Conversations

A Facebook friend has been posting about flock technology and the proliferation of cameras that capture license plate information into something resembling a national database. This caused me to research just a bit about this technology and to be honest I’m concerned about where we might be headed from multiple viewpoints including technology, government adoption, and law enforcement implementation, bad actors, civil liberties even if this technology is effective in curbing crime, et cetera. I have several questions for you. First, would you be willing to forecast likely scenarios where this will be in the next three to five years? I’m okay with multiple scenarios as long as you give me your best estimate as to their likelihood.

Yes. Your concern is reasonable, and I think the most useful way to examine this is not as a question of whether Flock is “good” or “bad,” but as a question of what happens when a genuinely useful law-enforcement technology becomes inexpensive, ubiquitous, interconnected, and politically difficult to roll back.

One clarification first: it is a little misleading to say that Flock Safety currently operates one national government database. What exists is potentially more consequential in another way: thousands of cameras and agencies can participate in interconnected searchable networks. A plate sighting contains the plate image, vehicle characteristics, time and camera location. Flock says the normal retention period is 30 days, searches require justification and are logged, and customers control sharing.

That means the raw ingredients already exist for something resembling a distributed national vehicle-location system, even if nobody has formally created a database called “The National Vehicle Tracking Database.”

My 3–5 year forecast

These aren’t mutually exclusive; I expect pieces of several to happen simultaneously.

Those percentages overlap intentionally.

What I think is most likely

By 2030, I expect Flock and competing ALPR systems to be much more extensive than they are today, but also considerably more regulated.

The reason is simple: they solve real crimes.

If a child is abducted in a gray Honda Accord and police know the plate, a system that can say “that car passed this intersection 11 minutes ago” is extraordinarily useful. The same applies to stolen vehicles, armed robbery suspects and missing persons. Police departments therefore have powerful practical incentives to adopt these systems.

And that usefulness creates political durability.

At the same time, resistance is already appearing. For example, Connecticut has enacted restrictions including a shorter 21-day retention period and limitations concerning immigration and reproductive-rights investigations. Some communities have terminated Flock contracts, while police chiefs have publicly defended them as important investigative tools.

So I don’t expect either extreme — nationwide prohibition or completely unrestricted deployment.

I expect a messy American compromise.

The development that concerns me most isn’t the camera

It’s historical searchability.

One camera photographing your car isn’t particularly remarkable. Police officers have always been able to see cars traveling on public roads.

Ten thousand cameras remembering what they saw is fundamentally different.

Suppose eventually a system can answer:

Show me every location at which vehicle ABC-123 was observed during the last six months.

That question transforms thousands of individually mundane observations into a reasonably detailed picture of someone’s life:

home → workplace → church → doctor → political meeting → girlfriend’s house → gun shop → addiction-treatment center → union hall → psychiatrist → protest → hotel.

And importantly, the computer doesn’t need to know what those places mean initially. Other databases and AI can supply that context afterward.

That distinction — observation versus retrospective reconstruction — is where I think much of the constitutional fight will eventually occur.

Flock currently says its standard retention period is 30 days, although longer retention up to a year can be offered where an elected official or governing body approves it.

Thirty days is an important protection.

But it is a policy and architecture choice, not a technological limitation.

Storage gets cheaper every year.

There’s another technological change coming

Within five years, I don’t think we’re going to talk nearly as much about license plate readers. We’ll talk about vehicle identification.

Computer vision can increasingly identify:

• make/model/year/color
• bumper stickers
• roof racks
• damage patterns
• wheels
• trailers
• vehicle shape
• direction and timing of travel

Flock already says its ALPR system captures vehicle characteristics in addition to license plates.

So imagine police don’t know your plate.

They search:

“White 2019–2022 Ford F-150, black toolbox, dent in passenger rear quarter panel, seen within five miles of this robbery.”

That is extremely valuable police technology.

But now add AI:

“Find that truck everywhere it appeared during the previous month and construct its probable travel history.”

The incremental technological steps between today’s system and that system aren’t very large.

The “bad actor” problem may actually become the decisive issue

We tend to frame surveillance as:

government versus citizen.

