Earthquake Early Warning: What the Evidence Shows
Key takeaways · 16 min read
- Even an ideal, delay-free warning system could give no location more than about 7 seconds of warning before very strong shaking in a modelled magnitude 8 earthquake.
- For light shaking the same model gave almost 90% of locations some warning and 17% more than a minute: the weakest shaking gets the most notice.
- The US Geological Survey says a zone roughly 25 km around the epicentre currently cannot be warned before strong shaking starts.
- In Japan’s 2011 Tohoku earthquake, warnings reached the Tohoku district more than 15 seconds before strong shaking, but surveys find most people who receive alerts do not take physical protective action.
Earthquake early warning is not prediction. Nobody can say that a fault will break tomorrow. What these systems do is detect an earthquake that has already started, estimate how big it is and how hard it will shake each place, and send an alert faster than the damaging waves travel. The gain is measured in seconds, sometimes tens of seconds, and very occasionally more than a minute.
Public systems now run in Japan, Mexico, South Korea, Taiwan and along the US West Coast, and Google’s Android network sends alerts in dozens more countries. The promise is easy to grasp: a few seconds to get under a desk, stop a train or close a valve. The question that matters for anyone deciding how much to rely on an alert is how many seconds, and for whom.
This article looks at what the research says about warning times, why the places that shake hardest usually get the least notice, how real systems have performed in Japan, California and Türkiye, what people actually do when their phone goes off, and which injuries a warning can realistically prevent.
How early warning works
An earthquake sends out two main kinds of wave. The faster P wave usually does little damage; the slower S wave and the surface waves that follow carry most of the strong shaking. Sensors near the rupture pick up the first waves, a computer estimates location and magnitude, a ground-motion model predicts shaking at each place, and an alert goes out by phone, broadcast or direct link. Radio signals travel far faster than seismic waves, which is where the seconds come from.
Each system sets thresholds. The Japan Meteorological Agency issues public warnings when it expects a maximum seismic intensity of 5-lower or more on its own intensity scale. The US ShakeAlert system, run by the US Geological Survey, sends Wireless Emergency Alerts for earthquakes of magnitude 5 or more with expected shaking of intensity IV or higher, and app alerts from magnitude 4.5 and intensity III. The Puget Sound earthquake of September 2026, at magnitude 4.2, was below both.
When a public alert goes out
Public alerting thresholds for three early warning systems.
| System | Public alert threshold | Delivery |
|---|---|---|
| Japan Meteorological Agency | Expected maximum intensity 5-lower or more (JMA scale) | TV, radio, mobile phones |
| ShakeAlert, US West Coast | Magnitude 5+ and expected intensity IV+ | Wireless Emergency Alerts |
| ShakeAlert, US West Coast | Magnitude 4.5+ and expected intensity III+ | Phone apps and operating-system alerts |
| Android Earthquake Alerts | Magnitude 4.5+ (alerts in 98 countries) | Android phones |
Sources: Japan Meteorological Agency, earthquake and tsunami information leaflet; Kohler MD et al., Seismological Research Letters (2020), doi:10.1785/0220190245; Allen RM et al., Science (2025), doi:10.1126/science.ads4779.
The physics of the deadline
The key study on what is physically possible was published in Science Advances in 2018 by Sarah Minson and colleagues at the USGS. They imagined a perfect system with no delays in data, computing or delivery. Even then, there is a hard limit: a system cannot know that shaking will be strong far away until the earthquake has grown large enough to produce it, and big earthquakes take seconds to grow. By the time that is clear, the strong waves are already on their way.
Their worked example was a magnitude 8 rupture starting off northern California and running south towards San Francisco, the kind of event often used to sell early warning. If San Francisco wanted alerts only for very strong shaking, about 20% of gravity, the ideal system could give about 8 seconds. Accepting alerts for weaker levels bought more time: 19 seconds at 10%g, 33 seconds at 5%g and 48 seconds at 2%g.
