What does "AI" on a security camera actually mean?

Every camera box now says AI, from the £30 doorbell to the enterprise dome, and the word is doing a different job on each of them. The spec sheet rarely says which, so it's worth laying the tiers out plainly. None of them is a con exactly; they answer very different questions, and only one answers the question you'll actually have after an incident.

Tier one: motion detection (not AI at all)

Motion detection is pixel change. The recorder compares one frame with the next, and if enough pixels differ it fires, which is why wind in a hedge and a headlight sweep both count as events. It's been in recorders for decades and it isn't AI, however it's badged. As a trigger for recording it's fine; as an alert it's the reason people mute their camera notifications inside a week.

Tier two: the on-camera classifier (what the badge usually means)

When a camera box says AI, this is usually what's inside: a small model running on the camera that labels motion as a person, a vehicle or an animal. It's real machine learning and a real improvement, because the hedge stops triggering and the person still does. But that's where it ends. It can't tell you what happened, only that a person-shaped thing moved, so “person detected on camera 3” at two in the morning still means squinting at a clip to find out whether it was a burglar or the neighbour's teenager taking a shortcut.

Tier three: scene understanding

The third tier reads the scene rather than classifying the blob. A vision-language model writes a plain-English description of what's actually going on in the frame, the people, what they're wearing, what they're doing, and that written record changes what footage is for. It becomes searchable (“person in a red jacket near the till”, typed weeks later, jumps to the moment), and alerts can be about events rather than categories: a fall, a concealment at the till, someone in a zone they shouldn't be in outside hours, a queue past four people. An event is judged over a window of recent frames, so a fall reads as someone going down and staying down, not one ambiguous image.

What to ask before buying an AI camera

The badge won't tell you the tier, but the answers to a few questions will.

• Does it tell you what happened, or just that something moved? Ask to see an actual alert. “Person detected” is tier two; a sentence about the event is tier three.
• Can you search what it saw last Tuesday? A classifier keeps no account of the scene, so it can't answer. If the footage is searchable in plain English, something wrote descriptions worth searching.
• Does it speak open standards? A camera with an RTSP stream can feed any system, now or later. One locked to a vendor's app keeps the intelligence, and your footage, wherever the vendor wants it.

The intelligence doesn't have to live on the camera

The quiet part is that tier three doesn't need to be in the camera at all. A camera's job is to point at the scene; the understanding can run elsewhere. Svid adds scene understanding to whatever camera you already own: it captures the RTSP stream server-side (or takes an uploaded file), has a vision model describe the footage, and makes the result searchable and watchable by checks that alert on events. No new hardware, and the AI on the box stops mattering.

1. 01

   Create an account at app.svid.ai

   Four questions about your setup suggest a watch level and a set of checks. Svid is pay-as-you-go (top up from $5), with the full price shown before you pay, and each check is a flat one-time $5.

2. 02

   Connect a camera or upload a clip

   An RTSP stream is captured server-side, so any camera that speaks the standard works as it is. No stream to hand? An uploaded file goes through the same pipeline.

3. 03

   Pick the events worth hearing about

   Checks are chosen individually: Theft Watch on the till camera, Vehicle Log on the yard, nothing at all on the corridor nobody worries about.

4. 04

   Search it, and let it alert you

   Everything described is searchable in plain English, and alerts land on screen, as push, by email, by webhook or on WhatsApp.

Are AI cameras worth it?

It depends what the badge means. A person and vehicle classifier is worth having over bare motion detection if false alerts are drowning you. If the question is what happened, only scene understanding answers it, and that can be added to the cameras you already own rather than bought as new hardware.

Is motion detection AI?

No. It's a pixel-change comparison that's been in recorders for decades. Useful as a recording trigger, but there's no model and no understanding involved, whatever the box says.

Do I need to replace my cameras to get real AI?

No. Scene understanding can run off the camera entirely: any RTSP stream or uploaded file can be described, indexed and watched by checks. The camera provides the picture; the intelligence lives elsewhere.

What's the difference between person detection and scene understanding?

Person detection tells you a person-shaped thing moved. Scene understanding writes down what's happening in the frame, in words, which is what makes footage searchable and lets an alert describe an event rather than a category.