AI Trained on Four Million Fossils Is Rewriting the Story of Mass Extinctions

A New Kind of Fossil Record

For generations, scientists have painted Earth's five great mass extinctions as sudden, catastrophic events — moments when life on our planet was nearly snuffed out in geological eyeblinks. A sweeping new analysis is now challenging that picture in a profound way. Paleontologists have trained an AI model — similar in design to the large language models that power tools like ChatGPT — on fossil records instead of words, then used it to reconstruct a timeline of changes in Earth's biodiversity over hundreds of millions of years in unprecedented detail.

Developers used data from a massive geological compilation known as OneStratigraphy to train the machine-learning model on statistical patterns within the fossil record. Instead of predicting text, the resulting system — called OneStrata — predicts when specific fossils were likely deposited in sedimentary rock. The result is the most comprehensive global fossil record ever assembled, and what it reveals is surprising: some of history's most famous extinction events may not have been the sudden catastrophes we long believed them to be.

Slow Burns, Not Sudden Crashes

This big-data approach to interpreting the rock record — the field known as stratigraphy — could spur a rethink of major upheavals in the history of life. The model suggests that three of the "Big Five" mass-extinction events were gradual declines in biodiversity, rather than the sudden cataclysms described by other reconstructions. That's a striking departure from the dominant narrative.

The "Big Five" extinction events are the End Ordovician extinction about 445 million years ago, the Late Devonian about 375 million years ago, the End Permian event 251 million years ago, the End Triassic roughly 200 million years ago, and the End Cretaceous event 65 million years ago — which marks the end of the dinosaurs. The new model suggests that at least three of these iconic events unfolded not as sharp breaks in the history of life, but as prolonged, creeping collapses. The OneStrata system's high-resolution data indicate that three major extinction crises actually unfolded as slow reductions in species counts, rather than the abrupt disasters portrayed by alternative models.

To make the fossil record manageable, stratigraphers have traditionally divided fossil sequences for each species into time intervals of millions of years — making it easier to assemble many sequences into a unified whole, but producing a blurry view that obscures changes in biodiversity occurring over shorter time intervals. The OneStrata model sidesteps this limitation by learning patterns directly from the data, achieving a resolution that older methods simply couldn't reach.

A Computational Leap Forward

Previous efforts using supercomputers successfully pieced together thousands of fossil sequences to estimate species appearance and extinction with a resolution of 26,000 years across 11,000 marine species — but scaling that method to include more species at shorter intervals required an impractical amount of computational resources. The new machine-learning framework bypasses that limit by relying on statistical patterns learned during training. In other words, AI accomplished what raw computing power alone could not.

The model is described as "a real revolution for studying stratigraphy and past diversity change" by Shuzhong Shen, a paleontologist at Nanjing University in China, who presented the research at a meeting of the Geological Society of America in Denver, Colorado, on October 12. Not everyone is ready to declare the old picture dead, however. While the OneStrata reconstruction appears to be the most detailed global record so far, some scientists question its necessity for every application — with Matthew Clapham, a paleontologist at the University of California, Santa Cruz, noting that researchers already possess high-resolution reconstructions of biodiversity changes at specific locations that effectively illuminate past extinction events.

Why It Matters Beyond the Deep Past

The stakes of this debate extend well beyond academic paleontology. We are currently in the beginning stages of the so-called "sixth mass extinction," with rates of species loss that are frighteningly high, even when compared with the highest rates measured in the fossil record over the last half billion years. Understanding whether past extinctions were sudden shocks or slow declines could reshape how scientists model — and potentially anticipate — the biodiversity crisis unfolding today.

If three of the Big Five were gradual processes rather than overnight disasters, it raises an unsettling possibility: a mass extinction can be well underway before it looks like one. The OneStrata model is still being scrutinized by the broader scientific community, and debate over its findings will likely intensify. But its core contribution — bringing the resolution of AI to the blurry fossil record — represents a new frontier in how humanity reads the deep story of life on Earth, and perhaps how it writes the next chapter.