目錄
A box-shaped camera mounted beside a utility pole at an intersection may look much like ordinary traffic equipment, but it records more than license plates. Flock Safety’s automatic license plate recognition system stores the plate, time, location, and vehicle image. It can also search by body type, make, color, and other visible features. Records obtained by Texas media also mention details such as bumper stickers and dents. Viewed in isolation, one camera is still a local law-enforcement tool. But when different police agencies open their data to other organizations and use Nationwide Lookup across jurisdictions, cameras once scattered across cities can become an extremely large network of vehicle-movement data.
On the evening of August 27, 2026, Texas Governor Greg Abbott ordered state agencies to stop using state funds for Flock cameras. A spokesperson formally confirmed the order the next day. The timing coincided with an investigation that The Texas Tribune was preparing to publish. Since 2023, the Texas Motor Vehicle Crime Prevention Authority had used an auto-insurance fee originally increased to fight catalytic converter theft to invest at least $30 million in expanding the Flock network, helping state and local agencies install at least 3,200 cameras.
This cannot be reduced to “AI cameras were suddenly banned.” Abbott’s order concerns state funding. It does not automatically remove every existing Flock camera in Texas, nor does it prohibit local governments from continuing to use federal or other funding. But it is a useful case for examining a larger governance problem. As a data system originally procured for public safety adds more cameras and broader cross-jurisdictional sharing, the difficult question is no longer merely “Who can install a camera?” It becomes who can search, how far a search reaches, how long data remains, and whether restrictions created in one place actually travel with the data.
Why Did Texas Suddenly Freeze Flock Funding? Start With the Source of the $30 Million
A One-Dollar Auto-Insurance Surcharge Eventually Became at Least 3,200 Flock Cameras
In 2023, the Texas Legislature passed a law that increased a related charge on each auto-insurance policy by one dollar. The purpose was to give the Motor Vehicle Crime Prevention Authority more funding to address the wave of catalytic converter theft occurring at the time. The law received unanimous support in both chambers, but legislators interviewed by The Texas Tribune said the legislative discussion did not mention using the revenue to build a large Flock or similar license plate recognition network.
Three years later, the Authority had used at least 95 grants to help local law-enforcement agencies buy and maintain about 2,000 Flock cameras. There was also a three-year, $15.9 million arrangement for the Texas Department of Public Safety to add nearly 1,200 cameras. In early August, the Authority had separately approved about $3 million for another 583 cameras along Texas toll roads. Altogether, the state-level expansion of at least 3,200 devices was tied to a fee originally increased in the name of fighting vehicle crime.
After Abbott ordered the state-funding pause on August 27, the state spokesperson’s explanation preserved an important boundary: many local Flock cameras are actually funded by the federal government. The governor cut funding for Texas agencies; that does not mean every local Flock system in Texas stopped at the same time.
Local Governments Had Already Started Leaving Before the State
Pflugerville Shut Down 28 Cameras, While Hood County Unplugged and Covered Three
Before the governor acted, local governments in Texas had already begun addressing their Flock contracts. On August 25, the Pflugerville City Council voted unanimously to end its partnership with Flock and stop operating 28 automatic license plate recognition cameras. The city had been negotiating a renewal, but after public records showed that many nonpartner agencies had searched related data through sharing relationships, officials reframed the issue as one of trust and data governance, not merely whether cameras had helped solve cases.
Hood County took a more direct approach. Its Commissioners Court had already decided to terminate the contract, and on August 11 it unanimously ordered the three covered cameras disconnected and covered while awaiting formal removal by Flock. Constable John Shirley, who carried out the order, later posted photographs of the covered cameras.
On the same day, Kendall County voted 5–0 not to renew a contract covering about 55 to 60 cameras around Boerne. These local governments were not rejecting all ALPR technology. Kendall County officials specifically said the core dispute was trust in Flock and data governance, not a belief that no license plate reader should ever be used.
Across the United States, contract departures also accelerated compared with only a few months earlier. Statistics that the advocacy group Secure Justice gave Ars Technica show that 214 cities and counties have stopped using or declined to renew Flock since 2021, with 90 of them concentrated in August 2026. This is not an official national government count, and databases do not define cancellation, suspension, and nonrenewal in exactly the same way. Still, it confirms that the backlash is no longer limited to a few isolated local cases.
