Perfect World
← Transcript library

The Controversial Flock Cameras Tracking Every Car — Full Breakdown

Dave's Garage228K viewsJul 13, 202622:504,035 words
▶ Watch on YouTube⬇ Download transcript (.txt)

Summary

A neutral AI summary of this video is queued and will appear here.

Full transcript

There's a camera sitting beside a road somewhere near you right now that doesn't care if you're speeding, doesn't write tickets, and might not even be recording actual video. But the moment your car rolls past, it can read your license plate, guess the make, model, color, and body style, log the exact time, location, and direction, fire that data up over the cellular network, and make the whole event searchable in just seconds. One photo, not very exciting. But link thousands of those cameras into one database, and each photo becomes a breadcrumb. Connect enough breadcrumbs and you've got an interesting trail. And that in a nutshell is the promise and the controversy of flock safety cameras. Police departments say these systems help recover stolen vehicles, find missing people, catch getaway cars, and turn a vague gray SUV witness description into real investigative leads. Privacy advocates counter that the same network quietly builds a searchable record of where everyday people drive without asking permission, without probable cause, and without a warrant. Today, I'm not here to cheer for one side or dunk on the other. Instead, I'm going to tear the technology apart conceptually, following the bits from the roadside into the cloud, see what exactly the system records, including what it doesn't know, and examine the real benefits, the real risks, the safeguards, the loopholes, and the tough legal questions. Now, I fancy myself something of a law-abiding libertarian, which means I guess I think there should be as few laws as are necessary, but people should at least follow the ones that we have. And that's part of the problem with Flock cameras because, as you'll see, the camera itself is almost the least interesting part. The best known Flock product is the Falcon automated license plate reader or ALPR. Physically, it is a compact roadside camera that can be installed without trenching in power or network cabling. Many installations use a solar panel for power and an LTE cellular modem for connectivity, which means a camera can be placed at a subdivision entrance, an intersection, a parking lot, or a rural road without waiting for somebody to pull Ethernet through the sewer system. Flock handles much of the installation and operates the system as a subscription service rather than merely selling a camera in a box. That architecture is one reason the system is spread so quickly. Traditional municipal camera projects become miniature public works programs involving poles, permits, power drops, fiber, network closets, and a conference room full of people debating who owns the conduit. But a self-contained solar and cellular node skips much of that. It's essentially an internet of things device with a badge nearby. And if you can get power to it and it's in cell range, that's all it needs. Conceptually, the image processing pipeline works much like any other modern machine vision system. First, the camera needs an exposure that freezes a moving vehicle without turning the plate into a white reflective rectangle. The software then identifies a vehicle and plate region within the frame, corrects for angle and perspective, separates the individual characters, performs optical character recognition, and normalizes the result into a candidate plate number. At the same time, additional computer vision models classify the rest of the vehicle. Is it a sedan, a pickup, an SUV, or what is it? Is it red, gray, blue, or a color best described as rental car beige? What manufacturer and model does it most resemble? Are there visible roof racks, bumper stickers, damage, or other distinguishing characteristics? Flock refers to this collection of characteristics as a vehicle signature, or a vehicle fingerprint. Its current search tools can use plate numbers, partial plates, make, model, color, location, time, and natural language descriptions such as a white sports car with a racing stripe. The company also advertises location searches, convoy analysis, hot lists, custom alerts, and real-time routing tools. Now, it matters a lot to me personally where the characterization is actually done. If it's all happening on the device using a entirely local model, I'd be both impressed and relieved. Impressed because that's a lot of AI juice to squeeze for a little box like this, and relieved because it would then mean that the images of you in your car never leave the box. If the block cannot and does not store them and they can never leave, I could give this thing a pass. But there are just two problems. First, the precise division of labor between the processor inside the camera and the software in Flux Cloud is an implementation detail that is not publicly documented deeply enough for us to draw any reliable conclusions or make even a block diagram. But in so far as I can tell, every detection is forwarded as both an image and a set of extended metadata about that image. So, it's not just recording the fact that your license plate value is spotted on your blue BMW a mile from your house. It's also sending the image this time and every time that it sees you. And that database of images and metadata is where the system stops being merely a camera and starts becoming something more like Google for cars. Suppose a convenience store camera captures a gray SUV leaving right after a robbery. The plate is unreadable, but the vehicle has a roof rack and damage to the left rear bumper. An investigator can search nearby flock cameras for gray SUVs matching those characteristics during the relevant time window. Instead of manually watching 6 hours of footage from 20 cameras, the investigator gets a much smaller set of candidate vehicles. Doesn't prove that any of them was involved and it does produce leads. If the plate number is already known, the search becomes much more direct. Enter the plate and the system can return the points where the participating cameras observe that plate during the retention period. A single camera tells you that the vehicle passed one location at one