A finger bone in a Siberian cave. A genome 50,000 years old. An AI that reconstructed an entire unknown human species from DNA alone — before a single skull was found. The story of how artificial intelligence is rewriting the human family tree.
In 2008, a fragment of a child's finger bone was found in Denisova Cave in the Altai Mountains of Siberia. It was small enough to fit in a test tube. The DNA extracted from it would, within two years, reveal a previously unknown species of human being — one that had interbred with our ancestors, gifted altitude-survival genes to Tibetan populations, and vanished without leaving a single skull anyone had identified.
We now know more about this species — the Denisovans — from ancient DNA and machine learning than from any physical fossil. AI reconstructed their facial anatomy before a face was ever found. This is what the marriage of genomics and artificial intelligence looks like: it doesn't just tell us who our ancestors were. It shows us what they looked like.
The story of the Denisovans begins, improbably, with a molar. In 2000, a local man brought a large, unusual tooth found in Denisova Cave to a Russian archaeologist. No one knew what species it belonged to. Eight years later, a fragment of finger bone from the same cave was sent to the laboratory of Svante Pääbo at the Max Planck Institute for Evolutionary Anthropology in Leipzig. Pääbo's team had pioneered ancient DNA sequencing — extracting and reading genetic material from bones tens of thousands of years old.
When they sequenced the finger bone's mitochondrial DNA in 2010, they expected to find either a Neanderthal or a modern human. What they found was neither. The genetic sequence was clearly archaic — as old as Neanderthal DNA — but distinct from any known hominin. A new species had been identified entirely from the genetic code of a child's finger bone. The paper was published in Nature.
Subsequent nuclear DNA analysis revealed the full picture: Denisovans were a sister group to Neanderthals, diverging from a common ancestor roughly 400,000 years ago. They ranged across a vast swathe of Asia. And critically — they interbred with modern humans not once, but multiple times, in different populations and different locations. The genetic traces they left behind are still measurable in living people today.
Svante Pääbo received the Nobel Prize in Physiology or Medicine in 2022 for these discoveries. The Nobel Committee's citation noted that Pääbo's work had "established an entirely new scientific discipline" — palaeogenomics — that "provided the foundations for exploring what makes us uniquely human." Much of that discipline now runs on machine learning.
In 2019, a team led by Liran Carmel and David Gokhman at the Hebrew University of Jerusalem published a paper in Cell that achieved something that sounded, on its face, like science fiction: they used artificial intelligence to reconstruct the physical appearance of a Denisovan — a species known almost entirely from DNA.
The method rested on a phenomenon called DNA methylation. Methylation is a chemical modification to DNA that controls which genes are switched on or off in different tissues. Crucially, methylation patterns in ancient DNA are partially preserved over tens of thousands of years. And methylation patterns in bone-forming genes are predictive of skeletal anatomy: regions of bone that are larger or smaller, ridges that project or recede, proportions of the face and cranium.
The team trained machine learning models on the relationship between methylation patterns and skeletal anatomy in living and recently-extinct species — comparing known anatomy with known methylation signatures. They then applied these models to the Denisovan genome, inferring the anatomical profile of a species seen only in fragments.
The reconstructed Denisovan was archaic in significant ways: a wider skull, a longer dental arch, projecting jaw, large molars, and a generally more robust build than both Neanderthals and modern humans. The wide skull was confirmed when the Xiahe mandible — a partial lower jaw found in Tibet — was identified as Denisovan through protein analysis in 2019. The jaw's proportions matched the AI's prediction.
The DNA methylation method for inferring anatomy from genetics has applications far beyond paleoanthropology. The same approach — using epigenetic markers to predict physical traits — is being applied to ancient pathogen genomes, to reconstruct what extinct diseases looked like before the antibiotic era, and to predict disease susceptibility from genetic profiles. AI trained on evolutionary biology is informing modern medicine.
The Denisovans are one discovery in a rapidly accelerating field. Ancient DNA — extracted from bones and teeth found at archaeological sites around the world — is being sequenced at an extraordinary rate: from perhaps 50 new genomes per year in 2015 to several thousand per year by 2024. Machine learning algorithms process these genomes, identifying population structures, migration waves, admixture events, and ancestry that no amount of traditional archaeology could reconstruct.
What has emerged is a human family tree of staggering complexity. The ancestors of modern Europeans, for example, were not one population but at least three: western hunter-gatherers who lived in Europe for tens of thousands of years, early Anatolian farmers who arrived around 8,000 years ago, and Yamnaya steppe pastoralists from the Pontic-Caspian steppes who swept west in a massive migration beginning roughly 5,000 years ago. Each wave brought different genetics, different immune profiles, and different cultural innovations — all now detectable in ancient bones and teeth through computational analysis.
The Yamnaya migration is particularly striking. AI analysis of ancient genomes suggests it was one of the largest population movements in prehistoric Europe — potentially replacing up to 75% of the genetic ancestry of Neolithic Europeans in the space of a few centuries. The plague bacterium Yersinia pestis — DNA ancestor of the Black Death — appears in the same populations at the same time, suggesting that pandemic disease may have preceded or accompanied the migration. These are findings that would have been invisible to archaeology alone; they emerge only from large-scale computational analysis of ancient genetic data.
Approximately 1–4% of the genome of non-African modern humans derives from Neanderthal interbreeding. AI analysis of where these Neanderthal sequences have been retained — and where they've been eliminated by natural selection — tells a functional story. Neanderthal-derived sequences are enriched in genes related to innate immunity, keratin production, and metabolism. They appear to have provided adaptive advantages in the cold, pathogen-rich environments of Eurasia. Your immune system is partly inherited from a species that went extinct 40,000 years ago.
Ancient DNA is not the only domain where AI is rewriting history. In archaeology, a technology called LiDAR — Light Detection and Ranging — has transformed our ability to see what lies beneath jungle canopies and dense vegetation. LiDAR fires millions of laser pulses per second from aircraft, measuring return times to create precise three-dimensional maps of the ground surface below even dense forest cover.
The challenge has always been the data. A single LiDAR survey can generate billions of data points. Processing it manually to identify archaeological features — the subtle ridges of ancient walls, the regular depressions of plazas, the straight lines of causeways — took years. Machine learning has compressed that to weeks, and in some cases days.
In 2024, a study published by a consortium of researchers revealed over 1,000 previously unknown Mayan structures in northern Guatemala, identified by AI processing of LiDAR data collected over a survey area of several hundred square kilometres. The structures included monumental architecture — pyramids, causeways, ball courts — that had been invisible beneath forest cover for centuries. The Maya civilisation, long thought to have been a collection of relatively isolated city-states, appears to have been far more densely interconnected than previous evidence suggested.
Similar AI-driven LiDAR discoveries have emerged from the Amazon basin (complex pre-Columbian ring villages and causeway networks in the Llanos de Mojos, Bolivia), the Cambodian highlands (a major unknown settlement adjacent to Angkor Wat), and sub-Saharan Africa (ancient iron-smelting complexes invisible from the ground). Across each survey, the pattern is the same: what human analysts spent years looking for and didn't find, machine learning identifies in days.
The implications extend beyond archaeology. These discoveries are reshaping our understanding of pre-industrial human carrying capacity — how many people could live sustainably in environments we assumed were marginal. The Amazon was not empty forest before European contact. It was densely populated, intensively managed, and home to civilisations whose engineering is only now becoming visible through the trees.
Buy me a coffee