I’m at least as concerned about:

authorized user versus citizen.

Imagine tens of thousands of officers, dispatchers, analysts and contractors eventually having some degree of access.

It only takes one person asking:

Where has my ex-wife been going?

or

Where does that reporter go after work?

or

Where is this judge’s car at night?

Audit logs help enormously, and Flock says every search is recorded and reviewable.

But audit logs generally identify abuse after someone performs the search.

That makes me think a highly publicized misuse case is quite likely within the next several years. It could become for ALPR what famous misuse cases became for other government databases: the event that causes legislators suddenly to discover an issue privacy advocates had warned about for years.

And civil-liberties organizations are already documenting uses well beyond stolen-car searches. EFF says its examination of millions of searches found queries connected with protests and other activities that raise First Amendment and privacy concerns. That’s an advocacy organization’s characterization and should be understood as such, but the underlying question about mission creep is important.

There is also a cybersecurity problem

A database containing:

plate + place + date + time

is valuable.

A database containing billions of those observations is extraordinarily valuable.

Not merely to criminals. To foreign intelligence services, stalkers, private investigators, organized crime and anyone attempting to locate particular individuals.

Consequently, my concern isn’t simply:

“Will government misuse this?”

It is:

“Will we construct an extraordinarily valuable location dataset that someone eventually obtains who was never supposed to have it?”

Those are two different risks.

The constitutional question is fascinating

The Supreme Court has already recognized in other contexts that collecting location information over time can become constitutionally significant even though an individual’s movements occur in public.

ALPR presents almost the perfect test of that principle.

A police officer seeing my truck drive through Columbia at 2:15 PM clearly doesn’t require a warrant.

What about a computer retrieving every place my truck has appeared for six months?

One observation?

Ten?

Five hundred?

At some point quantity arguably changes the nature of the information.

I would be surprised if that question does not reach federal appellate courts repeatedly during the next five years.

The scenario I’d put my money on

Around 2030, I think America will probably have a vast ALPR infrastructure that remains legal and heavily used, accompanied by a growing regulatory structure something like this:

short default retention + mandatory search justification + immutable audit trails + serious penalties for misuse + restrictions on interstate/federal sharing + warrant requirements for sufficiently extensive historical searches.

That outcome preserves the thing police genuinely need:

“A stolen vehicle passed Camera 47 twenty minutes ago.”

while making this considerably harder:

“Show me everywhere Brad’s car has gone for the last six months.”

And that distinction is probably where I’d draw the policy line myself.

There’s a deeper reason I think this particular technology deserves attention. The dangerous capability doesn’t require anyone designing an authoritarian surveillance system. Each individual decision can be perfectly understandable:

A town installs 20 cameras to catch stolen cars.

The neighboring county does the same.

They share results because criminals cross county lines.

Other states join because criminals cross state lines.

Retention expands because detectives sometimes investigate crimes months later.

AI search improves because manually reviewing millions of records is inefficient.

Every individual step has a perfectly defensible explanation.

And twenty perfectly reasonable decisions can produce an infrastructure that nobody would have been politically willing to create in a single decision.

That, more than Flock itself, is what I would watch over the next five years.

Thanks. Please define ALPR. And yes, there isn’t a single database located somewhere, but then everything is so interconnected now days that I treat this interconnectivity as a single entity. But then, there is facial recognition, purchase information, browsing data, and even this chat. And some of the government abuse seems to target people who actively oppose elected officials. You put it all together and the picture of the future is one that makes past science fiction look pretty tame.

ALPR stands for Automated License Plate Reader (sometimes Automatic License Plate Recognition). It is a camera-and-software system that photographs vehicles, reads the license plate with computer vision, and typically records the plate, time, location, direction of travel, and often descriptive information about the vehicle. The important part is that the resulting record is searchable, not merely a photograph.

And I think your concept of treating interconnected databases as one entity is useful, with one important qualification: interoperability is not the same thing as universal access. Legal restrictions, permissions, incompatible systems, retention periods, and organizational boundaries still matter. But technologically, those barriers are increasingly easier to bridge.

