Flock’s OS Investigate shows where police AI gets slippery
Decrypt reports that Flock’s OS Investigate includes preloaded AI prompts for searching camera footage without a name or license plate. The practical issue is not sci-fi omniscience. It is how quickly pattern search can turn ordinary camera networks into investigative infrastructure.
TL;DR: The real concern with Flock’s reported OS Investigate tool is not that AI can magically identify anyone, it is that police search can move from known identifiers to fuzzy behavioral patterns at scale.
What did Decrypt report about Flock’s OS Investigate?
Decrypt’s report, “Flock’s New AI Police Tool Can Track Drivers Without a Name or License Plate,” says the underlying code for Flock’s “OS Investigate” includes 69 preloaded AI prompts. According to Decrypt, those prompts can turn Flock camera footage into a search system that helps identify people by how they move, not only by a name or a license plate.
That is the key shift.
Traditional camera search is usually anchored to something explicit. A plate. A vehicle make. A timestamp. A location. The operator starts with a known field and narrows from there. The reported OS Investigate setup points toward a different model: describe a pattern, let the system search across observations, then surface likely matches.
That does not mean the system is accurate. Decrypt’s report, at least from the material provided here, does not give accuracy numbers, false positive rates, deployment scope, auditing practices, or details on which agencies have access. Those gaps matter. A demo prompt and an operational investigative tool are not the same thing.
But the absence of those details does not make the concern imaginary. It makes the concern harder to evaluate.
Why is “without a plate” such a big deal?
Because identifiers create friction.
If police need a plate number, they need a reason to care about that plate. If they need a name, they need a person already in view. When search moves to behavior, clothing, gait, vehicle movement, or route pattern, the front door gets wider.
That can be useful. If there is an abduction, a violent crime, or a credible threat, searching for a partial description across many cameras may help investigators move faster. This is why these tools will keep getting built. The operational demand is real.
The problem is that fuzzy search also invites fuzzy suspicion. “Find vehicles that moved like this.” “Find a person who appears similar.” “Find someone who returned to this area multiple times.” Those queries can be legitimate in one case and abusive in another. The same capability that helps reconstruct a crime can also support dragnet searches, protest monitoring, or fishing expeditions.
This is where AI changes the politics of surveillance. Not by making cameras new. By making old footage easier to query.
A camera network used to be limited by human attention. Someone had to watch, scrub, compare, and remember. AI reduces that labor. Once the labor cost drops, the practical boundary shifts from “can we do this?” to “are we allowed to do this?” That is a governance problem, not only a model problem.
What should buyers and cities ask before using this?
The first question is simple: what are officers actually allowed to type?
Prompt libraries matter because they encode use cases. If Decrypt’s reporting is right that OS Investigate includes 69 preloaded prompts, those prompts are not just convenience features. They are policy in product form. They tell users what the system thinks is normal to ask.
The second question is about evidence handling. Is an AI match treated as a lead, a probable cause input, or something else? Are officers required to document the original query? Can defense attorneys inspect the query path? Are rejected matches logged too, or only the convenient hits?
The third question is independent measurement. A vendor saying the tool works is not enough. A police department saying it helped in a case is not enough either. These systems need tests on false positives, demographic effects, camera quality, weather, occlusion, and adversarial conditions. Especially if “how someone moves” becomes part of the search logic.
The catch for builders is that this is not just a public-sector issue. Any company building video AI, workplace monitoring, retail analytics, or fleet intelligence faces the same product boundary. Pattern search is powerful. It is also easy to normalize. If you are building this kind of tool, start with query logs, permission tiers, retention limits, human review, and explicit banned queries before you ship the flashy search box. The demo should not be easier to build than the safeguards.