Everything else about the category follows from that one difference.
Recording versus deciding
A conventional dash cam is a loop recorder. It writes video, overwrites the oldest video, and its intelligence is the crash sensor that decides which clip to protect. Retrieving anything means finding the device and scrubbing through footage.
An AI dash cam runs the model at the point of capture. It can therefore mark the moment something happened, keep that specific segment, and send a small record of it immediately. The video becomes an attachment to an event rather than the primary artifact.
That is why inference location matters so much in this category. Moving the model off the cloud and onto the camera is what makes near-live reporting possible at all, because a system that must upload footage before it can classify it cannot report anything quickly. On the road-detection side of this network, a detection surfaces within a couple of seconds of the camera seeing it, with the image following within minutes.
Only we know
The first version of this was not a camera. Data collection began on users' own phones, an iPhone 5 among them.
That is worth more than an anecdote, because it explains the shape of everything built since. A phone on a windscreen is a bad dash cam and an excellent teacher: it proves what a single forward camera plus consumer-grade inertial sensors can and cannot resolve, before anyone has committed a cent to silicon. Dedicated Wi-Fi camera models followed in 2017 to improve the experience, and then simplified single-body designs to cut cost and complexity.
Most companies in this category started with a device and looked for the data. This one started with the data problem and arrived at the device, which is a different order of operations and it shows in what the hardware chooses to spend money on.
What the AI actually does with the road
Two capabilities do most of the work, and neither is what a spec sheet leads with.
Locating, not just identifying. The road-facing detection model, which is a separate component from the device itself, identifies and locates 29 types of object, trained on 1 million annotated frames. Position relative to the vehicle is the part that has operational value.
Holding on when vision fails. Detections drop out in shadow and in bad weather. A tracking layer keeps the object's state through those gaps so it survives instead of flickering. This is the honest answer to why real-world performance differs from demo performance, and any evaluation that does not test into a low sun or heavy rain has not tested the hard case.
Judging distance with one lens. A single camera has no depth sensor, so time to collision is estimated from how quickly an object's bounding box grows across frames. It is a geometric inference, not a measurement.
The hardware questions fleet buyers actually ask
In roughly the order they get asked.
Will it fit my windscreen. The dual-facing devices in this family mount on an adjustable bracket covering 0 to 52 degrees of rake, plus vertical support. Windscreen angle varies enormously across a mixed fleet and this is a real compatibility constraint rather than a specification flourish.
Will it flatten my batteries over a long weekend. Sleep draw is 11 milliamps on the newer device and 19 milliamps on the older one, with hibernate at 1 to 2 milliamps. The newer unit draws less despite carrying more compute.
Where does footage live. 128 GB of embedded storage, and there is no SD card slot. No card means nothing to steal, lose or corrupt, and it also means retrieval is a software problem rather than a physical one.
What else is in it. A 6-axis inertial measurement unit, GNSS, and a Wi-Fi and Bluetooth module, the radio also serving an external Mark button.
How does it see at night. A mechanical infrared cut filter physically moves between daytime and nighttime spectral requirements. A moving part inside a sealed device, accepted because the imaging alternative is worse.
On the consumer side of the same technology
The consumer product is where a lot of this was proven. Drives carry a timelapse so a route can be reviewed quickly without pulling the whole video. Small feature, and a good illustration of the category principle: the value is in not having to watch the footage.
Inference runs on both the consumer and the fleet cameras in this network.
What an AI dash cam is not
It is not a crash predictor for a named individual. It observes events. Reading an aggregate score as a forecast about one driver asks more of the data than the data supports.
It is not equally confident about everything it reports. Different event types rest on different sensors, and some are inferred from motion rather than seen. Ask a vendor, event type by event type, whether an alert arrives with footage attached or as a data point alone.
It is not a privacy-neutral device. It records people. Any serious evaluation covers what is captured, what is retained, what leaves the vehicle, and who can retrieve it, before it covers detection accuracy.
Limits
Specifications above describe particular device generations and particular trained models, not the category. No detection accuracy, alert timing or safety outcome appears in this piece; independent testing of those exists and is in review. Installed-base and coverage figures are deliberately absent.
Measured, not imagined. The camera earns its keep by deciding what not to send.