A look at AI-related research for predicting earthquakes by using seismic data, which scientists say is similar to the audio data used to train voice assistants
SAN FRANCISCO — Countless dollars and entire scientific careers have been dedicated to predicting where and when the next big earthquake will strike. Tweets: @scottnations and @nytimes Tweets: Scott Nations / @scottnations : If AI can predict earthquakes it can predict stock market crashes. The science of the two is surprisingly similar. Both start with a “nucleation”. If friction overcomes tension you have a tremor or a correction. If not you get a quake or crash. http://www.nytimes.com/... @nytimes : Scientists are now analyzing seismic data 500 times faster thanks to artificial intelligence. Some hope these experiments will help predict future earthquakes. http://www.nytimes.com/...
Context & Ripple Effects
The 2018 piece captures an early moment in a now-familiar migration: machine-learning methods built for human speech being pointed at geophysics, with scientists reporting AI analyzes seismic data 500 times faster by treating quake signals like the audio streams behind voice assistants. The same transfer logic soon showed up elsewhere — researchers applying self-supervised learning to animal sounds, and AI pipelines built for ocean data including whale-song identification.
The arc since then has two edges. On one side, the sensing idea scaled to consumers: [[a:888121|Google's Android early-warning system has issued alerts for more than 1,200 earthquakes since 2021]]. On the other, the hype risk was already visible in One Concern's exaggerated claims about its disaster-response AI — a cautionary benchmark for anyone selling predictive quake tools.
First-order effects
- Seismologists gain a processing pipeline borrowed from voice-assistant research, cutting seismic-data analysis time dramatically in experimental settings and making continuous monitoring of fault zones computationally cheap.
- Earthquake research stops requiring bespoke models: techniques already trained at consumer scale on audio can be repurposed, shifting the bottleneck from algorithm development to labeled seismic data.
Second-order effects
- Consumer hardware becomes scientific infrastructure — the path later validated when Google turned hundreds of millions of Android phones into an earthquake-detection network issuing seconds-level alerts.
- Commercial entrants face a credibility test set by One Concern, whose overstated disaster-AI capabilities made funders, agencies, and journalists far more skeptical of predictive-safety claims.
Third-order effects
- If the transfer pattern holds, the boundary between consumer tech and earth science keeps dissolving: phones, oceans, and fault lines all become continuously recorded, machine-readable datasets feeding the same model families.
- Disaster prediction sits on a credibility spectrum — genuine detection advances like Android's alerts coexist with overhyped tools — pushing regulators and emergency agencies toward independent validation before AI predictions drive evacuations or insurance pricing.
The trend: Machine-learning methods built for human language and audio are being ported across the natural sciences, turning phones, oceans, and fault lines into trainable sensor networks.