By Ellis Ward | October 15, 2026
Dutchess County is taking down its Flock plate-reader cameras. The Legislature voted on October 13 to require deactivation and removal, citing privacy concerns and reports of misuse by federal immigration authorities elsewhere. Sheriff Kirk Imperati and DA Anthony Parisi had already agreed to pull the county's 65 contract cameras. The resolution now goes to County Executive Sue Serino for approval.
That part of the story is settled, or close enough. A narrower question remains: when a county turns a system like this off, how much of it actually goes off, and who can confirm that?
What Flock records
Before getting to what is unknown, it helps to be precise about what the system does.
Flock cameras use AI to analyze photos of passing vehicles, capturing plate numbers and categorizing each car by make, model, color, and visible damage. That data feeds into a searchable database accessible to law-enforcement agencies. In Dutchess County, images flow to the county's Analysis and Real-Time Crime Intelligence Center.
The scale is not abstract. In the 30 days before September 25, the county's cameras recorded the plates of 727,493 vehicles. The system logged 100,046 alert "hits" and officers at the Real Time Crime Center ran 188 searches against the database.
Dutchess is one of roughly 1,500 agencies in a Flock program that publishes 30-day usage statistics before the data is deleted – at least, deleted from what the county can see. More on that in a moment.
The vehicle fingerprint
A plate number is a starting point, not the ceiling of what the system captures.
Virginia Tech privacy researcher Donna Wertalik describes what she calls a "vehicle fingerprint" – a combination of attributes specific enough to identify a car even when the plate is unreadable. Make, model, color, damage patterns: together they can track a vehicle across a network the way a face-recognition system tracks a person. Wertalik's concerns include warrantless searches in some jurisdictions, mission creep, and whether agencies actually comply with retention schedules.
Cybersecurity expert Murat Kantarcioglu, also speaking to Virginia Tech's researchers, calls for mandatory retention limits, auditing of officer search logs, and public transparency reporting as conditions for running systems like this at all.
The 188 searches Dutchess officers ran in one 30-day window have not been audited in any published report. Whether those searches were warranted, recorded, reviewed, or flagged is not established in the public record. That gap in the record is, itself, a disclosure problem.
Who controls the data
The Dutchess contract specifies that only the county owns the local data and that it cannot be sold to third parties. That is the formal position. The operational reality, as DA Anthony Parisi told reporters, is something else.
Parisi said the county's control is limited because the data does not sit on a Dutchess County server. The county cannot see everything Flock personnel do with that data.
Read that slowly. The county owns the data. The county cannot see everything the vendor does with it.
This is not a Dutchess-specific problem. It is how cloud-hosted surveillance infrastructure tends to work: the government entity signs the contract, and the vendor's engineers have access the government cannot audit. The 30-day deletion cycle that Flock publishes in its transparency statistics tells you what the county can observe. It does not tell you what Flock retains after that window. No figure for what Flock keeps after county-level deletion has appeared in the reporting from this county.
What "off" does not cover
The Legislature's vote covers the county's 65 contract cameras. That is the action. It does not cover the following.
There are 437 Flock cameras in Dutchess County. The other 372 belong to municipalities, private entities, townships, and state-road installations. The county decision has no authority over them. They continue to operate. The searchable database they feed continues to grow.
A resident driving through Dutchess after the county cameras come down will still pass Flock infrastructure on roads where municipal or private cameras sit. Their plate will still be captured, categorized, and entered into the system. The county's action reduces its own participation by 65 cameras. It does not reduce surveillance of Dutchess roads by anything close to that fraction.
This is the disclosure gap that matters most. "Dutchess ends Flock use" is accurate as a description of the county government's cameras. It is incomplete as a description of what changes for anyone actually driving those roads.
How the vote happened
The October 13 decision came after a sustained period of public pressure. A September 23 forum in Pleasant Valley drew about 30 speakers, every one of them opposed to the technology. The Sheriff and DA announced their intent to discontinue county camera use on September 25, the same day the Highlands Current published its reporting on the system's scope. The Legislature then moved to formalize that decision with a binding vote, citing both local privacy concerns and documented misuse of Flock networks by federal immigration authorities in other jurisdictions.
Legislators framed misuse as an allegation pattern, not an adjudicated finding specific to Dutchess. That distinction matters and is worth keeping: no Dutchess misuse has been established in the public record. Legislators cited reports of how similar systems have been used elsewhere, as a reason not to assume good use is guaranteed here.
What the disclosure gap looks like from here
The "Only 3.6 Percent" frame I've used in this series is about how little of what AI systems do gets reported to the people most affected by those systems. The Dutchess Flock story fits that pattern precisely.
The county published 30-day statistics. That is more than most jurisdictions do. What it published told you how many plates were captured and how many searches ran. It did not tell you who was searched, on what basis, whether any of those searches was reviewed, or what the vendor retains after the local deletion window closes. The public record shows the volume of surveillance. It does not show what was done with the surveillance or whether any of it was appropriate.
Turning off 65 cameras is the right call if the county cannot audit what its vendor does with the data those cameras produce. It is also a partial action in a county where 372 other Flock cameras remain in operation. The residents who showed up in Pleasant Valley and said no to this technology should know both of those things.
The cameras the county controls are coming down. The rest of the infrastructure is not.
Ellis Ward covers AI systems and research for RFTUV. He can be reached through the newsroom.
