You slam your phone face-down on the table for the third time this week, and nothing happens. No vibration, no panicked screen, no automated voice asking if you need an ambulance.
Good. That's the system working.
Modern iPhones and Pixel phones carry fall detection that can dial emergency services if you take a hard enough tumble and don't respond. It works reliably. It also doesn't misfire every time you spike your phone onto a couch cushion. The reason both of those things are true comes down to one tiny chip and some genuinely clever signal reading.
What the accelerometer actually measures
An accelerometer measures proper acceleration: the force acting on a mass inside the chip, expressed in g-forces. One g is normal gravity at rest. Hold your phone still on a desk and the sensor reads roughly 1 g pointing downward. Pick it up fast and the number spikes.
Drop it, though, and something strange happens.
For the fraction of a second the phone is in freefall, the sensor reads close to 0 g. Gravity pulls everything equally, so the internal mass feels no force at all. The phone is, briefly, weightless, like a marble floating inside a falling elevator.
That freefall signature is the first marker. It's not enough on its own, though. A phone sliding off a table hits near-zero g too. So does a long jump in a pickup basketball game. The accelerometer has to keep reading what happens next.
The full signature of a fall
A genuine fall has a three-part shape in the data.
First, freefall: the signal drops toward 0 g, usually for somewhere between 0.3 and 0.8 seconds depending on the height. Second, impact: the signal spikes hard, often exceeding 3 g and sometimes hitting 6 or 7 g on a hard floor. Third, post-impact stillness. A person who has just fallen tends to move very little immediately after hitting the ground.
That last part is what separates a fall from a sharp downward gesture. Say you're conducting an orchestra and you throw your arm down hard on the downbeat. The accelerometer sees a fast movement and maybe a small impact as your arm decelerates, but then your arm keeps moving, adjusting, coming back up. The signal stays dynamic. No freefall plateau, no hard spike, no sudden stillness.
A phone tossed onto a bed shows a gentle impact and then bouncing oscillations. A phone dropped onto concrete shows the classic three-part pattern but then often keeps skittering across the floor, which is exactly why the post-impact stillness check matters: it's looking for a person who has stopped moving, not a device that's rolling away.
Some implementations add gyroscope data to the mix. The gyroscope measures rotational velocity, and a real fall almost always involves the body rotating in a way a deliberate gesture doesn't. Arms flail. Torsos twist. The combination of linear and rotational data tightens the accuracy considerably.
Where it gets genuinely hard
Consider two people, Maya and Rashid, both in their seventies, both carrying the same phone model. Maya trips on a step and falls sideways onto a hardwood floor. Rashid sits down heavily in a chair after feeling dizzy. Both events look, to the accelerometer, surprisingly similar: rapid downward acceleration, an impact, then stillness.
This is the honest difficulty, and anyone who tells you the algorithm has fully solved it is selling something.
Sitting down hard can produce an impact exceeding 2 g. The freefall phase is shorter but not absent. The post-impact stillness is the same. Engineers deal with this partly by setting thresholds (the freefall window needs to last at least a certain duration, the impact needs to cross a certain g-value) and partly by adding a human prompt: the phone vibrates and displays a screen asking if you're okay. No response within about 30 seconds, and it calls for help. The absence of a human action is doing real work inside the detection logic.
That framing raises a question that deserves a direct answer: what are you actually trusting when you turn this feature on?
Not a single yes-or-no sensor reading. The phone runs roughly 1 to 2 seconds of continuous signal through a pattern trained on thousands of fall recordings. The threshold tuning is a deliberate tradeoff, and getting it wrong in either direction costs something real. Too sensitive and anxious texts go out every time someone sits down hard. Too conservative and the feature fails the exact people it was built for. That's not a minor calibration question. It's an ethical one engineers are quietly making every time they push an update.
The accelerometer in your pocket samples between 100 and 200 times per second. It sees everything. The art is in knowing which everything matters.