Study applies AI to social media posts to estimate earthquake ground shaking intensity
You’ve probably seen the posts while scrolling on your phone: a shaky TikTok showing shelves emptying and lights swaying inside a convenience store, or a WHAT WAS THAT?? tweet that captures an earthquake in progress.
Researchers now show that Google’s large language model (LLM) Gemini can comb through these social media posts to provide estimates of earthquake ground shaking intensity corresponding to Modified Mercalli Intensity (MMI) Scale scores typically used to measure earthquake severity.
The MMI estimates provided by Gemini 1.5 Pro align well with data gathered by the U.S. Geological Survey’s “Did You Feel It?” dataset of public shaking reports and estimates based on seismogram data, said Harvard University researcher S. Mostafa Mousavi in his presentation at the Seismological Society of America’s 2025 Annual Meeting.
The findings suggest that ground shaking intensity scores developed by LLM could be useful in improving earthquake early warning system and rapid impact assessments for emergency responders, Mousavi said, while noting that more work needs to be done to understand potential biases in the social media data and to gather more precise location information from the posts.
In a study published in February, Mousavi and his colleagues used Gemini to process text, images and videos and even sounds like the rattling of objects, from earthquake-related posts in multiple languages to YouTube, X and TikTok.
The researchers used keyword searches to first identify these posts, focusing on posts with location identifiers. This dataset of 84 posts documented observed or experienced shaking from seven earthquakes in the United States, Taiwan, and Japan between 2011 and 2024.
The data consisted of screenshots and screen recordings, which allowed the scientists to directly evaluate Gemini’s ability to extract and analyze data in the unstructured way it is presented to humans.
Mousavi and colleagues say Gemini’s ability to process multimodal data—from images to text—was important in estimating ground shaking. But they also noted a curious result: Gemini models also seemed to demonstrate a simplified understanding of the relationship between earthquake magnitude, intensity and distance.
“These findings raise intriguing questions about the extent to which Gemini’s training has led to a broader understanding of the physical world and its phenomena,” the authors noted in their study.
Mousavi said that one way to use LLM in the future to estimate shaking intensity would be “to build specialized intensity-estimator agents using labeled data such as “Did You Feel It?” data collected by USGS or [European-Mediterranean Seismological Centre], and use them in real time.”