The pitch for an AI baby monitor is that it watches so you don’t have to — and tells you when something needs attention. The part that rarely appears in the pitch is how often it will tell you when nothing does. That number isn’t a detail. It decides whether the alerts are useful at all.
Any detection feature makes two kinds of error: false positives, where it alerts when nothing is wrong, and false negatives, where it stays silent when something is. Tuning a detector to miss less makes it alert falsely more often, and the reverse. The part most buyers never hear about is arithmetic: when the events a detector is looking for are rare — as the situations worth an alert usually are on most nights — even a detector that is right the vast majority of the time will produce more false alerts than true ones. That has a well-recognised consequence, known from clinical settings as alarm fatigue: people exposed to frequent false alarms respond more slowly, check less carefully, or turn alerts down or off. So “more detection” doesn’t automatically mean more reassurance. What makes an alert useful is design — how the threshold is set and whether it’s adjustable, whether the alert shows you why it fired, whether alerts are graded by urgency, and what happens when the connection drops. Detection rate alone is the least informative number a product can quote.
§01Two kinds of mistake, one dial
Every detector — a cry detector, a motion zone, a face-visibility check — ultimately compares what it sees against a threshold. Set the threshold sensitive and it catches more real events but also fires on shadows, rustling bedding, a sibling’s voice. Set it conservative and it stays quieter but misses more.
There is no setting that eliminates both errors. Better models move the whole trade-off in a favourable direction, but at any given level of quality, fewer misses means more false alarms. Every detection product has made a choice about where to sit on that dial, whether or not it says so.
§02Why rare events make false alerts dominate
This is the counterintuitive part, and it’s arithmetic rather than opinion.
Suppose a detector checks a situation 1,000 times across a night, and a genuine event it’s designed to catch occurs in 2 of those checks. Suppose it’s quite good: it catches real events 95% of the time, and falsely fires on only 1% of ordinary checks.
A detector that is “99% accurate” on ordinary moments still produces roughly five false alerts for every real one, because ordinary moments vastly outnumber the events it’s looking for. The rarer the event, the stronger this effect.
The numbers above are invented purely to show the shape of the effect, not taken from any product. The principle — that rare events make even a small false-alarm rate dominate — holds for any detector, and is the reason detection rate on its own tells you little.
FIG.01 — Illustrative only. The ratio depends on how rare the target event is and how often the detector evaluates — which is exactly why a product’s headline accuracy figure can’t tell you how many of its alerts will be real.
§03What repeated false alerts do
The consequence of frequent false alarms is well recognised in settings far more controlled than a nursery. In hospitals, where monitoring equipment generates large numbers of alarms, alarm fatigue is an established concern: when most alarms turn out to need no action, people learn — reasonably — to respond more slowly, to check less thoroughly, or to silence alarms altogether.
A home monitor isn’t a hospital system, and nothing here is a claim about any household’s experience. But the mechanism is general: an alert only helps if it’s acted on, and people act less on alerts that are usually wrong. A product that fires often can end up delivering less protection than one that fires rarely but reliably — because the second one is still being listened to.
There’s also a quieter cost. Public discussion has turned increasingly to whether detailed infant tracking reassures parents or gives them more to worry about. That’s not a question this article can settle, and we won’t pretend to. What can be said is narrower: an alert that frequently turns out to be nothing is, by definition, frequently interrupting for nothing — and that belongs in the calculation alongside whatever it catches.
§04Design choices that make alerts useful
Given that no detector is perfect, the differences between products lie in how they handle their own errors.
A fixed threshold suits some rooms and not others — a noisy household, a sibling sharing the room, a fan running all night. Being able to move the threshold lets a household choose its own point on the trade-off rather than inheriting one tuned for an average room. The same principle applies to simple sound activation, covered in the voice activation article.
ASK: can sensitivity be adjusted per feature, and what does each level change?An alert that comes with the clip or frame that triggered it lets a person judge in seconds whether it’s real. An alert that just says “event detected” forces either a trip to the nursery or a guess. Explainable alerts make false positives cheap, which matters more than reducing them slightly.
ASK: does each alert show what triggered it — and does showing that require uploading a clip?Not every detection deserves the same interruption. A system that separates “noteworthy” from “check now” — a quiet log entry versus a loud notification — keeps the loud channel meaningful. When everything is urgent, nothing is.
