A glass of milk in the morning is probably fine for you; for someone else it means a bad hour. Bread is fine for you; for someone else it inflames the gut, and that condition has a name — coeliac disease. Part of these differences is written in genes, and the share of people in a population carrying those variants was not the same ten thousand years ago as it is now.
Nobody used to be able to see that share change. You could only reason backwards from the genomes of living people: find a stretch with suspiciously little variation around it and guess that something had swept through. It is like walking into a wood, seeing a few unusually tall trees, and guessing that someone pruned them — you have no photographs, only the wood as it stands.
Now there are photographs.
From one graduation photo you cannot tell who grew three inches. Take one every year and there is nothing to guess — you just flip through them.
Natural selection, by definition, is a variant's share of a population changing directionally over time. So the most direct approach is to line up the genomes of people who died in different eras and read along the queue. As the ancient DNA revolution topic covered, pulling a genome out of bones and teeth tens of thousands of years old has gone from impossible to routine in twenty years — once enough data piles up, this becomes doable.
But there is a hard problem: a share that changed is not a share that selection pushed. A variant can become more common in a place for at least four reasons. Selection: its carriers happened to leave more descendants in that environment. Migration: a group that already carried it moved in. Admixture: two groups mixed, and the share got averaged. And drift — in a population that is not large, who has how many children carries a lot of luck, so the share wobbles on its own, with no force required. West Eurasia over these ten thousand years happens to be the age of mass resettlement: farmers out of West Asia, herders off the steppe, arriving and mixing in wave after wave. All four causes were running at once.
In April 2026, Ali Akbari, David Reich and colleagues published in Nature a study that does exactly this: model out the share that migration and admixture should have produced anyway, then look at what is left. The data are 15,836 ancient West Eurasian genomes (10,016 of them reported here for the first time) plus thousands of present-day ones, spanning more than ten millennia. The first-order conclusion is counter-intuitive — of all that frequency change, only about 2% is attributable to directional selection.
The second-order conclusion is the point. Inside that 2% sit 479 variants whose evidence is strong enough to name individually — where inference from modern genomes had firmly established perhaps a dozen or two (the famous one is lactase persistence, covered in the gene-culture coevolution topic). Another 7,600-odd loci are rated "more likely than not" to be real. A small fraction times a whole genome is still a large absolute number.
Open your fridge: milk, bread, cheese, a box of eggs. Fifteen thousand years ago not one of those existed. And it was not only diet that changed — people settled, packed together, lived up against livestock, stored grain in one place, and the rats and pathogens came with it.
When the environment changes, the answer to "who leaves more descendants" changes with it. Note that there is no such thing here as "evolving something in order to adapt to farming": nobody set a goal. Variation was already scattered through the population, most of it neither good nor bad; the new environment simply meant some carriers happened to have a few more children, and dozens of generations later their variant was common. The sieve changed. The material did not take orders.
The study's most striking single finding is this: selection got stronger after farming appeared, not weaker.
There is an immune-system gene called HLA-DQB1, and one variant on it is today the largest risk factor for coeliac disease (eat the gluten protein in wheat and the gut inflames). Along the time axis it climbed from near zero to about 20% in only four thousand years. Why it was pushed up is still debated — variants in immune regions usually have something to do with whichever pathogens were circulating, and being useful against a pathogen and being intolerant of gluten can perfectly well be two faces of the same variant. No variant comes with upside only.
Another is blood type B, from near zero to about 8% over six thousand years. The third is the most instructive: on a gene called TYK2 there is a variant that raises the risk of tuberculosis. It rose from about 2% around 5,500 years ago to about 9% by 3,000 years ago — and then fell back to roughly 3%. The direction reversed.
That turning curve does more work than any argument: having been favoured is no guarantee of staying favoured. Selection answers only to the environment of the moment, and when population density, livestock and the prevalence of a pathogen shift, the same variant goes from credit to debit. So "this variant was selected" always needs the clause "under what conditions".
"Civilisation switched natural selection off" does not survive this time axis: farming, settlement and dense populations are themselves new sieves, not an off switch for sieving. But do not slide to the other extreme either — these data cover the past ten thousand years. They say selection did not stop over that stretch, not that you can read off what is being selected today.
There is a gene called CCR5, and what it builds happens to be the handle the HIV virus grabs to get into a cell. Some people are born missing a small piece of it, and those missing both copies are essentially never infected. That deletion is not rare in Europeans, and along the time axis it was being pushed up.
But HIV is a twentieth-century disease. So whatever pushed it up back then cannot possibly have been HIV.
This is the step in the whole business where it is easiest to fall. What a variant "is associated with now" is an association measured in today's environment, in this group of people; the reason it was pushed then can be something else entirely. Two routes make that mismatch near-inevitable. One is pleiotropy — a single variant usually touches several things at once, the one selected need not be the one you measure. The other is hitchhiking — chromosomes travel down in blocks, so a genuinely favoured variant drags a string of neighbours up with it, and those neighbours, which did nothing, light up alongside it in today's statistics.
The paper is at its most careful exactly here. It also reports another class of result: those scores that add up thousands of small-effect variants shifted, over ten thousand years, by about one standard deviation of present-day variation — predicted body fat and schizophrenia risk going down, predicted cognitive-test performance going up. And immediately after comes the sentence that matters: these predictors were all built in modern industrialised societies, and how they relate to the traits that actually affected reproduction back then is unknown.
That is not politeness. Such scores already lose accuracy noticeably when moved to a population of different ancestry — a repeatedly confirmed fact — and the gap to be crossed here is not one population but ten thousand years, over which environment, nutrition, disease and what even counts as "cognitive-test performance" all changed. Add two hard boundaries: the study covers West Eurasian populations only, so nothing extends to humanity as a whole; and "selected" has never been a synonym for "better" — that tuberculosis curve rising and falling is the ready-made counterexample. Reading results like these as one population being innately superior is a substitution made at exactly this step, and it is where this stretch of scientific history has gone wrong most often.
The evolutionary psychology topic named a failure mode: see that a trait is useful now and reason backwards to its having been selected for that — plausible-sounding, and impossible to falsify. The contribution of ancient DNA time series is not that it finds evidence for stories like that, but that it turns them, for the first time, into questions data can refuse: did that variant actually rise back then? Go and look.