The stronger the shaking, the shorter the notice
Best-case warning time for San Francisco in a modelled magnitude 8 rupture, by alert threshold (seconds).
Source: Minson SE et al., Science Advances 4(3):eaaq0504 (2018), doi:10.1126/sciadv.aaq0504. Assumes a delay-free system that knows the rupture perfectly; real systems add seconds of latency.
Across a generic magnitude 8 scenario, the pattern held everywhere. Almost 90% of locations could receive some warning for light shaking, and 17% more than a minute. For very strong shaking, no location received more than 7 seconds. The authors concluded that early warning is most useful to people willing to be alerted at lower shaking levels, accepting many alerts for earthquakes that never grow large, in exchange for a head start when one does.
Long warnings are for light shaking
Share of locations receiving warning in a modelled magnitude 8 strike-slip rupture, ideal delay-free system.
Source: Minson SE et al., Science Advances (2018). A theoretical upper bound, not a measured performance.
The blind zone
Close to the epicentre, the strong waves arrive before any alert can. The USGS calls this the blind zone and says current networks and communications leave a radius of about 25 km, or 15 miles, around the epicentre where alerts before strong shaking are not possible. The Japan Meteorological Agency gives the same warning in its public guidance: in areas close to the focus, especially for shallow inland earthquakes, the warning may not arrive before strong tremors hit.
This is the uncomfortable centre of the subject. The people nearest the rupture usually face the most violent shaking and the greatest risk of building damage, and they are the ones least likely to get any warning. The USGS puts it plainly: substantive warning times occur mainly at relatively low shaking levels. Early warning works best for people some distance away, who are often less at risk. That does not make it useless, but it does mean an alert is a bonus, not a plan.
Japan and the 2011 earthquake
Japan has the longest-running nationwide public system, issuing warnings to the public since October 2007. In its first years, nearly 80% of the agency’s intensity predictions were within one unit of what was observed, though there were cases where warnings did not arrive in time for strong shaking. The real test came on 11 March 2011, with the magnitude 9.0 Tohoku earthquake.
According to an analysis by Mitsuyuki Hoshiba and colleagues at the agency, the public warning reached the Tohoku district before the S waves and more than 15 seconds before strong shaking of intensity 5-lower or greater everywhere in the district. Other researchers estimated that several million people near the epicentre got about 15 to 20 seconds before the most severe shaking. Because the rupture started far offshore, this was a favourable case for warning time.
The system also missed things. For Tokyo it predicted intensity 4, but the city recorded 5-upper, an underestimate the authors put down to the enormous length of the fault rupture. Point-source methods struggle to size a rupture hundreds of kilometres long while it is still breaking. After 2011 the agency added methods that predict shaking directly from nearby observed ground motion, rather than relying only on a single estimate of location and magnitude.
ShakeAlert in practice
ShakeAlert began statewide public alerting in California in October 2019 and later extended it to Oregon and Washington. Between October 2019 and September 2023 it created 95 events with magnitude estimates of 4.5 or more, the public threshold, and 94 of them were real earthquakes. The threshold was deliberately set below the level where damage is likely, to allow for the system underestimating magnitude.
Producing an alert is not the same as people receiving it in time. A study of the USGS “Did You Feel It?” surveys found many reports of shorter warning times, late alerts and missed alerts than ShakeAlert’s publication times implied, with delivery delays that were substantial and highly variable. The MyShake app measures its own delivery: across many events, the median of its 20th and 50th percentile delays was about 2 and 3 seconds.
When it works, the reach is large. On 24 June 2026, a magnitude 5.6 earthquake east of Willits in northern California prompted advance alerts to more than a million people, according to the state government. For smaller urban earthquakes directly beneath the population, an analysis of MyShake found that about half of recipients got the alert before the estimated arrival of the S wave, and up to 90% before peak shaking.
