Every smartphone has an accelerometer, used to rotate the screen or count steps. It is far less sensitive than a seismometer, but good enough to record the shaking of nearby moderate and large earthquakes. The MyShake project in California (Kong et al., 2016) and Google’s Android earthquake alerts system (Allen et al., 2025) showed that many phones together can detect earthquakes.

The physics of early warning

An earthquake sends out two main wave types. The faster but usually weaker P wave travels at roughly 6–7 km per second in the crust; the slower but more damaging S wave at about 3.5 km per second. If sensors near the epicentre detect the P wave and raise an alert, the alert can reach people farther away over the internet before the S wave does.

The time gained is short: a few seconds at 50 km from the epicentre, around ten seconds at 100 km. Detection, computation and delivery take several seconds, so very close to the epicentre the alert usually arrives after the shaking; this is the blind zone. Still, a few seconds can be enough to drop, cover and hold on.

The hard part: noise

A phone moves all day: in a pocket, on a desk, in a car. Sensor networks therefore usually listen only to phones that are still and charging. The phone notices a sudden change with a simple trigger that compares short- and long-term averages (STA/LTA), then a classifier decides whether it looks like an earthquake or human motion. One phone triggering means nothing; the server checks whether many phones in one area trigger in an order that fits the speed of the P wave.

Estimating the magnitude early and correctly is the hardest part. Google’s 2025 study reported that during the 6 February 2023 Pazarcık earthquake in Turkey its system greatly underestimated the magnitude, so its strongest alert reached very few people. Large ruptures last tens of seconds, and the first seconds of data say little about the final size.

Deprem’s station mode

The Deprem iPhone app (coming soon to the App Store) has a station mode:

  • It runs only when you turn it on, while the iPhone is charging and still; the phone samples acceleration 50 times a second.
  • The signal is processed on the phone: gravity is removed, the 1–20 Hz band is filtered, and an STA/LTA trigger and an on-device classifier run.
  • The server receives no raw sensor data, only a short summary: the trigger time, an area of about 5 km² and the peak acceleration. Details are in our privacy policy.

Honest limits:

  • We run in shadow mode. The server records detections and compares them with the AFAD, Kandilli and EMSC catalogues, but sends no alerts. Before alerting we must measure how accurate the method is.
  • iOS does not let apps read the accelerometer continuously in the background. An iPhone is therefore a sensor only while the app is open in station mode; most iPhone users are alert receivers, not sensors. Android allows continuous listening through a foreground service, so a future Android app will be the backbone of the network.
  • A useful network needs many phones. Without enough stations in an area, nothing can be detected.

Keep in mind: Station mode does not send alerts today. Do not rely on it for warnings; follow AFAD for official information.

Community detection is experimental; AFAD is the official authority for earthquake information and warnings in Turkey. What you can use today are the earthquake alerts based on the expected shaking at your saved places, and the live map.