The short answer
In our study an AI model reviewing short thermal clips made no false fall calls in 76 hours of ordinary life in a lived-in room (395 clips; at most about 1.2 a day at 95% confidence). Two detectors trained the conventional way on public laboratory data would have raised 14 and 24 false alarms a day in the same room. The price of the AI's caution is missed falls: it caught 69% of laboratory falls and 4 of the 9 staged falls the camera could see in the room, where the trained detectors caught more.
Why false alarms matter more than they look
Detection rate is the number vendors lead with, but in a home the number that decides whether a system is used is how often it is wrong. A family that is called every day about a fall that did not happen stops answering, and an agency coordinator who gets 24 alerts a day for one room cannot act on any of them. A fall detector that cries wolf is turned off, unplugged, or ignored, and then it catches nothing.
What we measured
Our research prototype, a 160 by 120 pixel thermal camera and a microphone, ran in one lived-in room from 6 to 13 September 2026. It woke on sound with movement or on large movement, and cut an 18-second clip around each wake. Every clip from 76 hours of ordinary living (395 clips, with our own staged falls and test sessions set aside) was sent to Google's Gemini 3.8 Flash with a three-line instruction and no training on falls. We also ran two detectors built the conventional way on a public laboratory dataset of thermal falls, a boosted-tree classifier and a hand-tuned descent rule, on the same clips.
What we found
| Screener | False alarms per day at home | Laboratory falls caught | Visible falls caught in the room |
|---|---|---|---|
| AI reviewer (Gemini 3.8 Flash, no training) | 0 (95% upper bound 1.2) | 69% | 4 of 9 |
| Boosted trees, trained on laboratory falls | 14 | 100% | 6 of 9 |
| Descent rule, tuned on laboratory falls | 24 | 94% | 7 of 9 |
Laboratory falls: the held-out subjects of the public dataset, at its 32 by 24 pixels. False alarms: 395 clips from 76 hours of ordinary living.
One afternoon we set aside because we were testing the system ourselves held one more clip the AI called a fall: someone dropped behind the bed and came back on all fours. Counted as ordinary living, that would be 1 false call in 80 hours. The AI also called none of the 499 laboratory clips of people walking, sitting, lying on the floor and getting up a fall.
What the AI missed, and why
The AI missed falls mostly for one reason: its instructions. One of its three lines tells it that a person who goes out of view, behind furniture or below the frame, has not fallen, a line written to stop false alarms. Falls that ended at the edge of the picture were declined under that rule, often after the AI had described the fall itself. Without that one sentence it caught 83% of the laboratory falls instead of 69%, and 9 of the 9 visible falls in the room, but it also called 2 of 2 people deliberately lying down on the floor a fall. Where to draw that line is a design choice between missed falls and false alarms, not a fact of the technology.
What it means if you are choosing a system
- Ask for false alarms per day, in real homes. A detection rate from a laboratory says little about life at home; our trained detectors caught nearly every laboratory fall and would have alarmed 14 to 24 times a day.
- Ask what happens before anyone is called. A check-in with the client, or a person reviewing the alert, is what keeps the remaining false alarms from reaching a family at 3 a.m.
- Ask whether an alert can be checked. When our AI called a fall, it said when it happened, within a second of the real moment in 99% of cases, so a person can verify it in seconds.
- Ask what it misses. Every system trades missed falls for false alarms. A vendor who says it has neither has not measured both.
Questions this answers
Do AI fall detectors give false alarms?
All fall detectors do sometimes. In our study an AI reviewer made none in 76 hours of ordinary life in one room, while two conventionally trained detectors would have raised 14 and 24 a day in the same room.
Is zero false alarms realistic for a product?
Not as a promise. Our result is one room over 76 hours, with a 95% upper bound of about 1.2 false calls a day, on a research prototype. Other homes, pets, visitors and televisions will produce cases this study did not see.
Why not use the detector that catches every fall?
Because in a home it would have called a fall 14 to 24 times a day, and an alarm that rings that often is ignored or turned off.
The paper
This finding comes from Can an AI tell from thermal video whether someone fell?, which describes the methods, the data and every limitation we know of.
Honest scope
- The falls were staged, by younger adults, in one laboratory and one room. Real falls of older people are different.
- The room recordings come from a research prototype, not the product's final sensor, mounting or software.
- These are research results, not a promise of how often a product will catch a fall. No system catches every fall.
Senecta makes the pods this research is for, so this is our own study, not an independent evaluation.