Phones as seismometers
Most earthquake-prone countries have no dense seismic network. Google’s Android Earthquake Alerts system uses the accelerometers in ordinary phones instead. A 2025 paper in Science by Richard Allen and Google colleagues reported three years of operation: an average of 312 earthquakes detected a month, alerts in 98 countries, and about 18 million alerts a month. Of 1.56 million people who answered a feedback survey, 36% received the alert before shaking, 28% during and 23% after.
Before, during or after?
When Android alert recipients said the alert arrived, relative to the shaking they felt (share of survey respondents).
Source: Allen RM et al., Global earthquake detection and warning using Android phones, Science (2025), doi:10.1126/science.ads4779. Self-reported by 1.56 million respondents; the percentages do not sum to 100.
The same paper was frank about a failure. In the magnitude 7.8 earthquake in Türkiye on 6 February 2023, the system’s estimate peaked at only magnitude 4.9, and it sent lower-level alerts to about 512,000 phones within 64 km. Replaying the event with the improved algorithm produced a magnitude 7.4 estimate and the stronger alert for 10 million phones within 158 km, with up to 35 seconds of warning. The improvement came after the deaths, not before.
Türkiye, February 2023: what happened and what could have
Android Earthquake Alerts in the magnitude 7.8 earthquake, as operated and as replayed with the later algorithm.
| Measure | On the day | Replay, improved algorithm |
|---|---|---|
| Highest magnitude estimate | 4.9 | 7.4 |
| Strongest alert sent | Lower-level alerts to 512,411 phones within 64 km | Take-action alerts to 10 million phones within 158 km, up to 35 seconds ahead |
Source: Allen RM et al., Science (2025). A separate smartphone app, Earthquake Network, reported warnings of up to 58 seconds for some users who went on to feel intensity IX or above (Finazzi F et al., Scientific Reports, 2024).
A later test went better. In the magnitude 6.2 earthquake off Marmara Ereğlisi, near Istanbul, in April 2025, the system detected the event 5.3 seconds after it began and alerted more than 16 million users. Warning times for light shaking reached up to 150 seconds, with a median of 56. Near the source, the paper notes, warning windows were shorter, the same pattern the physics predicts.
What people do with the seconds
The recommended response in the United States and many other places is drop, cover and hold on: get down, take cover under a sturdy table, and hold on until the shaking stops. Early surveys after 2011 were very positive; one study reported that some 90% of people who received the warning were able to take some action. Later, more careful surveys in Japan tell a more sobering story.
Kazuya Nakayachi and colleagues surveyed people who received two Japanese warnings in 2018. Most stayed where they were, not for safety reasons but because they were bracing themselves mentally. A 2022 national survey of 491 people who had received warnings found that only 35% took concrete protective action. Among Android users who received the stronger take-action alert, 28% reported dropping, covering and holding on. A 2025 review concluded that taking the recommended action is not a common response.
Warned is not the same as protected
What surveys found people did after receiving an earthquake early warning.
Japan, 2022 national survey: 35% took concrete action
Android take-action alert recipients: 28% dropped, covered and held on
Japan 2018 surveys: most people stayed still and braced mentally
Android recipients: 30% shared the alert with others
Israel, false alert to over 1 million people in 2024: public remained largely willing to follow guidance
Alerts that arrive early raise trust and use next time
Sources: Ohtomo S et al., Journal of Disaster Research (2026), doi:10.20965/jdr.2026.p0024; Allen RM et al., Science (2025); Nakayachi K et al., Risk Analysis (2019), doi:10.1111/risa.13306; Yagoda-Biran G et al., Scientific Reports (2026); Mousavi SM et al. (2026), on the 2025 Marmara Ereğlisi earthquake.
False alarms are a real cost but seem to be tolerated. In October 2024 Israel’s system mistook a large military explosion in southern Lebanon for a magnitude 5.2 earthquake and alerted more than a million people. A survey of 1,043 residents afterwards found considerable tolerance for false alerts and strong support for being warned even of non-damaging, felt earthquakes, though tolerance was lower than for false missile alerts.
