Flock’s new AI changes the question from plate reading to relationship mapping

For years, the public description of Flock Safety was relatively easy to understand: a camera photographs a passing vehicle, reads its number plate and stores information that investigators can search later. That description is no longer enough.

Flock’s newer AI tools can work from descriptions and movement patterns rather than requiring an officer to start with a known registration number. Reporting by WIRED on the company’s OS Investigate system found capabilities for identifying vehicles from descriptions, analysing where they travel and surfacing potential “associates” based on vehicles that repeatedly appear together. The important change is not that a camera has become better at reading a plate. It is that the database can begin to produce relationships.

The deployment is already enormous

PoliceAI News records Flock Safety ALPR as an operational deployment covering more than 5,000 US law enforcement agencies. The tracker records the system as a commercial, subscription-based network rather than a single national police database. As of August 2026, the available evidence puts the network at roughly 120,000 cameras across 49 states, although company and agency figures move as new contracts are signed.

That scale matters because an individual camera is not the real capability. The useful asset is the network created when thousands of cameras produce searchable records of vehicle movements. A department investigating a burglary may only need one plate read. A network can potentially answer a much broader question: which vehicles were repeatedly in the same places, at the same times, before anyone knew there was a connection?

This is why the deployment tracker is useful alongside the technology catalogue. It shows not just that Flock exists, but where it sits in the wider police AI landscape. Flock is currently one of only a small number of individual technologies with multiple documented deployments on the tracker, and the US record is heavily concentrated around operational surveillance systems.

From “find this car” to “find a pattern”

Traditional ALPR investigation usually begins with something concrete. An officer has a registration number, a time, a location or a vehicle description. The system helps determine whether the vehicle was recorded elsewhere.

Pattern-based search changes the starting point. An investigator can potentially begin with a location, a vehicle characteristic or a behaviour and ask the system to identify candidates. That sounds like a small change in the interface. It is not.

Searching for a known plate is an investigative query. Searching for vehicles that repeatedly appear near a location is a form of discovery. The first asks whether a particular hypothesis is supported. The second can generate the hypothesis in the first place.

That distinction matters because the risk of error changes with it. If an officer searches the wrong registration, the mistake is relatively visible. If an AI system produces a list of possible vehicles, associates or witnesses from patterns, an investigator may never know which assumptions produced the list unless the system preserves and exposes that reasoning.

The “associates” problem

The word “associate” deserves particular caution. Two vehicles appearing repeatedly near each other does not establish that their occupants know one another. They might belong to colleagues, neighbours, family members, delivery drivers, taxis or people who simply share the same route.

Yet once a system labels a relationship as potentially meaningful, that label can become a lead. The investigator may then search names, addresses or other records attached to the vehicles. The result is a chain in which an uncertain inference at the beginning can produce increasingly specific personal information at the end.

This is a familiar problem in other forms of predictive and analytical policing: the output can look more authoritative than the evidence that produced it. The technology does not need to make a final accusation to influence an investigation. It only needs to determine which people or vehicles receive attention first.

The US model makes governance harder

Flock is not operated like the UK's National ANPR Service. UK ANPR is a nationally structured police capability with national controls and retention arrangements. Flock is a commercial system purchased by individual US agencies and cities.

PoliceAI News records that distinction because it changes the governance problem. There is no single national customer deciding how every Flock camera is used. Individual agencies determine their own contracts, policies and access arrangements, while the network can allow searches across participating agencies.

The Flock technology record contains several documented examples of why that matters. Mountain View Police Department discontinued its deployment after its chief disclosed that federal agencies had accessed the city’s camera data without the city’s knowledge. An Illinois audit also found federal immigration authorities had gained access to state plate-read data through an arrangement that raised concerns under state law.

Those cases do not establish that every Flock deployment is poorly governed. They establish something more useful: access architecture is itself an important part of the technology. A camera can be locally installed while the information it produces becomes part of a much wider investigative environment.

Misuse is not a theoretical risk

The recent debate has also moved beyond abstract privacy arguments. The Washington Post identified dozens of police officials accused, charged or convicted of misusing Flock or other licence-plate-reader systems, including searches involving former partners and acquaintances. In one Indianapolis case, an officer was found to have conducted thousands of searches involving vehicles connected to people close to him.

Flock responded in August by making additional safeguards mandatory, including requirements around case numbers and automated monitoring for abnormal search behaviour. The company's response is significant because it demonstrates that technical capability and operational safeguards cannot be separated cleanly. If a system is powerful enough to reconstruct movements, it also needs controls capable of detecting when an authorised user is applying that capability for an unauthorised purpose.

That is one reason PoliceAI News records discontinued and paused deployments separately. Adoption is not the end of the story. A system can be technically successful and still be withdrawn because the legal, governance or public-trust conditions around it do not hold.

Texas shows how procurement becomes part of the debate

The current US backlash also exposes another part of the Flock story: how surveillance infrastructure gets funded.

In Texas, reporting by The Texas Tribune found that at least $30 million from a state vehicle-crime prevention programme had been directed towards expanding the Flock network, helping fund thousands of cameras. The funding originated from a $1 increase to vehicle insurance costs introduced to combat catalytic-converter theft. Governor Greg Abbott subsequently ordered state agencies to pause funding for Flock cameras amid the growing scrutiny.

That sequence is worth examining because procurement decisions can become difficult to reverse once a network has reached scale. A camera contract is not just a camera purchase. It creates infrastructure, data flows, investigative habits and expectations about what police can discover.

The Flock technology record therefore tracks more than the basic capability. It records the legal basis, oversight arrangements, access issues, documented withdrawals and changes made by the vendor. That context is often more useful than a simple list of what the software claims to do.

What should agencies be able to answer?

The most useful test for a system like this is not whether it can find a vehicle. It plainly can. The harder test is whether an agency can explain and defend the complete investigative chain.

Who authorised the search? What was the original investigative purpose? What data did the system use? Which other agencies could access it? How long was the underlying data retained? What did the AI infer rather than directly observe? Can an investigator see why a particular vehicle or “associate” was surfaced? Was the result independently verified before action was taken? Was the search audited afterwards?

Those questions are not anti-technology. They are the conditions that allow technology to remain useful when its capabilities become more powerful.

The capability is moving faster than the terminology

Calling Flock an “automatic licence plate reader” is increasingly incomplete. The camera still reads the plate, but the investigative product is becoming a broader vehicle-intelligence platform: location history, vehicle characteristics, search across participating networks and AI-assisted inference about patterns and relationships.

That evolution is precisely why a deployment tracker is more useful than a static technology list. The important fact about a police AI system is not only what it did when it was first purchased. It is how its capability, reach, access arrangements and oversight change afterwards.

Flock is a useful case study because the underlying technology is not exotic. It is already operational at a scale measured in tens of thousands of cameras and thousands of agencies. The difficult policy questions are therefore no longer hypothetical.

The question facing police leaders is becoming much narrower and more practical: if AI can turn millions of vehicle observations into inferred relationships, what rules should govern the moment when an investigative lead becomes a person of interest?