Bite marks on fossil bones show leopards hunted early humans

Early humans weren’t always hunters – sometimes, they were the hunted. A new study using artificial intelligence has traced the marks on two-million-year-old Homo habilis fossils back to an unexpected predator: leopards.

The researchers applied a machine-learning method to analyze tooth impressions on bones from Tanzania’s Olduvai Gorge, one of the most famous sites in human evolution.

The results show that early members of our genus were still prey in parts of East Africa.

Reading bites in the bone

The work was led by Professor Manuel Domínguez-Rodrigo of Rice University. His work is focused on paleoanthropology and artificial intelligence in respect to zooarchaeology, vertebrate taphonomy, and methods that reduce subjectivity in fossil interpretation.

Tooth marks are not random scratches. In taphonomy, the study of how bones change after death, patterns on bone can reveal who ate whom and when.

Earlier computer vision efforts already pushed this field forward. A 2024 analysis showed that algorithms could sort tooth marks by species with high accuracy.

Species-level identification matters. It helps test whether early humans scavenged from big cat kills or carved their own meat first.

Adaptive AI meets archaeology

The new models lean on meta-learning, an approach that trains a system to adapt quickly from few examples, rather than memorizing one task.

A key variant is model agnostic meta-learning (MAML), a method that tunes a network so that just a few gradient steps yield strong results on new tasks.

The MAML method is flexible across problems, and is used as a general strategy for fast adaptation.

The team also used few-shot learning, a family of methods that work with very small labeled datasets by focusing on transferable features.

Behind the leopard predictions

Researchers photographed 1,296 tooth marks from four carnivores, including leopards, lions, hyenas, and crocodiles. They trained ensembles on these images to learn subtle mark shapes and textures.

Performance was measured with the macro average F1 score, a single number that balances precision and recall across classes.

The top MAML model reached 84 percent on this metric, which shows not just correctness, but even treatment of all carnivores.

The best accuracy – 85.13 percent – came from the Xception based setup. It also kept taxon-specific F1 scores at or above 80 percent, including 88 percent for leopards.

These figures matter because earlier deep learning systems struggled with rare classes like crocodiles. The new approach raised the floor while keeping the ceiling high.

Humans as leopard prey

The method was then applied to two Homo habilis fossils, OH7 and OH65, found at Olduvai Gorge. The marks on both matched leopard feeding behavior, with OH7 showing particularly strong probabilities.

The results suggest that early members of our genus were not yet dominant predators. In at least some parts of East Africa, leopards still had the upper hand over humans.

The finding challenges the traditional story of human evolution as a clean shift from prey to hunter.

If leopards could still bring down Homo habilis, then the rise to top-predator status came later – or varied by species and place. It’s a reminder that the path to dominance was neither quick nor uniform.

Hidden danger in plain sight

Even with our modern dominance, the findings serve as a reminder that predation risk once shaped the human mind.

Living as potential prey likely favored sharper vigilance, early-warning cooperation, and the ability to read threats before they struck. Those adaptations may underlie aspects of social awareness that persist today.

The leopard’s bite marks on early human bones are not just relics of violence. They record the ecological pressures that helped produce the very intelligence later used to escape such fates.

Expanding the AI fossil record

The models look strong, but the field context remains messy. Diagenesis – chemical and physical alteration of fossils over time – can blur surfaces, and overlapping agents can confuse even well-trained algorithms.

Open science will help here. The image bank and code are public in a Harvard Dataverse dataset, which lets other teams check, extend, and challenge the results.

A broader library, especially with more crocodile and lion marks from varied settings, should tighten estimates. Better controls on image color and lighting will also reduce bias.

The study is published in Royal Society Open Science.

—–

Like what you read? Subscribe to our newsletter for engaging articles, exclusive content, and the latest updates. 

Check us out on EarthSnap, a free app brought to you by Eric Ralls and Earth.com.

—–