What you’re describing is something larger than ALPR. It is the convergence of several information streams that historically existed separately:

• Where you go: ALPR, cellphone location, toll systems, security cameras.
• Who you associate with: communications metadata, social networks, location correlations.
• What you look like: facial recognition and increasingly sophisticated video analytics.
• What you buy: credit/debit transactions, loyalty programs, online purchases.
• What you read and search for: browsing histories, search engines and online services.
• What you say and believe: social media, emails, texts, public comments—and potentially conversations with AI systems, subject to the particular service’s privacy policies and legal protections.

The last step is what makes the future qualitatively different from traditional surveillance: AI can correlate it.

Twenty years ago, having all those databases was one problem. Someone still had to search them, reconcile identities, establish timelines and figure out what the information meant. That imposed an enormous practical limitation on surveillance.

AI radically reduces that limitation.

A future system doesn’t necessarily need a human analyst spending three weeks investigating someone. Conceivably, an authorized user could ask something resembling:

Construct a timeline of this person’s movements, purchases, communications, associates and online activity during the previous year, and identify unusual changes in behavior.

That is an entirely different capability.

Your concern about surveillance of political opponents also identifies what I regard as one of the most important issues. We shouldn’t evaluate such systems by asking whether we trust the people currently operating them. Administrations change. Police chiefs change. Prosecutors change. Laws change.

The better institutional-design question is:

Would I be comfortable giving this power to someone I strongly distrust?

That’s a surprisingly useful test.

And there is historical precedent for taking that concern seriously without assuming that every current government use is abusive. American governments have at various times improperly surveilled civil-rights leaders, antiwar activists, political organizations and other lawful groups. The lesson isn’t that contemporary officials necessarily intend to repeat those abuses. It is that institutions should be designed on the assumption that eventually somebody will be tempted to abuse them.

There is another difference from 1984 and much older science fiction that I find particularly interesting.

Orwell imagined surveillance as expensive and conspicuous: microphones, cameras, government employees and an enormous state bureaucracy watching people.

We are creating something almost opposite.

Surveillance becomes cheap, automated and largely invisible. Much of the infrastructure wasn’t even created primarily by government. We voluntarily carry location-aware phones, drive connected automobiles, make electronic purchases, use cloud services, install doorbell cameras and conduct enormous portions of our lives online. Government doesn’t necessarily have to construct the surveillance apparatus; in some circumstances it can obtain information from systems that already exist, although the legal requirements for doing so vary considerably.

That leads to what I think is the central question underlying your concern.

The problem isn’t really whether Flock cameras are acceptable.

It is whether there is some point at which the aggregation of individually reasonable information becomes something fundamentally different.

A photograph of your truck on a public highway seems innocuous.

A credit-card purchase seems innocuous.

A Google search seems innocuous.

A Facebook post seems innocuous.

A face appearing on a security camera seems innocuous.

Any individual item tells relatively little.

But combine enough of them and you can potentially construct something approaching a behavioral model of a human being.

And once AI becomes capable of continuously interpreting those records, surveillance changes from “What did this person do?” toward something more consequential:

“What is this person likely to do?”

That’s the point where some rather tame-looking science fiction starts becoming relevant—not because we’ve inevitably arrived at a dystopia, but because our technological capability is beginning to outrun many of the legal concepts developed when surveillance was labor-intensive and expensive.

There’s also an important counterweight that I wouldn’t want us to lose as we explore this. These same technologies can find kidnapped children, identify murderers, locate missing elderly people, expose government misconduct, establish that an accused person wasn’t at a crime scene, and solve crimes that previously would have remained unsolved. The tradeoff isn’t imaginary.

Which makes the difficult question not “surveillance or no surveillance?”

It’s “How much retrospective knowledge about an innocent person’s life should anyone—government or private industry—be capable of assembling, under what circumstances, and who gets to decide?”

I suspect that question will become one of the major civil-liberties issues of the next decade.

A thought that keeps coming to me is that a search warrant should be required before most searches are permitted. Exceptions might be allowed for expediency, such as a kidnapping or robbery, but even these perhaps should be subject to review. Is that feasible or even currently considered?

