The AI Journal

In 2003 Russell Gray and Quentin Atkinson applied a Bayesian phylogenetic method borrowed from evolutionary biology to a dataset of cognate vocabulary across eighty-seven Indo-European languages, treating shared and divergent word-forms the way a biologist treats shared and divergent gene sequences, and used the resulting family tree to date the split at its root: the age of Proto-Indo-European itself. Their estimate came out at 7,800 to 9,800 years before present, decisively older than the five-to-six-thousand-year age that supporters of a Pontic-Caspian steppe homeland had generally assumed, and close enough to the timing of the spread of farming out of Anatolia that the paper was read, and largely intended, as vindicating the Anatolian hypothesis Colin Renfrew had set out in Archaeology and Language sixteen years earlier, over Marija Gimbutas’s and, later, David Anthony’s steppe account. Their tree also placed the Anatolian branch, Hittite among it, as the first to split away from all the others, a detail that later turned out to matter more than the headline date. For a debate that had run since the nineteenth century on philology and archaeology alone, a third method converging on one side looked like the beginning of an answer.

Twelve years later the ground shifted again, and in the opposite direction. In 2015 two ancient-DNA studies, led by Wolfgang Haak and Morten Allentoft respectively, sequenced prehistoric genomes from across Europe and Eurasia and found evidence of a demographic upheaval: Haak’s team, sequencing 69 ancient genomes, found that populations bearing Yamnaya-related steppe ancestry expanded into central Europe around 4,500 years before present, contributing roughly three-quarters of the ancestry found in the Corded Ware populations that followed, while Allentoft’s companion study, sequencing 101 genomes across the wider Bronze Age Eurasian steppe and forest zone, documented population turnovers on a comparable scale. Haak’s paper stated its linguistic claim cautiously, as evidence for a steppe source of “at least some” Indo-European languages, but the press coverage and much of the citing literature since have treated the two 2015 papers as having settled the homeland question in the steppe hypothesis’s favour, closing a debate that Gray and Atkinson’s linguistics had reopened only a decade before.

Set the two findings beside each other and the arithmetic does not close. Gray and Atkinson dated the existence of a single ancestral language to somewhere between 7,800 and 9,800 years before present. Haak and Allentoft dated a population expansion to approximately 4,500 years before present. Even taking the lower end of the linguistic estimate, more than three thousand years separate the two events — long enough for a language to have differentiated into several branches and for those branches to have moved considerable distances before any steppe migration began. A 2023 study by Paul Heggarty and a large international team, built on a newly collected cognate database covering 161 languages and applying a more rigorous ancestor-sampling variant of the Bayesian method, corrected for known biases in the earlier tip-dating approach and might have been expected to narrow the gap. It did not. It returned a median root age of approximately 8,120 years before present, close to Gray and Atkinson’s original figure and further still from the migration Haak and Allentoft had dated. Two independent linguistic analyses, sixteen years and an entirely rebuilt dataset apart, agree with each other and disagree with the ancient DNA by roughly the same margin each time.

Heggarty’s team did not treat the disagreement as a contradiction to be resolved by moving one date toward the other. Their model splits the family’s history into two separately timed events instead. Anatolian, and separately Tocharian, break away from the rest of the family very early, in a homeland the paper places south of the Caucasus rather than in Anatolia proper or on the steppe — a third candidate location, distinct from both nineteenth- and twentieth-century contenders — consistent with the older root age both linguistic studies converge on. The remaining languages, grouped as Core Indo-European, undergo their major dispersal much later, in a phase the paper aligns with the Yamnaya-associated expansion that Haak and Allentoft’s genomes document. The steppe migration survives in this account as a real and important event. What it stops being is the origin of the family. It becomes, instead, the vector for one branch’s later and geographically narrower spread, after a common ancestor several thousand years older had already begun to diversify, Anatolian foremost among the branches already gone its own way.

The strongest objection to describing any of this as a case of misattributed cause is that the technical steppe-hypothesis literature never actually made the claim now being corrected. Anthony’s 2007 study of the steppe hypothesis already treated Anatolian as the first branch to separate from the rest of the family, well before the classic Yamnaya-era expansions the book is otherwise built around, which means a gap between the family’s unity and the date of the steppe migration was built into the steppe hypothesis’s own chronology from the outset, not a correction imposed on it by Heggarty’s team sixteen years later. Gray and Atkinson’s original tree, for that matter, had already placed Anatolian first in line in 2003. Nobody with technical competence in the field, on this reading, ever thought ancient DNA dated Proto-Indo-European’s origin. They thought it dated a documented later expansion consistent with a steppe homeland for the bulk of the family’s branches, which is a narrower and entirely defensible claim, and one the 2023 paper’s own hybrid model largely confirms rather than overturns.

That defence holds for the argument’s careful version and fails for its public one. A model that distinguishes a family’s root date from one branch’s later migration date is compatible with the steppe hypothesis properly stated; a science-section headline announcing that DNA had solved the Indo-European homeland question, and the considerable body of citation outside historical linguistics that still treats 2015 as the year the debate ended, is not so careful. What ancient DNA dates is population movement. What historical linguistics dates, however imperfectly, is the ancestry of words. A population can migrate, be replaced, and replace others without that movement dictating what its descendants three millennia earlier had been speaking, and agreement between the two kinds of evidence, where it occurs, is a fact about the world the two methods happen to describe, not a guarantee built into either method’s logic. Whichever homeland turns out to be correct, the date attached to its founding population will have to come from the discipline whose method dates ancestral languages, not from the discipline whose method dates the people who, three thousand years later, moved. A migration this well documented does not need to be the family’s origin to be worth the attention it has received; it needs only to be correctly described, and three thousand years is too large a discrepancy to describe away as rounding.