Flock Safety is usually described as a license plate reader company. That undersells what its system does. Reading plates is only the first step. The rest of the system is an AI pipeline: models that classify vehicles, a natural-language search engine over camera footage, and algorithms that flag travel patterns for police.
That pipeline is now under federal scrutiny. On September 23, 2026, a Senate Judiciary subcommittee held the first congressional hearing on AI surveillance camera makers like Flock. Documented misreads have led to innocent drivers being stopped at gunpoint and, in one case, jailed.
This article explains what each layer of Flock’s AI does, where the errors come from, and how to evaluate the company’s accuracy claims.
What AI Does Inside the Flock System
Flock’s AI works in three layers:
- Recognition: reading the plate and classifying the vehicle carrying it.
- Search: letting officers query stored images with plain-language descriptions.
- Pattern analysis: flagging vehicles whose movements match patterns Flock associates with crime.
Each layer adds capability, and each adds its own way to be wrong.
Vehicle Fingerprint: Classifying Cars, Not Just Plates
A traditional plate reader converts a plate image into characters. Flock adds a second model that describes the vehicle itself.
Flock calls this “Vehicle Fingerprint” technology. The company says it identifies make, model, and color, plus distinguishing details such as mismatched paint, bumper stickers, dents, and temporary plates. Flock also says its cameras capture vehicles moving up to 100 miles per hour at distances up to 75 feet, regardless of lighting.
This is what makes the system more powerful than older plate readers. A police officer doesn’t need a plate number. A description of the car is enough. It also means the system stores far more information about every passing vehicle than a plate alone.
FreeForm: Natural-Language Search Over Camera Footage
FreeForm is Flock’s plain-English search tool. Instead of filling in fixed fields, an investigator types a description. Flock’s own examples include “blue SUV with a racing stripe” and “white F-150 with a ladder in the back.”
This is the same shift happening across AI video surveillance: moving from recorded footage people review by hand to footage a model can search on request.
FreeForm Searches Can Include People
FreeForm isn’t limited to vehicles on every camera. On video cameras, Flock says FreeForm searches clothing, colors, and objects. On license plate cameras, it’s limited to vehicles. 404 Media reported in July that officers were using FreeForm to search for people by clothing and tattoos.
An AI Model Moderates the Searches
Flock screens FreeForm queries with a second AI model. According to code reconstructed by WIRED, the model scores an officer’s description of a person against eight categories of sensitive content before the search runs, and returns one of three verdicts: allow, block, or warn.
WIRED’s reporting also shows the limits of that design. When a California officer typed “American flag,” the search was blocked when aimed at a person, then ran across 11,000 cameras when aimed at vehicles instead. The filter judged who the search targeted, not what the search revealed.
Pattern Analysis: When the AI Flags Your Travel
Flock’s newest tools go beyond answering an officer’s question. They raise flags on their own.
- Multi-State Insights: Flock alerts police when suspect vehicles have been spotted in multiple states over the past 30 days, which it says helps identify patterns tied to crime networks.
- Plate swap detection: Flock markets alerts for vehicles whose plate doesn’t match the vehicle it’s on, a tactic used to evade plate readers.
- Flock Nova: EPIC’s amicus brief says Nova integrates commercial data broker and open-source intelligence sources, which one Flock employee described as helping to “jump from plate to person.”
Critics argue this is the most consequential change. The ACLU warned that an algorithm could now point police toward a driver just because the driver’s travel pattern looks suspicious to the model.
Where Flock’s AI Gets It Wrong
The 93% Claim, Worked Out
The Institute for Justice reports that Flock claims its cameras accurately capture 93 out of every 100 plates, across more than 20 billion plate readings a month.
Take those two claims together. If 7% of 20 billion monthly reads are not accurately captured, that’s roughly 1.4 billion imperfect reads every month.
That calculation needs a caveat. It isn’t clear what “accurately capture” means in Flock’s marketing. It could mean a missed plate, a partial read, or a wrong character, and most bad reads never trigger any alert. But the scale matters. Even a small error rate becomes a large number of mistakes across a network this size.
Three Places an Error Can Enter
Wrongful stops don’t all come from the same failure. Separating them clarifies who is responsible.
| Failure point | What goes wrong | Documented example |
|---|---|---|
| Character recognition | The model misreads a character on the plate | In Toledo, a Flock camera read a 7 on a man’s plate as a 2, producing a false stolen-vehicle hit. Police stopped him at gunpoint and sent a dog after him. |
| Vehicle matching | An alert is matched to the wrong vehicle | In York County, South Carolina, a deputy got an alert about a stolen dark BMW nearby and approached a man’s black BMW with his gun drawn, according to a lawsuit. |
| Vehicle matching | A camera links the wrong vehicle to a crime | Lindsay Isaacs testified at the Senate hearing that she spent nearly two weeks in jail after a Flock camera mistakenly tied her vehicle to a crash that killed three people. |
| Bad source data | The plate is read correctly, but the watch list is wrong | An LAPD inspector general audit found 161 vehicles falsely labeled stolen over two months, linked to Flock’s hot list alerts. |
Flock has pushed back on many of these incidents. The company argues its cameras worked correctly and that the fault lay with stale entries in stolen-vehicle databases, not with its character recognition.
