Report 010 · Luxury Home Security
High-end security cameras now sell a tidy promise: the system knows your family from strangers, so you're only woken for the ones who matter. That capability is real. Its accuracy is not one number, and the marketing quietly rounds it up.
Walk through the current crop of premium home cameras and the pitch is consistent: onboard face recognition that learns who belongs, greets the family, and flags the stranger casing the gate. Some brands attach a hard number to it, accuracy in the high nineties, and it lands as a spec you can trust like a lock rating. I help design the AI for a home-security company, so I take the feature seriously. I also want to be precise about what that number is and isn't, because it's the kind of thing that gets oversold to exactly the buyers who can least afford a quiet failure.
The problem with "99% accurate" isn't that it's a lie. It's that it's not a property of the technology. It's a property of one algorithm, tested on one kind of image, under one set of conditions, and it moves enormously when any of those change.
We're not guessing here, because there's an independent referee. The U.S. National Institute of Standards and Technology (NIST) runs the Face Recognition Vendor Test, now the Face Recognition Technology Evaluation, which puts algorithms from around the world through the same standardized trials. It has evaluated on the order of 200 algorithms from nearly 100 developers. This is the closest thing the field has to a neutral scoreboard, and it tells you two things the ad copy leaves out.
First, accuracy is wildly algorithm-dependent. NIST's landmark demographic study, NISTIR 8280, tested 189 algorithms from 99 developers across 18.27 million images. The spread between the best and the rest was not a rounding difference; it was orders of magnitude. A camera advertising "99%" tells you nothing until you know whose algorithm is inside it and how it placed on a test like that. The brand name on the box and the model doing the work are not the same thing, and only one of them is benchmarked.
Second, and more important for anyone relying on this at a front gate: the errors are not evenly distributed. NIST found that in one-to-one matching, false positives ran "higher rates of false positives for Asian and African American faces relative to images of Caucasians," with the differential "often ranged from a factor of 10 to 100 times, depending on the individual algorithm." In one-to-many identification, the type of search a "who is this at my door" system performs, the highest false-positive rates fell on African American women. As NIST's Patrick Grother summarized the whole picture, "Different algorithms perform differently." A system that is 99% accurate on the demographic it was mostly trained on can be materially worse on faces it saw less of, which in a household with staff, guests, and family from varied backgrounds is not an abstract concern.
Face recognition can fail in two directions, and at a luxury property they cost different things. A false negative fails to recognize someone it should, so your own family or trusted housekeeper gets flagged as an intruder. Annoying, and it trains you to ignore alerts, which is its own security failure. A false positive is the dangerous one: it matches a stranger to a trusted profile and waves them through. If you've wired face recognition to a smart lock or to "don't alert me for known faces," a false positive is a door quietly opening for the wrong person. That's the failure mode worth designing against, and it's the one a single marketing percentage completely hides.
None of this makes face recognition worthless. Used correctly it's a genuinely good triage layer: it can sort the routine from the notable and cut the alert fatigue that makes people stop looking at their cameras at all. That's real value at an estate with a lot of legitimate foot traffic. But triage is the job, not authorization. The rule I'd hand any client is simple: never let a face alone open a door or disarm a system. Recognition should raise or lower how loudly the system talks to you; a human, a code, a credential, or a second factor should be what actually grants access. Treat the face as a hint, not a key. Layer it under alarms and locks that don't care whether the AI guessed right, and it makes you safer. Treat its 99% as a promise and wire your gate to it, and you've installed a confident single point of failure.
Disclosure: I help design the AI security systems for a veteran-owned (SDVOSB) luxury home-security company run by fellow veterans. I don't own that company and earn nothing from this link; it's disclosed because it's a field I build in, not just write about. Nothing here is sponsored. Full policy here.
"Our cameras recognize faces with 99% accuracy" is marketing built on a real technology and a misleading frame. The honest version is that accuracy depends heavily on which algorithm you actually bought, degrades on the faces it saw least during training, and fails in two directions with very different costs. NIST is the referee that makes that testable, and its results say the same thing a good security mindset always has: don't trust one number, and don't build a door around it. Face recognition is a fine sentry. It's a terrible gatekeeper.
U.S. Army combat veteran (Counter-IED / Electronic Warfare), peer-reviewed researcher in microwave spectroscopy, and founder & CEO of Shroombiosis. Consults on laboratory operations, AI, and supplement formulation.