What Do Flock Cameras Record? The Issue Is More Than a Plate Number
Vehicle Signature Can Enable Searches Without a Complete License Plate
The core of a conventional ALPR is still converting a license plate image into searchable text, but Flock adds Vehicle Signature. Search criteria listed in its official explanation include body type, make, color, roof rack, rear rack, custom wheels, and other visible features. Investigators may therefore be able to narrow candidates by appearance even without a complete plate number. Texas investigative reporting also says the cameras store details such as bumper stickers and dents that can help identify vehicles.
That is why calling it simply “a license plate camera like a speed camera” understates its use. Flock does not rely on facial recognition to identify drivers, and the company stresses that the cameras focus on vehicles, not faces. But if the same car is photographed repeatedly at different places and times, those records can still be used to reconstruct a route.
Flock currently says its system includes more than 120,000 deployed cameras in 49 U.S. states and more than 7,000 law-enforcement customers. What makes this scale sensitive is not that every customer automatically shares with every other agency. It is that customers can establish sharing relationships, and some agencies can use Nationwide Lookup to search data shared by other participating agencies.
Is Flock a Database That Lets “Any Police Officer Search the Entire Country”?
It Is Not Completely Unrestricted, but the Previous Nationwide Lookup Did Extend Searches Beyond What Many Local Governments Expected
This is the point that needed the most correction in the original draft. Flock’s current official position is that customers own their data and decide whom to share it with, while outside Access Requests require authorization. The Texas Tribune describes Nationwide Lookup as allowing participating Law Enforcement Agencies to search data opened by other participating agencies around the country. In other words, “having a Flock account automatically reveals every camera in the United States” is not accurate.
Yet the practical governance problem cannot be dismissed. The Palo Alto Police Department later explained that when Flock added Nationwide Lookup in late 2023, many users, including the department, did not understand that the change could allow out-of-state or federal agencies to search their camera data. In 2026, Flock also acknowledged that some California Agencies had unintentionally allowed Out-of-state Agencies to access data in 2025 because of system and configuration problems. Because early Logging capabilities were inadequate, the cause can no longer be fully reconstructed in some cases.
The more precise governance question is therefore not whether permission settings exist at all. It is whether sharing settings match the policies local agencies believe they configured, and whether purchasing organizations truly understand when system changes alter the boundaries around their data.
Why Did One Texas Search Reach 6839? It Was Actually 6,809 Networks and 83,345 Cameras
A Johnson County Search Turned Cross-State Data Sharing Into a Concrete Case
On May 9, 2025, a deputy with the Texas Johnson County Sheriff’s Office used Flock to search for a woman who had self-administered a medication abortion. A publicly obtained Audit Log shows that the second search looked back one month and covered 6,809 Flock Networks and 83,345 cameras. The reason field included language about abortion and finding the woman. The search also reached camera data in states including Washington and Illinois.
Sheriff Adam King later said the police were performing a Welfare／Safety Search because they feared the woman was bleeding heavily and needed to be taken to a hospital, not trying to arrest her for an abortion. But case records released later show that police did open a death investigation and discussed possible legal questions with prosecutors. The case therefore cannot be shortened to “Texas police used Flock to arrest a woman for an abortion,” nor can it be left as “a simple welfare check.” A fuller description is that the police’s public account of the purpose and the investigative content reflected in case documents do not completely align, while the search undeniably crossed thousands of networks.
The governance conflict arose because the searched cameras included Illinois. State law restricts using ALPR data to track, investigate, or punish conduct involving reproductive healthcare or immigration in another state. The issue was therefore not only what Texas police searched for, but why data generated in Illinois could enter that out-of-state search.
What Did the 262 Illinois Searches Prove?
Mount Prospect Police Did Not Search 262 Times; Other Agencies Searched Its Data 262 Times for Immigration-Related Reasons
After the Illinois Secretary of State audited the Mount Prospect Police Department’s Flock use, it found that from mid-January through the end of April 2025, other Law Enforcement Agencies had made 262 searches of Mount Prospect ALPR Data using an immigration-related reason. Illinois Officials considered those accesses violations of a state law enacted in 2023. The state eventually identified 46 out-of-state agencies that had conducted searches inconsistent with state requirements and required Flock to disable the relevant Access.
The case clearly illustrates that “a law exists” and “a system technically blocks violations” are different things. Illinois did not lack rules. The opposite was true: state law already required certain interstate access to honor restrictions, but the Flock Sharing／National Lookup configuration at the time did not reliably translate that restriction into a technical block before every search. Illinois and Flock later added filters, blocked out-of-state agencies, and expanded Audit mechanisms.