Yes. What you’re proposing is legally and technically feasible, and versions of it are already being considered and enacted. I also think it may ultimately become one of the principal compromises between eliminating ALPR systems and allowing essentially unrestricted searches.

The important distinction would be between collecting the photograph and searching historical data. Courts have generally treated photographing a visible license plate on a public road as ordinary observation, not a Fourth Amendment search. But the constitutional status of querying a huge historical database of those observations is much less settled. The Supreme Court’s Carpenter decision, requiring warrants for substantial historical cellphone-location data, provides an obvious analogy.

Your proposal could therefore look something like this:

1. Routine collection: Cameras can record plates passing public locations, subject to a short mandatory retention period.
2. Immediate public-safety exception: Police could search without a warrant for an active kidnapping, Amber Alert, fleeing violent suspect, stolen vehicle, imminent threat, etc.
3. Ordinary criminal investigation: Searching historical ALPR records associated with a particular person or vehicle would normally require a warrant based on probable cause.
4. Emergency searches would be audited afterward. An officer would have to document why the emergency exception was used, and inappropriate use could result in discipline or exclusion of evidence.
5. Longer retention would require judicial authorization. Otherwise, records automatically disappear after perhaps 21 or 30 days.

Interestingly, Connecticut has just moved partway toward exactly that kind of architecture: its new law limits normal ALPR retention to 21 days, while permitting longer retention when a warrant has been obtained, along with restrictions on certain sharing and uses. Washington has likewise been legislating around a 21-day general retention period with warrant-related exceptions for retaining data longer.

And this issue is closer to home than you might expect. South Carolina’s legislature currently has an ALPR/privacy bill, H.4675, expressly addressing access, storage, retention and vehicle-feature recognition. It even specifically defines AI recognition of things such as body damage, roof racks, decals, bumper stickers and paint patterns—the development we were discussing earlier.

Where I’d modify your idea slightly

I wouldn’t require a warrant for every database query.

Suppose somebody reports a stolen Buick at 10:03 and at 10:08 an officer enters the stolen plate into the system. Requiring a judge to authorize that search probably provides little additional protection while impairing one of ALPR’s most legitimate uses.

Instead, I’d make the legal threshold increase with the intrusiveness of the question being asked.

Searching:

“Has stolen vehicle ABC-123 passed a camera during the last hour?”

is considerably different from:

“Show me everywhere ABC-123 has traveled during the past month.”

And that is different again from:

“Show me every vehicle that regularly appeared near ABC-123 and identify their owners.”

That last capability starts producing an association network, not merely locating a vehicle.

I’d therefore favor something resembling a sliding scale: immediate searches for specifically defined emergencies and stolen/wanted vehicles; documented reasonable suspicion for limited investigative queries; and a warrant for retrospective movement tracking, association analysis, or extensive cross-jurisdictional searches.

And your after-the-fact review idea is particularly important

An emergency exception without meaningful review can gradually become:

“I checked the emergency box.”

A better system would make every warrantless search create a permanent audit record containing the officer, vehicle queried, stated justification, databases searched, time range requested and results accessed.

Then require genuinely independent audits—not merely the police department auditing itself.

Recent events illustrate why. Several departments have faced allegations of officers improperly using ALPR systems; investigations have resulted in firings, resignations and arrests. The fact that audit logs can uncover misuse is encouraging, but it also demonstrates that authorized access itself is a vulnerability.

There is a fascinating constitutional principle lurking underneath your suggestion:

The government shouldn’t necessarily need permission to see your automobile in public, but perhaps it should need permission to remember everywhere it has seen it.

Our Fourth Amendment jurisprudence developed when human memory and manpower placed natural limits on surveillance. A police officer could follow someone without a warrant under many circumstances—but following every citizen simultaneously, remembering every observation, and searching those memories years later was practically impossible.

Computers have removed that natural limitation.

I increasingly think that access control, warrants, short retention periods and mandatory independent auditing will matter more than arguing about whether cameras themselves should exist. Once the information infrastructure exists, the crucial civil-liberties question becomes who may ask it what questions, under what circumstances, and who will know that they asked.

And your warrant idea provides something technologically sophisticated safeguards often don’t: a person outside the investigative organization has to say “yes.” That old-fashioned bit of institutional friction may turn out to be extremely valuable in an AI-enabled surveillance world.