time. Multiple cameras can show a direction or a route. More cameras produce a more complete route. And this is the semantic fault line running through the entire debate. Flock describes its core ALPR product as taking point-in time images of vehicles on public roads rather than continuously tracking people. And that is technically accurate. There's no transmitter attached to your axle and the Falcon does not follow you home like a tiny airborne probation officer. But critics respond that querying many point in time observations is still a form of tracking. And that is also technically accurate because just as a movie is merely a sequence of still images, a travel history is merely a sequence of timestamp locations. The distinction becomes less meaningful as camera density or image frequency arises. For those of you who took a little sequel in college, think of it this way. The individual records are not where the power lives. The power lives in the join across tables. The license plate reader also operates as a real-time alerting system. Police can subscribe to what are commonly called hot lists. These may include license plates associated with stolen vehicles, wanted suspects, Amber or Silver alerts, missing people, or other active investigations. Flock says it system integrates with criminal justice databases such as the National Crime Information Center. When a matching plate passes a camera, the system can notify officers or analysts immediately. From a software perspective, this is a classic publish and subscribe system. The hot list contains events of interest. The roadside cameras publish vehicle detections. The platform compares each event against the subscriptions and pushes an alert when there's a match. And speed matters, not of the vehicle, but of the report. A stolen vehicle report that's checked manually tomorrow is just an investigative record. But a detection delivered to an officer 30 seconds after the car passes an intersection can become an interception. And an interception can end or prevent a crime. Not in a minority report kind of way, but in their there was actually a cop there when I needed one way. Police departments using these systems describe them as valuable for recovering stolen vehicles, locating missing or endangered people, responding to crimes in progress, corroborating timelines, and identifying vehicles when witnesses can provide only a partial description. Some cities report successful recoveries and arrests, although attributing a change in overall crime rates specifically to cameras is much harder than documenting individual cases in which a detection was actually useful. And that distinction is important. A screwdriver can unquestionably repair a machine, but proving that buying more screwdrivers reduce mechanical failures throughout the city is a much larger statistical problem. Case utility and crime reduction are not the same metric. The camera also does not know who is actually driving. Now, a separate law enforcement data system may connect that plate with the registered owner, but the registered owner might not be behind the wheel. The vehicle might be borrowed or rented or sold without updated records, carrying a stolen plate or displaying a temporary plate printed by a homing jet printer that has begun to reconsider its life choices. Flock says that its core LPR system does not use facial recognition and is designed to collect vehicle information rather than identifying the occupants. Some municipalities also state that their deployments are not connected directly to the DMV or commercial vehicle history databases. But because law enforcement can use other authorized databases, a plate observation can still become associated with the person during an investigation. So calling license plate data either a personally identifiable or completely anonymous misses the nuance. Your plate is not your face, but neither is it a random number with no connection to human activity. It's more like an IP address. By itself, it identifies a device or vehicle with additional records and context that can often be connected to a person. And then there is accuracy. ALPR is an inference system, not a divine revelation engraved on a stone tablet. Rain, glare, darkness, plate frames, dirt, motion blur, unusual fonts, damage, viewing angle, and visually similar characters can all affect the read. A seven can become a two. The letter O can become a zero. One state's plate design can look like it was created specifically to defeat machine vision and possibly human vision as well. And in some parts of Canada, plates aren't even rectangles, they're bears. model. Internally, the recognition model will normally produce a candidate result and some measure of confidence. Setting a strict confidence threshold reduces false matches but can miss legitimate plates that would otherwise be read. Relaxing the threshold catches more possibilities but increases false positives. This is the same precision versus recall trade-off that appears in spam filters, medical screening, antivirus software, and every other detector that must decide whether something looks enough like a target. From a programming standpoint, heruristics are useful, but they are by definition imperfect. And that is why a responsible deployment treats a plate reader alert as a lead requiring confirmation, not as probable cause descending directly from the cloud. Here in Redmond, Washington, for example, they say that alerts must be verified before any action is taken. Other policies likewise require officers to visually confirm the plate and check that the hot list entry remains still valid. This procedural distinction matters enormously. If the computer says a stolen black Range Rover just passed by, an officer should verify that the photograph actually shows the correct plate, that the alert has not expired, and that the vehicle description makes sense. An algorithmic match plus human confirmation can be a useful tool. An unverified match followed by a high-risisk felony style stop can turn one mistaken character into a very frightening encounter or worse. Documented ALPR errors and incorrect hot list entries have resulted in innocent motorists being detained and sometimes they're at gunpoint. Those incidents do not mean every alert is unreliable, but to the extent that an exception disproves the rule, they also demonstrate why the human validation step cannot be treated as optional ceremonial paperwork. Now, we arrive at privacy, where the details matter more than the slogans. Flock's default retention period for license plate data is 30 days. The company says images and metadata are encrypted during transmission and at rest with cloud data protected using AE26 encryption. According to its policy, information remains on the roadside camera only temporarily and is removed after upload and is hard deleted from the cloud when the retention period expires unless another period is required by law or established in the customer agreement. 