The Seconds Before The Shaking

AAn earthquake announces itself twice. The rupture first sends out a compressional wave that travels through rock at roughly six kilometres a second and does very little damage, and this is followed by slower shear and surface waves which between them account for almost all destruction. A dense network of instruments near the fault can register the first arrival, work out where the rupture began and broadcast a message over fibre and radio at the speed of light. The interval this buys is not generous. A city two hundred kilometres from the epicentre may gain something like fifty seconds; one at fifty kilometres may gain ten. Public systems built on the principle have operated in Japan since 2007 and now cover parts of Mexico, Taiwan and the western United States. Their designers have never claimed to be predicting earthquakes, and the distinction is not a quibble: what is being announced is not a future event but the arrival, somewhere else, of shaking that has already begun.

BHalvard Rennie, a seismologist at the Tarrant Geophysical Observatory, has spent a decade arguing that the published figures flatter the technology. Between 2015 and 2022 his group reconstructed the timing of 214 earthquakes of magnitude 5.0 and above, setting the moment at which each alert was issued against the moment heavy shaking reached every settlement in the affected area. Inside a radius he puts at about twenty-five kilometres from the epicentre — the blind zone — the message arrives after the shaking, or never arrives at all. The awkward feature of that radius, he observes, is that it encloses the ground on which buildings actually fall. In his sample the districts that recorded the heaviest damage were given a median of two seconds of warning, while districts that escaped structural damage altogether were given a median of nineteen. The alert, as he puts it, is delivered in inverse proportion to the need for it.

CA second objection concerns how much the opening seconds can be made to reveal. Magnitude is estimated from the first part of the wave, yet a large rupture goes on growing for tens of seconds after it begins, so the earliest reading describes a smaller event than the one under way. Rennie reports that in 31 of the 46 events above magnitude 6.0 in his sample the magnitude first broadcast was at least 0.4 below the value eventually assigned, and that in four cases the shortfall passed a full unit. The consequence is not merely arithmetical. Because most operators hold back alerts below a set threshold, an underestimate at the outset can delay or cancel a warning for exactly the events that matter most. Rennie is careful, even so, to present the finding as a defect in one method of estimation rather than as an argument against warning systems in general.

DInes Machado, who directs the Vallena Institute for Hazard Response, regards the whole line of criticism as misdirected, on the ground that a few seconds are worth whatever their recipients have been trained to do with them. Her group questioned 2,640 employees in three cities after moderate earthquakes between 2019 and 2023. Where the workplace had run a drill within the previous six months, 68 per cent reported having taken cover before the shaking reached them; where no drill had been held the figure was 22 per cent. Machado stresses that the advantage was confined to buildings in which the alert was carried by a loudspeaker, since many staff had silenced the handsets on which it would otherwise have arrived. She concedes two weaknesses. The behaviour was reported by the workers themselves rather than filmed, and her respondents were office staff in a single sector, so nothing follows from her figures about schools, hospitals or the street.

EThe argument has consequences well beyond the seminar room. An automatic response needs no comprehension and can use a signal too brief for any person: lifts travel to the nearest floor and open, high-speed trains begin to brake, valves shut on gas mains, and in a few hospitals surgical equipment is halted. One national rail operator reported in 2022 that braking triggered by alerts had cost it 41 unnecessary stoppages in a single year, an expense the company judged acceptable. Machado points out, uncomfortably for her own emphasis on training, that most of the benefit documented so far has gone to machinery rather than to people. Rennie replies that this is precisely his case, since the blind zone matters less to a valve than to a commuter: a valve has nothing to understand first. Both treat the habit of publicising the seconds gained in one dramatic event, instead of the spread of outcomes across all events, as having damaged public understanding.

FWhat is emerging is less a verdict than a division of labour between the two accounts. Where an action can be carried out by a machine, two or three seconds are enough and the alert is worth issuing even close to the fault; where an action requires a person to interpret a sound, decide and then move, the useful threshold rises to something nearer ten seconds, which in practice restricts the gain to populations at a distance from the rupture. Neither researcher now defends the position attributed to each of them a decade ago, when the disagreement was reported as a quarrel over whether such systems worked at all. The question that stays open is whether warnings reduce injuries when a whole population is counted, and it will not be settled soon, because hospitals record injuries in categories that do not distinguish those who moved from those who did not.