David Reich – Why the Bronze Age was an inflection point in human evolution

We’re back with David Reich from Harvard to talk ancient DNA and what it reveals about us. To start us off, how do you describe your work?

I’m a geneticist using ancient and modern DNA to map how people were related in the past and how that connects to people today.

Our last conversation really resonated, and you’ve got a new study that digs into biology, not just migrations. Set the stage for what you set out to learn.

The field long hoped to track biological change through time, but small sample sizes made that hard; one genome is great for ancestry, but it’s just one data point for a trait. Now with large ancient datasets, we can watch allele frequencies move through time to test where selection really acted.

Why focus on frequency change?

When environments shift—new diets, pathogens, altitude—adaptive variants nudge upward, and with enough data you can see movements too big for chance. The signal is tiny without big numbers, so scale is everything.

The standard view held that recent human evolution was quiet; comparing continents shows little absolute divergence. Our partitioning shows about 98 percent of frequency shifts owe to drift and structure, yet the remaining sliver reveals widespread, real selection.

But if one population replaces another, isn’t that selection too?

Replacement can reflect culture or tech and it moves the whole genome together, which masks locus-specific adaptation. We look across many small place-time windows with minimal mixing and ask if the same allele edges upward everywhere; big steppe-era swings, for example, are mostly migration, not adaptation.

On results: you report around seven thousand two hundred candidate sites at fifty percent confidence, so roughly half are likely real—right?

Yes, about three thousand six hundred at that threshold, and many more weaker signals sit below it. The genome looks alive with subtle selection even if most change is demographic noise.

Signals cluster in immune genes and also in metabolism, while behavior-linked sites don’t enrich because their effects are spread across many loci.

So that’s not saying behavior wasn’t selected, just that your hits skew to immune and metabolic biology.

Exactly; behavioral traits are highly polygenic with tiny effects, so we’re underpowered there, but other tests show selection touched them too. The big picture is that selection intensified, especially in the last five millennia.

Farming, crowding, and close contact with animals created new pressures, and immune and metabolic selection ramps up in that era. It suggests a recent environment–genome mismatch being rapidly corrected.

Give us a few concrete examples.

A tuberculosis-risk variant surged to roughly ten percent, then fell sharply in the last three millennia, hinting at a pathogen shift; the B blood-group allele rose at the expense of A. Skin lightening accelerated about four to two thousand years ago, and traits like hemochromatosis show reversals with stronger swings in the north than the south.

You once saw little selection over a few centuries in African American genomes, yet the Bronze Age shows strong effects—how do we square that?

A few generations is too short to register modest selection, but compound it over millennia and you see it. Polygenic pigmentation peaks between two and four thousand years ago, and cognitive-performance predictors show strong Bronze Age selection and near-zero in the last two thousand years.

How big are these effects, and what’s migration’s role?

Migration drives huge jumps—European hunter-gatherers score far below modern means on today’s cognitive predictors, early farmers jump up, and steppe groups shift down—while our method teases out a smaller, consistent directional push from selection on top of that. So you see both a big demographic tide and a steady adaptive breeze.

That challenges the idea that advancing societies reduce selection for intelligence; your data says the opposite in early complex societies.

Data overturned my priors; the schooling and IQ predictors also track life-history traits like later childbearing, planning, and body composition. Iceland even shows recent selection against the schooling predictor, which underscores how context and fertility trade-offs can flip the sign.

What exactly is changing in the genome, and is this robust across populations?

The schooling predictor entwines with age at first birth, adiposity, walking pace, and wealth, suggesting a shared underlying dimension like executive function or delay of gratification. Using a Chinese study as an external yardstick, the same variants’ effects align with European ancient trajectories, which makes a Europe-only artifact unlikely.

Why wasn’t this maxed out among foragers if general problem-solving was always valuable?

What helps depends on ecology and values, and many traits trade off with fertility, so the optimum moves. Even liabilities like schizophrenia risk may track spectra where subclinical tendencies can be useful in some cultural settings and harmful in others.

You also see selection trimming body fat and related risks since agriculture—why?

Polygenic scores for body mass, fat distribution, and type 2 diabetes fell by about a standard deviation over ten thousand years, fitting a thrifty-genes story as food access stabilized. Hunter-gatherer feast–famine cycles operate on short timescales, while agrarian famines recur over years, so selection targets different dynamics.

Is there headroom to push traits like intelligence higher, which matters for how we think about AI?

Complex traits hold ample standing variation, so selection can move means a lot, though you pay trade-offs elsewhere.

It’s wild that such diversity comes from a small out-of-Africa source.

For polygenic traits the clay is abundant, and with today’s vast populations mutation supply isn’t limiting; the Bronze Age surge isn’t about hitting a size threshold, because the selection strengths we see would work even in small groups.

So was Bronze Age acceleration driven by bigger populations making selection more efficient?

Not really; effects of this magnitude beat drift even in thousands, so timing and environment explain more than census size.

Your 2016 scan found no universal hard sweeps tied to the so-called cognitive revolution—what does that imply?

We see no recent, species-wide single-gene sweeps in that window, so the cultural quickening likely rode many small genetic shifts or pure culture rather than a magic mutation.

If the genetics were in place earlier, why did farming wait until after the Ice Age?

The Holocene brought unusual climate stability, and across very different regions that seems to have unlocked agriculture multiple times from long-held capacities.

That’s surprising given how varied the birthplaces of farming are.

It is startling, but the consensus is many lineages with similar cognitive toolkits sat on a long fuse that lit when conditions steadied.