It is therefore too absolute to say that “local policy is meaningless in the face of a national network.” A more accurate conclusion is that “if local policy is not implemented as a technical control, it may still fail during cross-jurisdictional sharing.”
Why Did Richmond Also Block Federal Agencies From Flock?
The 2025 ATF Incident Shows That Problems Can Also Occur Through an Already Authorized Account
In June 2025, the Richmond Police Department discovered that an Analyst with the Bureau of Alcohol, Tobacco, Firearms and Explosives, or ATF, had been granted Access to the Richmond Flock System and used it for searches related to Immigration Enforcement, violating the Richmond Police Department’s own Operational Standards. Police immediately revoked the Analyst’s permissions and later decided to stop allowing Federal Agencies direct access to Richmond’s License Plate Reader Program.
This case differs slightly from Illinois. It does not prove that Flock completely bypassed Richmond’s Permission. Rather, someone received access they should not have had, or later used granted access for a prohibited purpose, and the problem was not found until an Audit／Alert. Even when Sharing must be configured, governance must still address who can grant access, what purposes permit a search, and whether searches are reviewed before or after execution.
Why Misuse by 50 Officers Undermines the Assumption That “Users Act in Good Faith”
In early August, The Washington Post reviewed police and court records and identified at least 50 law-enforcement personnel who had been charged or accused of improperly using Flock or other License Plate Reader Systems. Forty-six cases involved Flock. In 26 cases, police or prosecutors believed officers used these systems to monitor wives, girlfriends, former partners, a former partner’s new partner, or other private targets.
One case involved former Police Chief Michael Steffman of Braselton, Georgia. Investigative records show that he searched for the license plates of his former girlfriend and her daughter about 600 times. Flock already had Audit Logs and offered an optional Audit Assistance feature that could detect unusually frequent searches for one vehicle and Off-duty Searches. But when The Washington Post reported the issue, only about one-third of law-enforcement customers had enabled it.
The real issue in this controversy is therefore not whether AI recognition is accurate. Even when every search result is correct, a person with a legitimate account can use the system for an illegal or improper purpose. Accuracy cannot solve Insider Misuse. That requires least privilege, purpose restrictions, anomaly detection, and Audits that are actually enforced.
What Remedies Did Flock Announce in August?
Seven-Day Retention Does Not Force Every Existing Customer to Switch to Seven Days
On August 13, Flock CEO Garrett Langley announced a series of Privacy, Accountability, and Security changes. The most visible was shortening the new recommendation and Default for ALPR Data Retention from 30 days to seven, paired with a new Evidence Mode that lets investigators preserve specific records when justified by a case. Flock said its analysis showed that more than 90% of searches without complete license plate information occur within one week of data generation.
But this cannot be described as “Flock changed every retention period from 30 days to seven.” The company simultaneously made clear that existing customers can retain their previously configured Retention Period. Customers adopting the seven-day recommendation can pair it with the new Evidence Mode. This is a change to the Default／Recommended Policy, not a platform-wide forced reduction of every historical setting.
Flock also announced Offense Filtering, allowing cities to decide which purposes are permitted when they share data with other agencies based on Crime／Offense Type. A city might allow searches for stolen cars, missing people, or Violent Crimes while blocking Immigration Enforcement. All Law Enforcement Customers must also enable Audit Assistance by the end of 2026. When the system detects conduct that meets anomaly criteria, it can proactively suspend the user’s Search Access pending Administrator Review.
Case Code, previously optional, will also become mandatory for all Law Enforcement Searches by year-end. Emergency Searches can be exempted but will be flagged for administrator review. MFA became mandatory for Flock login in early August as well.
Are the New Measures Enough? The Real Difference Is Between “Default” and “Mandatory”
Not all of Flock’s changes can be dismissed as cosmetic. Mandatory Audit Assistance, Proactive Lockout, Mandatory Case Codes, and shorter Default Retention are stricter than the previous approach, which left many controls as choices for local customers. The Washington Post reported that only about one-third of law-enforcement customers used Audit Assistance while it was voluntary. Making it Mandatory is itself a substantive governance change.
But critics’ concerns remain. After an Audit Tool flags anomalous conduct, many responses still require further investigation by an Agency Administrator. Offense Filtering also requires local agencies to configure their own policies in the system. Existing customers’ Retention Periods do not automatically become seven days because of the new announcement. Flock has begun turning some Guardrails from features customers may use into features customers must use, but the purchasing agency’s own Policy, Configuration, and Enforcement still determine a substantial part of the outcome.