What injuries a warning can prevent
The answer depends on what hurts people. In the 1994 Northridge earthquake in Los Angeles, a study of coroner and hospital records found 171 serious injuries in Los Angeles County: 33 deaths and 138 hospital admissions. Most deaths were caused by building collapse. Most hospital admissions were caused by falls or by being hit by objects. In the 1990 Luzon earthquake in the Philippines, falling objects were the leading cause of injury, at 34%.
Those are the injuries a few seconds can plausibly prevent: getting down before you are thrown, moving away from a tall bookcase or a window, taking cover from things falling off shelves. A seismologist, Max Wyss, has pointed out the limit: taking cover under a table protects against falling objects, not against a collapsing building. Early warning cannot save people inside a building that fails. Only stronger construction does that.
Automated actions may matter more than human ones. Trains can brake, lifts can stop at the nearest floor, and gas valves and factory lines can shut without anyone deciding anything. These need no reaction time from a person and can use even two or three seconds. For households, agencies such as the Japan Meteorological Agency stress the things done long before any alert: fixing furniture to walls, planning where to take cover, and practising it.
What the evidence suggests
Earthquake early warning works, within limits that are set by physics rather than by engineering. It reliably buys tens of seconds for light and moderate shaking at a distance, and only a few seconds, or none, for the most violent shaking close to the fault. Real systems add delivery delays, and they can underestimate the largest earthquakes, as Türkiye showed. Even when an alert arrives, many people do not act on it. The seconds are most valuable for machines, and for people who have already decided what they will do.
What to do
Steps sorted by when they matter. Each one says how strong the evidence is: Measured Official advice Untested
Before
- Where public alerts exist, as with ShakeAlert and the MyShake app on the US West Coast, set them up on your phone. Official advice see Sources
- Accept alerts for lighter shaking. That is where warning time is longest; for the most violent shaking, no location got more than 7 seconds. Measured see Sources
- Decide in advance what you will do when an alert arrives. The seconds are most valuable to people who already know. Measured see Sources
What not to do
- Do not wait for an alert before reacting to shaking. Delivery delays are substantial, and alerts can be late or missed. Measured see Sources
Questions people ask
Can earthquake early warning predict an earthquake?
No. It detects an earthquake that has already begun and races the seismic waves. Nobody can currently predict when a specific earthquake will happen.
How many seconds of warning will I get?
It depends on distance and on how strong the shaking will be. Far from the epicentre it can be tens of seconds; within roughly 25 km, often none before strong shaking.
Why did I feel an earthquake but get no alert?
It may have been below the alert threshold, as in Puget Sound in September 2026, or you may have been too close to the epicentre for an alert to arrive first.
What should I do when I get an alert?
Follow your local official advice. In the United States that is usually drop, cover and hold on; in Japan, the agency advises staying calm and securing your personal safety.
Is a phone-based system as good as a seismic network?
It can be fast where phones are dense, as in Istanbul in 2025, but it badly underestimated the 2023 Türkiye earthquake. Performance depends on the algorithm and the event.
The short version
- Even an ideal, delay-free warning system could give no location more than about 7 seconds of warning before very strong shaking in a modelled magnitude 8 earthquake.
- For light shaking the same model gave almost 90% of locations some warning and 17% more than a minute: the weakest shaking gets the most notice.
- The US Geological Survey says a zone roughly 25 km around the epicentre currently cannot be warned before strong shaking starts.
- In Japan’s 2011 Tohoku earthquake, warnings reached the Tohoku district more than 15 seconds before strong shaking, but surveys find most people who receive alerts do not take physical protective action.
- A warning can help prevent falls and injuries from falling objects; it cannot protect people in a building that collapses.
This article summarises published research and agency guidance for general information. It is not safety or engineering advice for any particular building. Earthquake early warning is not available everywhere and may arrive late or not at all. Follow official alerts and advice from the US Geological Survey and ShakeAlert partners, the Japan Meteorological Agency, or your national and local authorities, and plan where you will take cover before an earthquake happens.