What big open questions are you chasing on archaic–modern relationships?

Genomes cluster Neanderthals with Denisovans, yet Neanderthals share tool traditions and uniparental lineages with modern humans due to deep interbreeding, which is a mismatch I’m trying to reconcile. I want to understand how those shared threads fit alongside the genome-wide tree.

There’s a small, roughly five percent, pulse of modern-human ancestry into Neanderthals, and I’m wondering if that sliver signals something bigger—that Neanderthals were, in important ways, culturally modern and part of the same Middle Stone Age turn as us.

When did that mixing happen, and is it true that parts of our genomes could be more recently connected to Neanderthals than to some other humans today? Wild if so.

Across the genome, many segments trace to common ancestors one to two million years back, so in places you can be closer to a Neanderthal on one parental line than to your other parent; that deep variation predates the Neanderthal–Denisovan split. This lens shapes how I think about global expansions that ancient DNA now lets us reconstruct.

After we wrapped, you jumped to the whiteboard with a new Neanderthal idea; we caught it on my phone, so video may help, but audio works fine.

The standard tree puts modern humans splitting from the Neanderthal–Denisovan line seven to eight hundred thousand years ago, plus a later five percent modern pulse into Neanderthals. I’m exploring a simpler story: a group invents Levallois-style tools three to four hundred thousand years ago, expands two to three hundred thousand years ago into Europe and Africa, mixes with locals, gets mostly replaced genetically, yet carries the toolkit and some genes forward; that could explain why Neanderthal maternal and Y lines align with us even though their whole genomes sit closer to Denisovans, and why a cave site in Spain shows nuclear DNA trending Neanderthal-like while the maternal line points elsewhere.

In range expansions, pioneers intermarry at the wave front and get swamped by local ancestry, so by the far edge they can be approximately ninety five percent local yet still transmit culture and some lineages. The same source group could have mixed within Africa too, leaving around eighty percent from an early modern lineage and roughly twenty percent from a deeply diverged African branch, with greater genetic distance there limiting gene flow compared to Eurasia; this could all be wrong, but it ties many loose ends.

Where was that archaic African branch, and what happened to the parts that did not fuse into modern humans?

Modern DNA in Africa shows deep structure older than a million years, followed by remixing a few hundred thousand years ago, with living groups carrying different proportions of those ancestries. We lack much ancient DNA from Africa, so we cannot map where each branch lived, but the signal of long-standing substructure is clear.

So in Neanderthal regions around three hundred thousand years ago, our lineage shares the technology without replacing the population—culture spreading faster than genes?

Genes move too, but they can dilute fast; think of Yamnaya ancestry in India, where a thin genetic trace still carried language and ideas. A five percent thread can mark a major cultural handoff.

Why would maternal and Y lineages be so overrepresented if the genome mix is small?

Lineal transmission can preserve one side if the expansion was matrilineal or patrilineal, and the other side could fix by natural or social selection. Male competition and mate choice can skew Y lineages, and differential fitness can winnow mitochondrial lines; it is the weakest link, but it could close the gap.

We likely met them again about seventy thousand years ago and then they vanish; why diffusion the first time but near replacement later?

Barriers grow with time apart, probably faster than linearly; the early Eurasian contact involved groups only about four hundred thousand years diverged, while the African contacts were closer to one point two million and likely much less fertile. The same cultural spark hits both regions, but gene flow tracks divergence, and the shared toolkit hints at a common source rather than independent invention.

Walk back your patched-model and epicycles analogy, and what makes your alternative hard to adopt?

We keep bolting fixes onto the sister-tree model to explain oddities like Neanderthal maternal and Y lines, much like adding epicycles to save geocentrism; a single wave spreading a toolkit and some lineages ties the threads with fewer assumptions. The leap is to link African deep structure with the Eurasian archaeological shift; bring genetics and archaeology together, and the timings line up more parsimoniously.

Could that tool shift reflect genetic change rather than culture alone?

Possibly; many gene lineages coalesce around four to five hundred thousand years ago, so mutations then could enable the new behaviors and be retained by selection during expansion. Archaeologists already see this earlier pulse as a major break, perhaps as formative as the later symbolic burst.

Do fixed differences separate people fifty thousand and three thousand years ago?

Clearer fixed differences start around three to four hundred thousand years ago, not within the last fifty thousand; that is when anatomically modern humans and recognizable Neanderthals appear, while the big behavioral flourish comes later.

I have learned to distrust my priors; I once tried to make the Neanderthal–non-African affinity disappear and failed, and I also thought recent selection was quiet. With far more ancient genomes and a new test that models relatedness and adds a simple directional selection term, we now see hundreds, maybe thousands, of positions shifting too consistently to be drift.

To recap, you soak up drift and demography with a relatedness matrix, then ask whether adding a constant selection pressure at a site improves prediction; what did that show?

Scaling to about sixteen thousand ancient individuals revealed at least 479 rock-solid selection signals, with thousands more likely; to validate, we showed that above a selection statistic near five, sites are roughly five times more likely to affect measured traits in the UK Biobank, and that plateau holds even when we control for background selection and allele frequency.

What let you scale ancient DNA output that much?

Sequencing costs collapsed, capture methods pull human fragments from mostly microbial mixes, and robotics turned dozens of genomes a year into thousands; we now target more than a million informative sites per sample. Pathogen reads are a bonus, but the human dataset has passed twenty thousand sequences, opening questions we could not touch a decade ago.

Awesome. Thanks for the time.