The real divide in camera AI isn't capability. It's control.

A bus is now the front line

Kansas City wants to put facial recognition on its public buses. The plan, run by the regional transit authority with a vendor called SafeSpace Global, would add cameras that scan riders' faces in real time and match them against three lists: banned riders, missing persons, and law-enforcement watch lists. The state of Missouri declined to fund it specifically over the facial-recognition piece. The city is going ahead anyway, on local and federal money.

The reaction was immediate. The ACLU, the Electronic Frontier Foundation, and digital-rights groups lined up against it. One privacy analyst called running face recognition on live public space "a line that until recently has never really been crossed." A transit official's defense was, more or less, that there have always been cameras on buses and this is "just new technology."

Here is where I land, and it is not the easy answer.

The problem was never that a camera can recognize a face

We build facial recognition too, so I am not going to pretend the technology is the villain. The problem with a project like the one in Kansas City is everything around the capability: where it runs, who controls the list it matches against, how long footage is kept, and what it is allowed to be pointed at.

That is the real divide in camera AI. Not capability. Control.

Same capability. Opposite architecture. Opposite outcome.

Same capability, two opposite systems

Take the exact same face-recognition model and drop it into two different architectures.

• A surveillance system runs the recognition in someone else's cloud, matches faces against a watch list nobody outside the agency can audit, retains the video for years, and can quietly be pointed at anyone. You have built a net.
• A controlled system runs the same capability on the customer's own hardware, where raw video never leaves the building, the customer owns the list and the retention policy, and it is scoped to a clear safety or access purpose. You have built a tool that serves the people inside the building instead of watching them.

Same model. Opposite outcome. The thing that changed was never the algorithm. It was who holds it and where it lives.

The data is where you see which one you've got

Watch the data, not the demo. The transit plan's own design is the tell: it says biometric face data is discarded on a non-match, but regular bus video is still archived on a local server for up to five years, and the watch list is defined by the authority itself with no legislative limit on what it can grow to include. As the ACLU put it, a narrow watch list today is the thing that quietly widens over time.

That is a control problem, not a capability problem. Facial recognition has already produced a documented string of wrongful arrests in the US, and federal testing has shown materially higher misidentification rates for Black faces and women. Those failures get dangerous precisely when the system is unauditable, retained forever, and pointed at the public by default. Change the architecture and you change the risk.

How we build it

At DHI, the capability runs on hardware you own, on your own site. Raw video never leaves the building. When recognition is part of a deployment, the customer owns the watch list and the retention policy, and it is scoped to a defined safety or access purpose rather than open-ended monitoring of the public. Nothing streams to our cloud, because there is no cloud in the path to stream to.

That is the same "privacy by architecture" idea we apply to every detection we run. The point is not to avoid powerful capabilities. It is to make sure the people who own the building also own the system watching it.

The fight that actually matters

The industry keeps arguing about whether this technology should exist. I think that is the wrong fight. The one that matters is who controls it and where it lives.

So when any camera AI gets pitched into a public space, ask the four questions that actually separate a tool from a surveillance net: Where does it run? Who controls the list it matches against? How long is footage kept, and where? What is it allowed to be pointed at? The answers, not the capability, tell you what you are really buying.

For background on the Kansas City plan and the debate around it, see the AP wire report (June 18, 2026) and KCUR's local coverage (June 25, 2026).