The End of Guesswork: How AI is Solving Paleontology’s Muddy Identity Crisis

For years, a set of tracks on Scotland’s Isle of Skye served as a silent battleground for paleontologists. Some experts looked at the weathered indentations and saw the heavy, three-toed gait of a carnivorous theropod. Others saw the softer, broader stride of an early plant-eater. This wasn't just a difference of opinion; it was a symptom of a deeper rot in the field of ichnology. When your primary data is a 170-million-year-old hole in the mud, human bias tends to fill the gaps that erosion left behind.

Now, an AI called DinoTracker has settled the score. By removing the "human" from the equation, the tool re-examined the Isle of Skye tracks and aligned them firmly with herbivorous species. It’s part of a broader shift in how we read the fossil record, turning subjective interpretations into hard, reproducible data.

Seeing Through the Mud

Paleontology has always had a "numbers problem." A single dinosaur left exactly one skeleton, but it could leave hundreds of thousands of footprints over its lifetime. Despite this abundance, trace fossils are notoriously messy. The consistency of the ancient mud, the angle of a slope, and the weight of the animal all conspire to distort a footprint until it is barely recognizable.

To cut through this noise, Gregor Hartmann of Helmholtz-Zentrum Berlin and Stephen Brusatte of the University of Edinburgh bypassed traditional machine learning. Most AI models are "supervised," meaning humans tell the computer what a "good" sample looks like. The problem? Human labeling bakes in existing assumptions. If a researcher thinks a track looks like a T-Rex, they label it as such, and the AI simply learns to mimic that human hunch.

DinoTracker uses an unsupervised neural network—essentially a Variational Autoencoder (VAE)—to find its own logic. It analyzed nearly 2,000 real fossils and millions of synthetic variations to map out a "shape space." Instead of looking for a "toe," the AI looks for clusters of geometric commonalities across thousands of dimensions. It identified eight key features of footprint variation—ranging from the precise spread of the digits to the specific pressure distribution of the heel—without any human telling it what to look for. The result is a system that matches expert consensus up to 93% of the time but does so with a cold, mathematical consistency.

The 200-Million-Year Stride

This isn’t just about sorting old photos; it’s about settling decades-old academic feuds and tracing the very blueprint of avian life. When the AI processed tracks dating back over 200 million years, it didn't just see dinosaur feet—it saw the origin of the modern bird.

The data revealed that the mechanical way these ancient reptiles interacted with the earth is nearly identical to the foot structures of modern birds. This isn't a new theory, but DinoTracker provides the first quantitative bridge. It shows that the "avian" way of walking wasn't a late-stage evolutionary perk; the geometry of the stride has remained remarkably stable for hundreds of millions of years. The AI effectively proved that the movement patterns of a Triassic predator are still alive today in the strut of a common crow.

Paleontology for the Masses

The team behind DinoTracker is now moving the technology out of the lab and into the pockets of the public. By releasing DinoTracker as a free app, they are effectively crowdsourcing the future of the field.

Processing the Backlog

In the past, massive sites like the Cal Orck’o cliffs in Bolivia—home to some 16,000 theropod tracks—represented a logistical nightmare. For a human team, measuring and categorizing that many prints would take decades of grueling manual labor. The AI can process those same sites in a weekend. This speed allows researchers to move past individual footprints and start looking at the "big data" of the prehistoric world: migration patterns, herd dynamics, and predator-prey ratios across entire continents.

Removing the Expert Filter

By focusing on "shape space" rather than rigid labels, the tool identifies the "weird" fossils—the outliers that fall between traditional categories. These are often the most important discoveries, representing transitional species that human researchers might have overlooked or forced into an existing label to avoid controversy.

As Gregor Hartmann puts it, the tool provides a "mathematical point of view." It doesn't replace the paleontologist; it just stops them from seeing what they want to see in the mud. For a field that has long relied on the "trained eye," the arrival of an unbiased lens is a long-overdue reality check.