, which began using Flock cameras in 2021 and has invested approximately $450,000 in Flock products since 2020, continues its contract with the company, citing its utility in solving crimes. The department acknowledges the imperfections of technology and emphasizes its commitment to human oversight, ongoing reporting of issues, and collaboration with Flock for improvement.
Growing Dissatisfaction with Flock's Performance
Flock's cameras have not only misread license plates but have also failed to capture vehicles entirely. Last autumn, Roseville police were investigating a mailbox break-in but a Flock camera missed the suspect's dark gray SUV. Similar failures occurred in early January during a deli break-in investigation and the following day when a gray sedan was observed speeding erratically. These frequent errors and missed detections have caused considerable frustration within the police department. Amanda Jones, a supervisor at the department's real-time crime center, documented numerous instances of missed or misread license plates. In April 2025, Jones reported to Flock employees that 16 out of 17 alerts from the previous week were inaccurate, including "blatant misreads" where the cameras confused dissimilar characters, such as '3' with '2'. A lieutenant copied on the email responded with a terse "Ouch!".
Flock's Efforts to Address Technical Flaws and Past Criticisms
In response to the department's concerns, Flock apologized for the inaccuracies and pledged improvements, proposing solutions like camera repositioning and software updates to rectify blurry plates and overexposed images. At one point, Flock's product director for machine learning suggested that officers request drivers to remove license plate frames that obstructed the camera's vision, a suggestion Roseville declined to implement. In 2024, Flock provided an analysis to Roseville's police department, attributing misreads to vehicles being too far from cameras, obscured plates due to trees or frames, and the software confusing similar characters like 'N' for 'V' or '1' for '4'. This is not the first time Flock's technology has faced scrutiny. In 2021, research firm IPVM independently tested Flock's license plate readers and found that the system frequently misidentified states in about one in ten reads and often misclassified vehicle types and makes. IPVM subsequently reported that Flock blocked their attempts to purchase cameras for further testing. U.S. Senator Alex Padilla of California expressed serious concerns about the inaccuracies and mass surveillance capabilities of such technology, advocating for strong safeguards to protect both public safety and constitutional rights.