That argument is partly correct and partly beside the point. An alert system is only as reliable as its weakest input, and drivers experience all three failures the same way.
The Institute for Justice counts at least 26 cases since 2018 in which Flock misreads led to innocent people being pulled over, held at gunpoint, jailed, or attacked by police dogs, most of them since 2023. Those are only the cases that left a record.
The Missing Human Check
The most important safeguard isn’t in the model at all. It’s whether officers confirm an alert before acting on it.
One review of the reporting noted that it doesn’t show whether departments train officers to verify a plate before treating a Flock alert as grounds for a felony stop, calling that the load-bearing question.
This is the same debate playing out across enterprise AI, where human-in-the-loop review is increasingly treated as a requirement rather than an option. A model flags. A person should decide.
Who Trains Flock’s AI
Flock’s recognition models learn from labeled footage, and some of that labeling happens overseas. 404 Media reported that Flock uses overseas workers from Upwork to train its machine learning algorithms, with training material showing how to review and categorize images of people and vehicles in the U.S.
The contracts matter here. Footnote4a found that Flock’s customer contracts grant the company an irrevocable, worldwide license to use customer data to provide its services.
For drivers, this means images of their vehicles may be used to train the system, not just to answer police searches.
AI Watching the Searchers: Flock’s New Guardrails
Flock’s response to misuse has been to point automated monitoring at its own users.
Audit Assistance flags unusual searches. The feature flags abnormal search patterns for review. More than a third of customers had adopted it voluntarily, and Flock is making it mandatory for all law enforcement customers by the end of 2026. Accounts showing abnormal behavior can now be suspended automatically, even before an administrator reviews them.
Structured fields replace typed reasons. Since December 2025, officers must pick at least one standardized offense type before any search, on their own cameras or on statewide and national lookups. The change followed an EFF analysis that found officers entering reasons like “LMAO” and “WEIRD KID,” including one search across more than 19,000 cameras.
Why structure matters. A dropdown gives audit tools consistent data to analyze. That makes misuse easier to detect, though it doesn’t prove any search was justified.
This is a familiar pattern in AI governance. The fix for an AI system’s risks is often more AI: anomaly detection applied to the humans using it. That only works if someone reviews the flags and acts on them.
Is Flock Using Facial Recognition?
Flock says no. The company states that FreeForm does not use facial recognition or person recognition.
That statement is narrower than it sounds. FreeForm can still return people by clothing and appearance on enabled video cameras. It just doesn’t identify them by face. How face-matching systems differ, and why they draw even stricter regulation, is covered in FaceCheck ID in 2026: Safety Tool or Privacy Trap?
How to Judge Flock’s AI Accuracy Claims
A single accuracy percentage tells you little about an AI system deployed at this scale. Vendors in many AI categories publish headline numbers that nobody independently checks, a problem this site found when testing accuracy claims for AI takeoff software.
For Flock, these are the questions that matter:
- What’s the error rate on alerts, not reads? A capture rate says nothing about how often a hot-list alert points to an innocent driver. The LAPD audit measured alerts, and that’s why its finding carried weight.
- Who audited it? A company’s own figure is marketing, not evidence.
- What must an officer verify before a stop? Look for a written policy, not just a promise.
- How long is data kept, and who can search it? Retention and sharing settle how far one misread can travel.
Those last questions are governance questions, not engineering ones. That’s the argument for treating AI deployment as a governance problem first: accuracy flaws can be patched, but a system with no audit trail can’t be held accountable.
Where the Oversight Fight Is Headed
The debate has moved from city councils to Congress.
Sen. Josh Hawley’s investigation began with an August 2026 letter asking Flock, among other things, how accurately it identifies vehicles and how often it issues incorrect alerts. At the September 23 hearing, Hawley said AI has made plate reader systems searchable by details such as vehicle damage, roof racks, and bumper stickers. He argued that most drivers captured by these systems have done nothing wrong.
The testimony focused on what the technology has become. Hawley said the country is “a long way from a simple automated license plate reader,” after the Institute for Justice’s Alasdair Whitney argued that cloud-linked cameras represent a drastic shift.
Local governments are already acting on accuracy concerns. The LAPD let its Flock contract expire after its inspector general’s audit.
The open question is whether an AI system that classifies, searches, and flags at national scale has error rates anyone has measured independently. No federal rules currently answer it. Hawley has called for federal protections for AI surveillance networks but has not introduced Flock-specific legislation.
What Flock’s AI Means for Drivers
Flock’s AI can find a car from a description, search thousands of cameras in seconds, and flag travel patterns nobody asked about. The same capabilities that make it useful to investigators make its errors more consequential.
The most useful question isn’t whether the AI is accurate. It’s what happens when the AI is wrong, and who is required to check before a driver pays for the mistake.
Related: AI Glasses for Accessibility: What They Can Actually Do in 2026