ASK: can different detections produce different kinds of notification, and can I choose which ones interrupt me?If alerts reach a phone through the internet, an outage means detection may continue with no way to tell you — covered in the on-device AI article. A well-designed system tells you plainly when it can’t deliver alerts, rather than going silently quiet in a way that looks identical to “all is well.”
ASK: when the connection drops, am I told — or does silence look the same as nothing happening?§05The numbers worth asking for
| Figure | What it tells you | Why it matters |
|---|---|---|
| DETECTION RATE | How often real events are caught | Necessary but insufficient — says nothing about false alerts |
| FALSE ALERT RATE | How often it fires when nothing is happening | Combined with how rare real events are, decides how many alerts are real |
| ALERTS PER NIGHT | How often it actually interrupts in typical homes | The most practical figure, and the one least often published |
| TEST CONDITIONS | Where and how the figures were measured | Lab results in controlled rooms rarely match real households — as with any spec claim |
TABLE.01 — A product quoting only the first row has told you the least useful thing. The third row — how often it actually interrupts in real homes — is the one most worth requesting.
§06Where our products sit
Our monitors don’t include automated detection or push alerts. They carry live video and audio over a dedicated FHSS link to a parent unit, and the judgement about what matters stays with the person watching and listening. That means no false alerts generated by software — and also no automated alerts at all, which is a real limitation for households that want a system to flag events for them.
We don’t think either approach is right for everyone. A well-designed detection system with adjustable thresholds, explainable alerts and honest offline behaviour can genuinely help. A poorly tuned one can generate enough noise that it’s switched off. If automated alerts are what you want, the questions in this article are the ones to take to any supplier — including the question of how many alerts a typical night produces. Specifications for our products are on the product pages.
§07Questions to ask about any alerting feature
- What’s the false alert rate, not just the detection rate — and how was it measured?
- How many alerts does a typical night produce in real homes?
- Can sensitivity be adjusted, feature by feature?
- Does each alert show what triggered it?
- Can I separate urgent alerts from informational ones?
- What happens to alerts when the internet drops — and will I know?
§08Frequently asked questions
Why does my AI baby monitor send so many false alerts?
Partly because of how detection works: tuning a detector to miss fewer real events makes it fire falsely more often, and no setting eliminates both errors. And partly arithmetic: when the events a detector looks for are rare, even a small false-alarm rate across many ordinary moments produces more false alerts than true ones. Adjustable sensitivity, and alerts that show what triggered them, are the design features that make this manageable.
What is alert fatigue?
It’s the tendency for people exposed to frequent false alarms to respond more slowly, check less thoroughly, or silence alerts altogether. It’s an established concern in clinical settings, where monitoring equipment generates large numbers of alarms. The general mechanism applies to any alerting system: an alert only helps if it’s acted on, and people act less on alerts that are usually wrong.
Is a higher detection rate always better?
Not on its own. Detection rate says how often real events are caught, but nothing about how often the system fires when nothing is happening. Because the two trade off against each other, a very high detection rate can come with a high false-alert rate. The more useful figures are the false-alert rate, how many alerts a typical night actually produces in real homes, and the conditions under which the figures were measured.
What makes a baby monitor alert useful?
Design more than raw accuracy. Adjustable sensitivity lets a household pick its own point on the trade-off. Explainable alerts — showing the clip or frame that triggered them — let a person judge in seconds whether an alert is real. Graded urgency keeps the loud channel meaningful. And honest offline behaviour means the system tells you when it can’t deliver alerts, rather than going quiet in a way that looks like nothing is happening.
Do True Bond monitors send alerts?
No. They carry live video and audio over a dedicated FHSS link to a parent unit, and the judgement about what matters stays with the person watching and listening. That means no software-generated false alerts, and also no automated alerts at all — a real limitation for households that want a system to flag events for them. For those households, the questions in this article are the right ones to put to any supplier.
Ask how many alerts a real night produces
It’s the most practical number in alert design and the one least often published. Our monitors don’t generate automated alerts — and we’ll tell you plainly whether that suits the household you’re buying for.
Ask about product fit → info@truebondtech.com · WhatsApp +86 189 2846 4489 · View productsThis series