Sometimes I think America is quite schizophrenic (for lack of a better term). The same people who would prohibit the kinds of surveillance we are discussing are the first to criticize law enforcement’s failure to connect all the dots when their is a mass shooting or other horrific crime.

Yes—and I think you’ve identified a genuine tension in a free society. I might call it the prevention/privacy paradox rather than schizophrenic: we often demand that government possess enough information to prevent the next atrocity, while simultaneously demanding that it not collect or use enough information to monitor innocent people.

After a mass shooting or terrorist attack, investigations sometimes uncover fragments that, in retrospect, look glaringly obvious: a threatening statement here, a police contact there, a weapons purchase, a tip to an agency, disturbing online activity. Then comes the understandable question:

“They had all this information. Why didn’t somebody connect the dots?”

But before the event, those dots existed among millions of other dots belonging overwhelmingly to people who would never commit a violent act.

That’s the part hindsight obscures.

Imagine an analytical system capable of connecting everything we’ve been discussing—ALPR sightings, criminal records, social-media posts, purchasing information, tips, perhaps other legally obtainable data. Give sufficiently capable AI access to it and say:

“Find people who appear likely to commit a mass shooting.”

Technologically, that is becoming increasingly conceivable.

But now comes the uncomfortable question: What do we do with the people it identifies?

Suppose it identifies 1,000 people and one really would have committed a mass shooting.

That’s an extraordinary law-enforcement success if we can identify that one person.

It’s also 999 innocent people whom an algorithm has effectively labeled potential mass murderer.

And because these events are extremely rare relative to the population, this is a classic base-rate problem. Even an impressively accurate predictive system can produce far more false positives than true ones.

There is an additional asymmetry in how society judges the results.

If government doesn’t investigate someone and he subsequently murders twenty people, we’ll hold hearings asking:

“Why didn’t law enforcement act on the warning signs?”

If government does investigate him and he never commits the crime, we may never know whether the intervention prevented something—or whether an innocent eccentric person was subjected to government surveillance for something he was never going to do.

That creates a powerful institutional incentive toward more collection and more intervention.

And that’s where your warrant idea from the previous discussion becomes especially interesting. Warrants aren’t merely privacy devices. They introduce friction.

Friction sounds undesirable when we’re talking about catching criminals. Constitutionally, however, some friction is deliberate.

A police officer thinks, “I need to search Brad’s house.”

The Fourth Amendment essentially says:

“Before you do that, explain to someone outside your organization why.”

That doesn’t guarantee the decision is correct. But it prevents the person who wants the information from being the sole judge of whether he should obtain it.

Technology is removing a lot of accidental friction

Historically, mass surveillance was constrained partly because it was expensive and difficult.

Following 10 suspects was possible.

Following 10,000 people required an army of investigators.

Following 100 million people was impossible.

Those weren’t constitutional protections. They were technological limitations.

Computers, cameras, cheap storage and AI are gradually removing them.

That means society may have to replace technological friction with legal friction deliberately.

And I think that leads to a principle that reconciles much of the apparent contradiction you’re describing:

Government should be very good at investigating people when there is a legitimate reason to investigate them, while being deliberately bad at investigating everybody just in case one of them eventually becomes dangerous.

The difficult argument is going to be over where “legitimate reason” begins.

There’s another complication. After the next horrific event, whatever safeguards we establish will suddenly look intolerably cumbersome. After the next surveillance scandal, whatever investigative powers we’ve granted will suddenly look intolerably broad.

American history has repeatedly swung between those poles.

Perhaps that isn’t entirely a defect. A society permanently maximizing security could become oppressive; one permanently maximizing privacy could become incapable of protecting its citizens. The continual argument between the two may actually be one of the mechanisms by which a liberal democracy stays somewhere between them.

And AI is going to make that argument considerably harder, because for the first time “connect all the dots” may cease to be merely a figure of speech. The machine may genuinely be capable of doing it.

Then our problem won’t be whether we can connect the dots.

It will be deciding which dots we have a right to connect before someone has done anything wrong.