30 days is not forever and that's a meaningful safeguard. A short rolling window limits how far into the past an ordinary search can actually reach. It also reduces the consequences of a database breach compared with preserving years and years of travel records. But default is not the same thing as universal. Local laws and contracts can require or permit different retention. Now, Flock says that an agency seeking extended retention beyond the legal requirements can obtain up to one year, but the company requires approval from an elected official or governing body. An evidence downloaded from the system for a legitimate cause can be retained under the AY's own separate evidence policies after it disappears from the live flock database. In other words, automatic deletion applies to the platform's rolling data store. It can't make a screenshot already exported into a case file spontaneously evaporate like an old Mission Impossible tape. Once it's in the wild, so to speak, it's a bell that can't be unrung. Flock also says that the customers own their data and that it does not sell customer data and that sharing is opt-in rather than automatic and that agencies control which organizations can receive access. Its security materials describe role-based access controls, audit logging, anomaly monitoring, SOC2 type 2 review, ISO 27001 controls, and tamperresistant search logs. I take some comfort in the notion that every search is logged. That's a substantial improvement over an invisible query box with no accountability. A log can and will record who searched when they searched, the stated purpose of the search, the case number if they are required to provide one, the offense type and which camera networks were actually queried. Flux transparency portal can expose portions of that information publicly, including anonymized usage statistics and sharing relationships. But you need to remember one key thing. A log is a witness, not a guard. Logging a misuse does not prevent it unless somebody reviews the log, recognizes the pattern, and imposes some consequences. Every Windows administrator knows this problem. You can enable exquisite auditing on a server, but if the event log is a digital addict that no human actually visits until after the burglary, it provides forensics rather than prevention. Flock has been adding more preventative controls including optional required case numbers, keyword filters intended to block searches that are prohibited by state law, labels identifying requests originating from federal organizations, and restrictions on federal access to certain state data. The company argues that local elected officials should determine the permitted uses and that the platform should provide the tools to enforce and audit those choices. So, they don't really want to be in the business of making up the rules. They just want to know what they are and follow them. It sounds like there's also evidence that policy controls can fail or prove incomplete. In 2025, an Illinois Secretary of State audit found that US Customs and Border Protection had gained access to Illinois plate reader data in violation of state restrictions. Illinois ordered the axis shut down and Flock paused its federal pilot while adding further controls. The episode can be read in two ways. Critics see proof that promise guard rails did not prevent the prohibited access while supporters can point that the audit trail detection shutdown and subsequent controls as the accountability system actually working correctly after a failure. And both readings contain some truth. Security engineers live in that uncomfortable territory all the time. Another concern is network sharing. A city may own 50 cameras but gain search access to detections from neighboring jurisdictions that have agreed to share. That can be extremely useful when a kidnapping suspect, stolen car, or robbery crew crosses municipal boundaries. Because if old movies have taught me anything, it's that criminals have shown remarkably little respect for the county line. But data sharing also changes the privacy equation. A resident may debate and approve 20 cameras locally only to discover that the effective searchable network is regional, statewide, or even larger. Critics argue that a federation of individually modest systems can become something no single community consciously approved of or even really considered in advance. Mind you, this is also the same architecture that made the internet useful. Independent networks agreed to exchange traffic and the combination became much more powerful than any individual network. The difference is that the internet packets usually belong to willing participants. Cars passing public cameras do not negotiate peering agreements. There's also an important product distinction. The classic Falcon LPR has generally been described as a vehicle focus system taking still images rather than acting like a conventional continuously monitored CCTV camera. But Flock's broader platform includes dedicated video cameras. And in 2025, the company announced optional live and recorded video capability for existing LPR hardware as well as colllocated video products. That means a community evaluating flock cameras needs to ask exactly which hardware, software options, and retention policies are enabled rather than relying on generic descriptions based on the brand name. Asking whether a Flock camera records video is therefore a little bit like asking whether a Windows machine runs SQL Server. Some do, some don't, and the logo on the case doesn't answer it for you. The legal question is evolving for much the same reason. Courts have