AI Fossil Model Suggests Three of Earth’s Big Five Mass Extinctions Were More Gradual

An artificial intelligence model trained on millions of fossil records suggests that three of Earth’s “Big Five” mass extinctions may have been more gradual, or weaker, than scientists have long believed. The model, called OneStrata, was developed by a team led by palaeontologist Shuzhong Shen of Nanjing University and presented at the Geological Society of America’s annual meeting in Denver, Colorado, on 12 October 2026, Nature reported on 20 October.

According to the researchers, only two of the five classic events, at the end of the Ordovician and the end of the Permian, appear in their reconstruction as profound global collapses in marine biodiversity. Other scientists welcomed the scale of the work but questioned how much it will change the field.

What Are the ‘Big Five’ Mass Extinctions?

Palaeontologists traditionally recognise five mass extinctions in the past 540 million years or so, when a large share of species disappeared in a geologically short time:

• End-Ordovician: about 445 million years ago.
• Late Devonian: about 372 million years ago.
• End-Permian: about 252 million years ago, the largest of all.
• End-Triassic: about 201 million years ago.
• End-Cretaceous: about 66 million years ago, which wiped out the non-bird dinosaurs.

That framework, the researchers argue in their conference abstract, was built on compilations of when groups of species first and last appear, sorted into time bins millions of years long, which “cannot resolve ecological dynamics operating on timescales of thousands to tens of thousands of years”.

How Does the OneStrata AI Model Work?

The history of life is read from layers of sedimentary rock, or strata. Researchers record the depth at which each species occurs at a site, and depth corresponds to time. The difficulty is combining sequences from tens of thousands of sites and species into a single global timeline, Nature explained.

Shen’s team had previously used a supercomputer to piece together a high-resolution timeline for 11,000 marine species over about 300 million years, published in Science in 2020. But conventional methods become impractical beyond roughly 10,000 species, the abstract says. OneStrata instead uses a transformer, the same broad type of AI architecture behind large language models such as ChatGPT, to learn a shared probabilistic representation of geological time.

• Stratigraphic sections: 31,044.
• Species: 183,514.
• Fossil occurrences: 2,845,815.
• High-precision dates: 449 uranium-lead (ID-TIMS) ages used as anchors.
• Time span: about 600 million years, from the Ediacaran to the early Neogene.
• Resolution: an average of thousands of years, according to the team.

Nature described the model as having been trained on nearly four million fossils; the figures above are those given in the conference abstract.

What Did the Model Find About Mass Extinctions?

In the OneStrata reconstruction, the end-Ordovician and end-Permian events “emerge as profound global biodiversity collapses, whereas the Late Devonian, end-Triassic, and end-Cretaceous extinctions appear comparatively weak or diffuse”, the abstract says. Nature reported that the model suggests these three events were gradual declines rather than the sudden catastrophes described in other reconstructions.

Shen called the model “a real revolution for studying stratigraphy and past diversity change”, according to Nature. The team, which includes Douglas Erwin of the Santa Fe Institute, describes OneStrata as a general “temporal infrastructure” that could integrate fossil, geological and environmental data, rather than a tool limited to biodiversity.

Does This Mean the Dinosaur Extinction Was Not Sudden?

Not necessarily. The study concerns marine biodiversity measured across thousands of sites, and a pattern that looks “diffuse” at a global scale may not contradict sharp, well-dated changes recorded at individual locations. Nature noted that scientists already have high-resolution records of biodiversity change at specific sites, including many that shed light on mass extinctions.

What Do Other Scientists Say?

The response has been cautious. Matthew Clapham, a palaeontologist at the University of California, Santa Cruz, told Nature that OneStrata appears to be the most detailed global record so far, but added: “We could do some extra things with it, but I don’t know if it would be Earth shattering.”

The results were presented at a scientific meeting, and Nature’s report did not refer to a published, peer-reviewed paper describing the full method. Further scrutiny of the data and model is likely before the findings are widely accepted.

Why Does It Matter?

Understanding how quickly life collapsed in the past matters for understanding today’s biodiversity crisis. The session at which the work was presented was titled “Mass extinctions and the current biodiversity crisis: Lessons from the past to inform the future”. If some past extinctions unfolded over longer periods, it could change how scientists compare them with current rates of species loss.

The work is also an example of AI methods spreading into the natural sciences, alongside the space and neuroscience discoveries we have covered recently, such as Euclid’s first new comet and a 1-gram wireless brain microscope.

What to Watch

• Publication: whether the OneStrata method and results appear in a peer-reviewed journal.
• Independent tests: whether other groups reproduce the pattern with their own data.
• Regional records: how the global picture compares with high-resolution local studies of each extinction.
• Open data: whether the model and reconstruction are made available to other researchers.

For more, see our Science News section.

Featured image: RealGatba / Wikimedia Commons, CC0 (the Ammonite Pavement at Monmouth Beach, Lyme Regis, Dorset).

Sources

• Nature — Vast new fossil record challenges story of Earth’s ‘big five’ mass extinctions (20 October 2026).
• Geological Society of America — Abstract: OneStrata (Shen, Zhang, Leng, Fan and Erwin), GSA Connects 2026 (12 October 2026).
• Geological Society of America — Session: Mass extinctions and the current biodiversity crisis (Part I) (12 October 2026).
• Science — Fan, J.-X. et al. A high-resolution summary of Cambrian to Early Triassic marine invertebrate biodiversity, Science 367, 272–277 (2020).