AI App Matches Dinosaur Tracks to Species With 90% Accuracy

When physicist Gregor Hartmann read a dinosaur book to his son at bedtime, he realized his AI research might solve a 200-million-year-old mystery. He contacted the book's author, paleontologist Steve Brusatte, and together they built DinoTracker, an app that identifies which dinosaur made which footprint.

The challenge is tricky because you never find a footprint next to the dinosaur that made it. For decades, scientists have classified tracks manually, sometimes leading to biases and mistakes based on assumptions rather than data.

DinoTracker trained on nearly 2,000 unlabeled dinosaur footprints, learning to sort them by eight key characteristics like toe spread, ground contact, and heel position. After more than a year of training, the model now matches human expert classifications 90 percent of the time.

"I wish there was an app like this when I first started studying dinosaur tracks," says Brusatte, who co-authored the study published in the Proceedings of the National Academy of Sciences. The app makes identification more objective and helps experts understand variation among tracks preserved in different environments.

Human experts still confirm the AI's classifications using additional context like the fossil's age and location. The model only sees the track itself, so paleontologists provide the final verdict.

Why This Inspires

DinoTracker confirmed something controversial that scientists had observed but couldn't prove without bias: some three-toed Triassic dinosaur tracks really do look like bird tracks. "The humans studying them were correct," Brusatte explains. "It wasn't just wishful thinking, or that they were seeing a bird shape in the tracks in the same way somebody imagines the face of Jesus on a slice of toast."

This discovery raises an exciting possibility. Birds or their ancestors might have appeared nearly 60 million years earlier than we thought, pushing back one of evolution's most important timelines.

The team made DinoTracker publicly available on GitHub for anyone to use, from professional paleontologists to amateur dinosaur enthusiasts. Co-author Paige dePolo calls it "a useful research resource" that helps figure out what tracks any footprint most resembles and what features drive that similarity.

"I think AI has a bright future in paleontology," Brusatte says. AI won't become an all-knowing oracle, but it can compile and classify data in ways free from human biases, becoming a new type of paleontologist that works alongside human experts.

A bedtime story just opened a new window into ancient history.