traditionally held that drivers have little expectation of privacy in a license plate displayed openly on a public road. And even I get that. But the modern question is not merely whether an officer may look at the plate. It's whether the government may automatically collect millions of public observations, retain them, aggregate them, and search them retrospectively. In a 2025 federal case from Kansas, a district court rejected a fourth amendment challenge involving nine flock detections over roughly 4 hours. The court distinguished those limited observations from continuous cell phone or GPS tracking and noted that the 30-day retention period reduced its concerns, but the same decision explicitly warned that a denser and more pervasive network could eventually approach the kind of systemic tracking that raises constitutional problems. That is a remarkably balanced description of the technology itself. The constitutional question may depend not just on what one camera does, but on scale, density, retention, search scope, and how the data is used, which leads to the most useful way to think about flock. The camera is a sensor. The cloud service is an index. The hot list is an alerting rule. The sherry network is a federation. None of these components inherently knows whether the person operating the search is trying to find a kidnapped child, recover a stolen truck, identify a robbery suspect, monitor a protest, investigate immigration status, locate someone seeking medical care, or satisfy a personal curiosity that has no legitimate law enforcement purpose. Because purpose does not live in the lens. Purpose lives in policy, permissions, oversight, and in the person that is allowed to sit at that keyboard. supporters see a machine that observes purely public facts more consistently than a patrol officer ever could. It does not become tired, distracted, or decide that one driver looks a little more suspicious than another. It records vehicles rather than skin color, and it can give officers a timely objective lead when minutes matter. Critics answer that decisions about where the cameras are placed and which plates enter the hot lists, which neighborhoods receive scrutiny, which agencies share data, and which alerts trigger stops are all human decisions. Automating observation can reduce one form of discretion while multiplying the reach of other forms. Supporters note that most detections are never actually searched and disappear automatically after the retention period. Critics reply that innocent people are still recorded first and filtered by suspicion later. Supporters emphasize encryption, audit logs, local ownership, opt-in sharing, and the absence of facial recognition in the core ALPR product. Critics point out that encryption protects data from outsiders, not from an authorized user performing an inappropriate search, and that a vehicle history can reveal sensitive associations without ever identifying a face. Neither side is arguing about imaginary capabilities. They're assigning different weight to the same engineering facts. And perhaps that is why the debate has become so heated. Flock cameras are not a hypothetical future surveillance technology humming inside of a lab somewhere. They are practical, affordable, easily deployed network appliances that solve real investigative problems while creating real governance problems at the same time. A community deciding whether and how to use them therefore has more meaningful questions available than simply asking whether cameras are good or cameras are bad. So what offenses justify a search? Must every search include a valid case number? Are searches for minor offenses allowed? Who may create a custom hot list? How quickly must stale entries be removed? Must every alert be visually confirmed? Which agencies may share that data? Are federal requests handled differently? Is access to sensitive locations restricted? Who reviews the audit log? Are misuse statistics published? Are they public? How long is data retained? Can exports live longer than that? Does the public know where the cameras are and which capabilities are actually enabled on them? As for me, I think the technology is incredibly impressive, as are the potential applications, but the ways in which the cameras can and sometimes are used are, as my kids say, a little sus. So, as usual, the devil is in the details of how the system is implemented. Answering those questions does not eliminate disagreement, but it moves the argument from bumper sticker philosophy into actual systems design. Back when I was working on operating systems at Microsoft, we learned that privilege by itself was not automatically good or bad. Kernel mode is necessary to make the computer work, but because the consequences of error are so much greater, kernel code receives stricter rules. Smaller interfaces, deeper review, and more aggressive auditing. A connected plate reader network deserves the same kind of thinking. The more powerful the query, the narrower the permission should become. The broader the sharing, the more visible the audit should be. The longer the retention, the stronger the justification should be. And the more consequential the alert, the more important the human verification becomes. So, this is not a verdict on the camera. It's simply how engineers handle a system with a large blast radius. A flock camera may help locate a missing child one day and record a thousand completely ordinary errands the next. Both facts can be true. Its usefulness comes from remembering events that would otherwise be forgotten. Its privacy implications come from exactly the same feature. The machine does not resolve that tension for us. It merely puts it into a database. If you have comments or questions about today's episode, please leave them in the video comments. And remember that we actually answer and discuss each week's best topics every Friday on Shop Talk on the Dave's Addict channel. I'll put a link to an episode up here. Please check one out. If you found today's episode interesting or entertaining, remember that I'm mostly in this for the subs and likes. So, I'd be honored if you would consider leaving me one of each before you go today. In the meantime, and in between time, hope to see you next time right here in Dave's Garage. Do it. Do it. Do it.