The governance question worth preserving is not merely whether a vendor offers privacy features. It is what those features default to, whether they can be disabled, who may change them, and whether the system automatically blocks an anomaly or merely records it.
What Is the Difference Between Seven and 30 Days of Retention? It Is Not Simply “One-Quarter the Harm”
The shorter the retention period, the less historical data is normally available for retrospective searches if an account is abused or a system is breached. Seven days can therefore shrink the exposure window substantially compared with 30. But it cannot be described as guaranteeing one-quarter of the harm. Actual risk also depends on how much data is collected each day, whether it has already been exported elsewhere, whether Evidence Mode preserves specific records, and when an incident occurs.
ALPR Retention Rules also vary widely among U.S. states. By 2026, at least more than ten states had explicit limits. Virginia and Washington use 21 days, Maine 21 days, Minnesota 60 days, while New Hampshire’s legal maximum is only three minutes unless the data meets a specific law-enforcement purpose. Retaining large quantities of license plate data for 30 days is therefore not a technical necessity. It is a policy choice balancing law-enforcement needs.
What Does This Actually Teach Us About AI Data Governance?
Separate the Scope of Data Collection From the Scope of Later Search
When many AI tools are introduced, the easiest first layer to check is which Permissions they currently have for Gmail, Drive, Notion, and Slack. The easier layer to miss is the second: after data enters the system, who else can search it, which Agents can continue using it, and whether connecting a new Connector makes previously separate data mutually visible.
This is where Flock is most instructive. One city buying a few dozen ALPRs and dozens of cities enabling Search Access to one another may involve the same cameras on the surface, but they have entirely different governance characteristics. The same is true of AI Agents. An Agent that can read only one Project Database should not be treated as the same risk as one with simultaneous Access to Mail, Calendar, Drive, Slack, and a CRM simply because both come from the same company.
Treat Retention Policy as a Product Setting, Just Like Permission
AI SaaS products often place data retention inside Privacy, Security, or Enterprise Settings, but Retention determines how much historical data remains recoverable after an error. Documents that truly require long-term preservation do not necessarily need the same retention period as Prompts, Meeting Transcripts, and Search Logs that a model handles temporarily.
When adopting an AI tool, ask not only whether data is used to train a model, but also how long it is actually stored, whether the period can be shortened, how long deletion takes to become final, and whether connected third-party services have their own Retention Rules. Training Policy and Retention Policy are different questions.
An Audit Log Only Shows What Happened Afterward; Sensitive Actions Are Better Restricted Before Execution
Flock already had an Audit Log, and many misuse cases were eventually discovered through Logs. But this did not stop the first improper search. The significance of August’s Proactive Lockout, Mandatory Case Codes, and Offense Filtering is that they begin moving some controls before execution instead of leaving everything for later investigation.
The same principle applies directly to AI Agents. For high-impact actions such as deleting data, sending Email, publishing externally, modifying a production Database, or exporting a customer list, an Audit Log makes it easier to understand what happened after an error. Reducing the error itself requires Human Confirmation, Write Permission, Scope Limits, or another control that actually blocks the action.
Taiwan Also Uses AI License Plate Recognition, but It Cannot Be Directly Equated With Flock’s Nationwide Commercial Sharing Network
Taiwanese police do use license plate recognition and vehicle-route analysis at scale. Public information from the Taoyuan Police Department, for example, says its surveillance system integrates license plate recognition, vehicle-route searches, model and color analysis, real-time alerts, and other intelligent video features. New Taipei City also has more than 6,000 license plate recognition devices that support historical footage and vehicle-route searches.
However, currently available public information does not support the claim that Taiwan has a system identical to Flock in which one private vendor opens local license plate data from across the island as a nationwide commercial search network. The two cannot be directly equated. Taiwan’s police systems differ in operators, database architecture, access permissions, and legal environment.
Retention and Access Governance are more useful points of comparison. A public New Taipei Police FAQ, for example, says ordinary surveillance video is retained for one month, license plate recognition video files for one month, and still-image files for six months. These concrete periods are more useful to discuss than simply asking whether Taiwan has Flock: why each system needs to keep data for that long, who may search it, whether searches are logged, and whether different purposes receive different permissions.
Four Data-Governance Checks to Make When Using AI Tools
The first is Access Scope. List the AI Agent, Notion Connector, Google Drive Integration, CRM Assistant, and other services you use. Check whether each can read only one folder or Database or inherits everything the account can already see. If the service supports a narrower Scope, do not open the entire Workspace merely for convenience.
