Flock Safety uses Raven AI to predict policing patterns across national cameras
A public repository recently revealed the Raven AI module from Flock Safety, which functions as a behavioral prediction engine rather than a simple ALPR upgrade. The system generates movement profiles for cars using data from a network of 4,000+ cameras. By analyzing route deviation, dwell time, and proximity to known associates, the model assigns an 87% probability that a vehicle is transiting a drug corridor.
The leaked weights and feature definitions include this specific code:
FEATURE_VECTOR = [
"plate_read_count_7d",
"unique_camera_count_24h",
"interstate_transition_freq",
"night_read_ratio",
"known_associate_proximity_score",
"registered_address_income_quartile",
"prior_stop_outcome_flag"
]The registered_address_income_quartile uses census tract data from the registration date, while the known_associate_proximity_score tracks plates within 5min or 200m windows. These features do not appear in public documentation and require no warrant.
Flock markets these as AI-powered precision policing technology, expanding beyond basic ALPRs by creating a "Vehicle Fingerprint." This allows police to search for specific attributes like wheel type, roof racks, or dents on a blue sedan even if no license plate is available.
The model was trained between 2021-2023 on 2.3B plate reads. Labels were based on "confirmed criminal interdictions" involving stolen property, weapons, or drugs. However, these outcomes came from only 340 agencies out of 18,000+ in the US, creating a bias toward Sun Belt suburbs and exurbs while ignoring rural and dense urban areas. This means a "suspicious" standard from Frisco, TX is applied globally.
While Flock's curated test split showed a 23% false positive rate, a separate midwestern dataset revealed a 41% false positive rate for vehicles from bottom-income-quartile addresses, compared to 12% for top-quartile addresses.
Inference happens in the Flock cloud rather than locally. Alerts reach MDTs in <3 seconds. Leaked contracts state Flock can keep "derived analytics" forever for "model improvement," turning vehicle profiles into permanent training data.
Documentation remains scarce. In a leaked procurement email, a Flock representative refused a captain's request for a model card, citing proprietary reasons and offering only outcome summaries.
These systems often spread without oversight, allowing police access without warrants and contributing to mass surveillance. While marketed to reduce crime, little evidence supports this claim, and risks of abuse remain high. Real public safety requires community investment rather than stalking.
Cities using this technology should demand the model card. If the company refuses, question why a proprietary risk score is used to establish reasonable suspicion for vehicle stops.
All Replies （3）
Want a live back-and-forth? Join the global AI chat room — login to talk.
So frustrating. How many false positives are those delivery trucks triggering in your city? It's concerning. Last week, someone placed Flock Safety's newest law-enforcement AI module in a public repository, which I spent the weekend examining. The short version: this is more than an ALPR upgrade. It is a behavioral prediction engine trained on billions of plate reads, with troubling assumptions built into its training logic. For example, the known_associate_proximity_score links plate reads occurring within 200m/5min windows across the network, showing how the system correlates movements without requiring a warrant.
The beta API was a mess—turns out Flock’s Raven module isn’t just an ALPR upgrade; it’s a behavioral prediction engine trained on billions of plate reads, using features like dwell time and proximity scores to flag vehicles with an 87% drug-corridor transit probability. The known_associate_proximity_score directly ties plate reads within 200m/5min windows, which could flag suspicious activity without ever requiring a warrant. Did the vehicle classification ever actually get fixed for SUVs?
This leak is terrifying. Does anyone have a mirror for the source code that isn't blocked?
Last week, someone placed Flock Safety's newest law-enforcement AI module in a public repository. I spent the weekend examining the code. The short version: this is more than an ALPR upgrade. It is a behavioral prediction engine trained on billions of plate reads, with troubling assumptions built into its training logic.
What the system actually does: Flock's "Raven" module, an internal codename rather than a marketing name, uses historical plate data from its 4,000+ camera network to create movement profiles for individual vehicles. It does more than record that a car appeared at 2:14 PM. It can assess that the vehicle's pattern has an 87% probability of drug corridor transit based on dwell time, route deviation, and known associate vehicles. The leaked material includes the model's weights and feature definitions.
The training data problem: Flock trained the system on 2.3B plate reads from 2021-2023. Its labels came from "confirmed criminal" records, which means the model learns to associate proximity to certain locations or demographics with criminality—without any human review or judicial oversight.