What if the first solid clue that birds existed was not a bone, but a three-toed footprint stamped into ancient mud? A new study suggests that some controversial “bird-like” tracks look closer to modern and fossil bird footprints than to other dinosaur groups.
The work, led by Gregor Hartmann of Helmholtz-Zentrum Berlin für Materialien und Energie with co-authors Tone Blakesley of The Open University, Paige E. dePolo, and Stephen L. Brusatte of the University of Edinburgh, uses artificial intelligence that learns without being told the “right” answers. The team also released a tool called DinoTracker, aiming to help scientists and enthusiasts compare puzzling prints with a large database.
Why dinosaur tracks matter
Footprints are “trace fossils,” meaning they capture what an animal did rather than preserving its body. Tracks can be more common than bones, and they can hint at where dinosaurs lived, how they moved, and what their environment looked like.
But footprints are messy evidence. If you have watched a footprint soften in wet sand, you have seen how the ground can change the shape. In the fossil record, mud texture, water content, and speed can all reshape a dinosaur footprint, which makes confident ID hard.
A cleaner kind of machine learning
Machine learning is often pitched as a way to classify dinosaur footprints, but many systems need humans to label training data first. When experts disagree, those labels can quietly bake assumptions into the software.
This study uses “unsupervised” machine learning, which means the program searches for patterns without being told which prints belong to which group. Earlier work has taken a more guided approach, including a 2022 Journal of the Royal Society Interface study that trained deep learning on labeled silhouettes to separate theropod and plant-eating dinosaur tracks.
Nearly 2,000 footprints in one dataset
The researchers trained their model on 1,974 footprint silhouettes spanning many dinosaur types and also modern birds. To make the system handle real-world variation, they generated many altered versions of the same tracks, mimicking effects like compression and small edge distortions in mud.
The model compresses each footprint into a small set of underlying features, then reconstructs the footprint from that compressed “shape code.” The team also tested the system on prints it had not seen during training, a basic reality check for any tool meant to work on new finds.
Eight features that separate track types
Instead of producing a single label, the model identified eight main ways footprints differ. In plain terms, it focused on how much of the foot pressed into the ground, how widely the toes spread, how the toes connect near the base, and how the heel and left-right weight show up in the print.
After the unsupervised step, the researchers compared the model’s groupings with expert identifications from published work. Agreement ranged from about 80 to 93%, which a university news release also summarized as “around 90%” consistency with human classifications.
Bird-like tracks that could rewrite a timeline
The most controversial result involves small, three-toed tracks from the Late Triassic and Early Jurassic that look strikingly bird-like. If true birds made them, birds may have originated roughly 60 million years earlier than the oldest widely accepted bird skeletons, a major shift in the timeline for how flight evolved.
In this analysis, most of those disputed tracks cluster closer to fossil and living birds than to non-bird dinosaurs. That does not settle the argument, because footprints reflect anatomy plus the ground underfoot, but it strengthens the case that these tracks are not just random look-alikes.
On the other hand, convergence remains a real possibility, meaning non-bird dinosaurs could have evolved bird-like feet and left bird-like tracks. Mud can also stretch or narrow toe impressions in ways that fool the eye, especially in wetter surfaces. A 2023 PLOS ONE study on bird-like footprints from southern Africa explains how tracks can look avian while still leaving room for other trackmakers.
A Middle Jurassic puzzle in Scotland
The model also tackled debated three-toed tracks from Middle Jurassic rocks on the Isle of Skye, formed on the muddy edge of an ancient lagoon about 170 million years ago. Scientists have argued over whether some prints came from meat-eating theropods or from ornithopods, plant-eaters that later include duck-billed relatives.
The system placed many of the Skye tracks nearer to theropod footprints, but it also suggested that some sit closer to ornithopods. That idea connects to earlier work on Skye’s tracksites, including a 2018 Scottish Journal of Geology paper describing the Brothers’ Point locality and its mix of dinosaur prints.
DinoTracker and the limits of AI fossils
To make the method easier to use, the team released DinoTracker on GitHub with open code and an app that can compare a footprint outline with the study’s database. It is the kind of quick check that could help prioritize which strange prints deserve a closer look, especially when time and tides are working against you at a tracksite.
The authors stress that the tool is not an all-knowing judge, and that caution is part of the point. The system relies on two-dimensional silhouettes, and sediment, motion, and preservation can all distort the same foot into different shapes.
Still, the approach fits a broader trend toward transparent AI in fossil science, echoed by a 2024 Earth-Science Reviews overview and a 2024 Integrative Organismal Biology review on how machine learning can help scientists study biological shape.
The original study was published in Proceedings of the National Academy of Sciences.