The second is Retention. Not every service offers a user-selectable seven- or 30-day setting, so “reduce everything to the minimum” cannot be applied universally. A more practical approach is to determine how long data is retained, whether the period can be shortened, which records cannot be shortened for legal or product reasons, and whether that period fits the actual work.
The third is the Sharing／Connector Boundary. Data that originally existed only in Notion may suddenly become searchable more broadly after connecting Slack, Drive, a CRM, and an External Agent. Each new Connector should therefore trigger another review of who can read the data. This should not be checked only once when the AI tool is first adopted.
The fourth is Audit and Preventive Control. A service with an Audit Log can at least support an investigation after an incident, but sensitive operations should also be checked for Admin Approval, Human Confirmation, Read-only Access, Usage Alerts, or Automatic Lockout. Not every AI SaaS offers these capabilities. If none exist, that is itself an important signal when deciding whether the tool is suitable for sensitive data.
What Texas Really Froze Was More Than a Camera Budget
Greg Abbott’s order did not make Flock disappear from Texas or remove all 3,200 state-supported devices overnight. The state cut future state-level funding. What happens next to local-government programs, federal funding, and existing contracts must be assessed separately.
The deeper issue is how a system that began by “helping police find stolen cars and vehicles linked to suspects” developed into data infrastructure capable of searching large quantities of vehicle records across cities and states. Local governments that had set procurement and sharing rules only began rechecking what they had actually agreed to after Audit Logs, court cases, and media investigations made specific uses visible.
Flock’s new August Guardrails also demonstrate one point: Retention, Sharing, Case Codes, Abuse Detection, and Lockout are not incidental details. They are product design. While an AI or data system is small, these options are easy to treat as settings administrators can handle later. Once the data network expands, the same Default may directly determine whether one search reaches dozens of cameras or tens of thousands.
Ordinary AI users do not need to treat a Notion Connector and a nationwide police-surveillance network as the same category of risk. But the governance logic is similar: know what the tool can see, how long data remains, whether it can be shared outward, and whether anomalies are blocked in advance or only recorded afterward. If these questions are not answered at adoption, greater convenience usually makes later separation more expensive.
Frequently Asked Questions
No. Flock’s Vehicle Signature can search by Body Type, Make, Color, and some visible body features. Texas investigative reporting also lists bumper stickers and dents, so vehicle characteristics may narrow results even without a complete plate number. Flock says, however, that its system focuses on vehicles rather than Facial Recognition.
That is too broad. Flock customers control Sharing, and whether law enforcement can find another region’s data through Nationwide Lookup depends on whether the relevant agencies participate and share. However, local police have sometimes not realized their data had entered broader interstate searches, and Sharing／Classification settings have allowed access by out-of-state agencies. The main issues are sharing configurations, notice of product changes, and whether local Policy is actually enforced.
Not every existing customer is necessarily on seven days. On August 13, 2026, Flock shortened its new recommendation and Default Retention from 30 days to seven, while also saying existing customers may keep their configured Retention Period. Customers adopting seven days can preserve records needed for a specific case through Evidence Mode. A city’s actual retention therefore still depends on local Policy and configuration.
Public Audit Records support it. On May 9, 2025, a Texas Johnson County Sheriff’s Office Flock Search for a woman who had self-administered a medication abortion looked back one month and touched 6,809 Networks and 83,345 Cameras. Police later said the purpose was to check her safety, but subsequent case documents show the incident was also handled as a death investigation, so both accounts need to be preserved.
Taiwan has large police license plate recognition and vehicle-route analysis systems. Taoyuan and New Taipei publicly describe using AI plate recognition, vehicle make and color analysis, and route searches. But there is not enough public evidence that Taiwan has a commercial data network identical to Flock, built by one private vendor and searchable across local agencies nationwide, so the two cannot be directly equated.
The Texas Tribune investigation says the Motor Vehicle Crime Prevention Authority has committed at least $30 million in related funds since 2023 to help state and local agencies build at least 3,200 Flock Cameras. About 2,000 came through at least 95 local grants, while nearly 1,200 were installed by the Texas Department of Public Safety through a three-year, $15.9 million arrangement.
No. On August 27, 2026, Greg Abbott ordered state agencies to stop using Texas State Funding for Flock Cameras. It was not a complete ban and did not require every local government to remove existing equipment immediately. Many local cameras use Federal Funding or local budgets, so continued use depends on each agency’s contracts and policies.