When an earthquake strikes the West Coast, sensors across the region feed the live data into a computer system called ShakeAlert. Within seconds, an algorithm decides whether to send out a warning via text messages and emergency alerts.

The problem, according to a new paper published in the journal Science, is that it’s impossible to say at an earthquake’s outset how severe it will become. That means that the designers of a warning system like ShakeAlert — or its international equivalents in Japan and Mexico — need to make a difficult decision: to either send out more alerts for tremors that are likely to be minor, or send rarer alerts that might come just seconds before a major quake.

Even a few seconds of warning can can make a difference, according to Annmarie Baltay, a geophysicist at the U.S. Geological Survey and one of the authors of the paper.

“For a person who might get an alert on their cell phone, or a school child in a classroom with a siren, they could get quite a few seconds to drop, cover, hold on, get under a desk,” Baltay told The Outline. “These are quick actions that could prevent casualties.”

The danger, though, is that frequent alerts for quakes that don’t turn out to be deadly could make residents of earthquake-prone areas complacent in the face of future warnings. In their next paper, Baltay’s team intends to study that tradeoff.

The danger isn’t theoretical. Researchers agree that a mega-quake is likely to strike California sooner or later — and in the case of the 2014 quake that struck San Francisco in 2014, ShakeAlert provided just six seconds of warning.