Further reading: Minson et al., Science Advances 4(3) (2018), on the physical limits of warning time. Allen et al., Science (2025), on the global Android system, including its failure in Türkiye. Nakayachi et al., Risk Analysis (2019), on how people in Japan actually respond to warnings.
- The Great Quake Debate, Susan Elizabeth Hough (2020). A USGS seismologist on the early arguments over earthquake risk in California and how modern seismology grew from them. Good history; it says little about early warning itself.
- Disasterology, Samantha Montano (2021). An emergency management researcher on how societies prepare for and recover from disasters. Strong on why plans fail; earthquakes are one hazard among many and get limited space.
- Doom: The Politics of Catastrophe, Niall Ferguson (2021). A historian’s sweep through earthquakes, pandemics and other disasters and why governments handle them badly. Wide-ranging and readable; its arguments are opinionated and not all specialists accept them.
Sources
Minson SE, Meier M-A, Baltay AS, Hanks TC, Cochran ES. The limits of earthquake early warning: timeliness of ground motion estimates. Science Advances 4(3):eaaq0504 (2018), doi:10.1126/sciadv.aaq0504. — US Geological Survey. Blind zone in earthquake early warning (May 2020). — Japan Meteorological Agency. Earthquakes and tsunamis: disaster prevention and mitigation efforts (leaflet, accessed October 2026). — Kohler MD et al. Seismological Research Letters (2020), doi:10.1785/0220190245. — Allen RM, Melgar D. Annual Review of Earth and Planetary Sciences (2019), doi:10.1146/annurev-earth-053018-060457. — Doi K. Soil Dynamics and Earthquake Engineering (2011), doi:10.1016/j.soildyn.2010.06.009. — Hoshiba M et al. Earth, Planets and Space 63 (2011), doi:10.5047/eps.2011.05.031. — Kodera Y et al. Frontiers in Earth Science (2021), doi:10.3389/feart.2021.726045. — Fujinawa Y, Noda Y. Earthquake Spectra (2013), doi:10.1193/1.4000127. — Lux AI et al. Bulletin of the Seismological Society of America (2024), doi:10.1785/0120230259. — Saunders JK, Wald DJ. The Seismic Record 5(2):239 (2025). — Marcou S et al. Seismological Research Letters (2025), doi:10.1785/0220250037. — Patel SC, Allen RM. Seismological Research Letters (2022), doi:10.1785/0220220062. — Office of the Governor of California. Press release, 24 June 2026. — Allen RM et al. Global earthquake detection and warning using Android phones. Science (2025), doi:10.1126/science.ads4779. — Finazzi F, Bossu R, Cotton F. Scientific Reports (2024), doi:10.1038/s41598-024-55279-z. — Mousavi SM et al. Performance and user response of Android’s smartphone-based alerts in the 2025 Marmara Ereğlisi earthquake. Nature Communications (2026). — Nakayachi K, Becker JS, Potter SH, Dixon M. Risk Analysis (2019), doi:10.1111/risa.13306. — Ohtomo S et al. Journal of Disaster Research (2026), doi:10.20965/jdr.2026.p0024. — Goltz JD, Nakayachi K, Vinnell L. Natural Hazards Review (2025), doi:10.1061/NHREFO.NHENG-2462. — Yagoda-Biran G, Nof RN, Zwebner Y. Scientific Reports (2026), doi:10.1038/s41598-026-37958-1. — Peek-Asa C et al. International Journal of Epidemiology 27(3):459 (1998), doi:10.1093/ije/27.3.459. — Roces MC et al. Bulletin of the World Health Organization (1992). — Wyss M. Natural Hazards (2012), doi:10.1007/s11069-012-0177-6. — McBride SK et al. Geophysics, doi:10.1190/geo2021-0222.1. — KOMO News, 29 September 2026.
