AI mass surveillance is an architectural outcome, not a feature. Four switches decide which side of the line your cameras sit on.

AI mass surveillance is the combination of four things, not one. First, centralized capture: video from many properties, cities, or agencies routes into a single pool that one operator can query. Second, identity indexing: faces, license plates, gait, or voice get OCR'd or embedded into a searchable vector so any person becomes a retrievable record. Third, cross-site correlation: the same identity can be traced across properties, cameras, and time. Fourth, indefinite retention: nothing is aged out by default. A property that installs AI on its own cameras does not automatically meet any of these conditions. AI mass surveillance is an architectural outcome of choosing all four switches, not an emergent property of using computer vision at all.

Only if the architecture routes video and identity data into a centralized pool the property does not control. A cloud-AI product that uploads every camera's feed 24/7 to a shared inference backend, applies face or plate recognition, and retains the data indefinitely across customers is, by construction, a mass-surveillance capable deployment. An edge-AI product that runs inference locally, emits only event thumbnails and metadata, does not index biometric identity, and does not pool data across properties is not. The label on the box is not the answer. The capture point, the retention policy, the identity layer, and the cross-site join are the answer.

Ask four questions and hold the answers in writing. One: where does inference run, on a device at my property or in a cloud region that also serves other customers? Two: does the system identify specific people or plates by biometric or OCR, or only detect classes like person, vehicle, package? Three: can my video frames, or any derivative of them, be queried alongside another customer's data? Four: what is the default retention period for frames, events, and identity records? If the answers are cloud, yes, yes, indefinite, the vendor is mass-surveillance capable regardless of their marketing. If the answers are edge, no, no, bounded, the vendor is not.

Usually yes, but the important property is not the chip location, it is the exit profile. A camera can have a neural accelerator on the sensor board and still upload every frame to a central ingest. The architectural question is what leaves the building and what stays. Cyrano's exit profile is event thumbnails, short clips associated with a detection, and structured metadata (time, zone, dwell, class, confidence). Full frames, continuous video, and biometric vectors do not exit. That is the sense in which the system is not mass-surveillance capable: the information required to mass-surveil is not generated or emitted in the first place.

Class detection is not biometric identification. A model that outputs 'there is a person in this tile' with a bounding box and a confidence score does not identify who that person is. A model that embeds a face into a 512-dimensional vector and compares it to a gallery of 10 million faces does. The difference is the presence or absence of an identity index. Cyrano's detection pipeline outputs classes (person, vehicle, package) and geometric attributes (position, dwell, zone entry). It does not construct or query a face, plate, gait, or voice index. For most property use cases (trespassing, package theft, tailgating, loitering), class detection plus zone rules is sufficient, and the identity layer is unnecessary.

It makes it narrower. That is the point. Mass-surveillance capability is a byproduct of the identity layer and the cross-site join. Refusing to build those means the system cannot be repurposed later into mass surveillance, and it also means the product has to be honest about what it cannot do. Cyrano cannot tell you that the person walking across the parking lot is the same person who walked across a different property's parking lot last month. That is a genuine limitation. It is also the feature that keeps a property deployment from becoming part of a pooled dataset. For the detections a multifamily operator actually has to act on in real time, the narrower product is the better product.

One HDMI cable plugged into the DVR or NVR's monitor output. The DVR has been rendering a multiview grid of up to 25 camera tiles on that port since the system was installed, and a guard monitor has been plugged into it. The install procedure is: unplug the monitor from the DVR, plug the Cyrano unit's HDMI input into the DVR, plug the monitor into the Cyrano unit's HDMI passthrough, connect ethernet, connect power. The guard still sees the same screen. Inference runs on the composite multiview frame on the device. Total time on a running DVR is under 2 minutes. No camera is touched. No RTSP is probed. No ONVIF profile is negotiated. No cloud stream is opened.

The legal exposure in most US states is not the AI layer. It is where cameras are pointed (interior units, common areas, bathrooms), what is disclosed to tenants, how footage is retained, and whether law enforcement can pull it. A property-scoped edge-AI system does not add new legal exposure beyond what the underlying cameras already had. What it does not do is expose the property to the policy debate that is driving the AI mass surveillance conversation: pooled cloud data, biometric indexing, cross-site correlation, and indefinite retention by a third party. If those are the parts of the conversation that concern you or your legal team, the architectural test in the main body of this page is the check to run before you sign a vendor contract.

Cyrano is $450 one-time for the edge unit plus $200 a month for software, for the whole property, across up to 25 camera tiles on a single HDMI port. The mass-surveillance-capable shapes (cloud-AI overlay products that upload every camera and a cloud-AI smart-camera platform that replaces every camera) are priced per camera per month, typically $20 to $120 each at a 16-camera property. The architectural choice to not generate the data required for mass surveillance is also, not accidentally, the cheaper choice: nothing is being uploaded, indexed, or stored off-property, so the compute bill is an edge-device depreciation cost plus a thin control-plane.