In Europe, I previously found a positive association between Early European Farmer / Anatolian farmer ancestry and educational-attainment PGS. That was not just an isolated observation: Akbari et al. later reported a similar farmer-related signal in their ancient-DNA analysis. So the European comparison is not merely that farmers expanded. It is that farmer ancestry appears to carry a detectable positive EA signal.

But Europe is not the only test case. China had its own great Neolithic agricultural cores, especially around the Yellow River and the eastern farming societies of Shandong. If farming populations really carried a distinctive trait profile, we should not expect the signal to stop at the Bosporus. We should see something similar in East Asia.

That is what I find here.

Using an imputed ancient East Asian dataset, imputed with GLIMPSE and then scored for several polygenic traits, I find that educational-attainment PGS rises through time in East Asia. The trend remains after controlling for ancestry, latitude, and sequencing coverage. Skin-colour scores also move in the expected direction: light-skin PGS rises and dark-skin PGS falls. Height, however, does not show the same robust increase.

The most interesting part is where the EA signal sits. The strongest positive independent ancestry effect is Yellow River / northern farmer ancestry. And the highest descriptive EA means appear in Shandong Neolithic farmer-related ancestry. In other words: the high-scoring signal is not random. It clusters around the old farming heartlands.

This suggest that the same broad pattern seen in Europe may also appear in East Asia: farming expansions may have carried more than crops.

Subscribe now

What Was Scored

I scored four traits:

• educational attainment, averaging standardized EA3 and EA4 scores

• height

• light skin colour, using the Kim et al. skin-colour GWAS

• dark skin colour, using the Pan-UKBB meta-analysis skin-colour GWAS

The genomes are ancient, imputed with GLIMPSE, and then scored using modern GWAS summary statistics. That immediately requires caution. Ancient PGS are noisy. GWAS portability is imperfect. Coverage, imputation quality, and ancestry structure can all matter.

So the key question is not whether a raw score rises or falls. The key question is whether the trend survives ancestry and technical controls.

The main model is:

Z_PGS ~ Date + ADMIXTURE + Latitude + Coverage

Date is in years before present. I report time effects as change per century forward in time.

I also fit interaction models:

Z_PGS ~ Date + ADMIXTURE + Date:ADMIXTURE + Latitude + Coverage

These ask whether the time trend differs by ancestry.

Main Regression Effects

The first figure shows the standardized regression coefficients for EA, height, light skin colour, and dark skin colour after controlling for time, ancestry, latitude, and coverage.

Figure 1. Standardized regression effects for EA, height, light skin colour, and dark skin colour. Each point is a standardized coefficient from the ancestry-controlled regression. Error bars show approximate 95% confidence intervals.

The East Asian EA Trend Looks Like the European One

The table below compares the time trend before and after controlling for ancestry. If the signal were just ancestry turnover, the EA trend should collapse after ADMIXTURE is added. It does not.

EA increases through time whether or not ancestry is controlled. Light-skin PGS also increases, while dark-skin PGS decreases. Height does not replicate the same pattern.

The EA increase is not large per century, but it is consistent and highly significant. It also remains almost unchanged after ancestry control: the ancestry-adjusted slope is only about 5% smaller than the unadjusted slope.

That looks like the European pattern. EA-associated alleles increase through time, and the trend is not simply erased by adding ancestry.

Skin colour also behaves as expected. Light-skin PGS rises. Dark-skin PGS falls. Height does not join the party: it is basically null without ancestry control and weakly negative after ancestry control. This is different from what I found in Europe, and likely explains the fact that East Asians are shorter than Europeans.

The EA Trend Is Not the Same in Every Ancestry

If a single homogeneous process were operating across all East Asian groups, the time trend would look similar in every ancestry component. It does not.

Adding Date × ancestry terms improves model fit for EA and dark skin colour, modestly for Kim light skin, and not convincingly for height. BIC penalizes the extra terms heavily, so I treat this as evidence of heterogeneity rather than final model selection.

The EA trend differs by ancestry. That matters because it tells us the signal is not just a single line through all ancient East Asians. Some ancestry backgrounds show stronger increases than others.

The next plot uses samples dated to the last 12,000 years to avoid a few very old samples stretching the x-axis. The Y variable is EA residualized on coverage and latitude. The X variable is raw date in years before present.

Figure 2. EA PGS time trend by dominant ADMIXTURE component, limited to samples dated ≤12k BP. The Y axis is EA residualized on coverage and latitude. The X axis is raw Date BP.

The faceted version makes the differences easier to see.

Each panel shows one dominant ancestry component. The labels report the main Date beta, the Date × component interaction beta, and the total beta. Positive values mean EA increases forward in time. Each panel shows one dominant ancestry component. The labels report the main Date beta, the Date × component interaction beta, and the total beta. Positive values mean EA increases forward in time.

Figure 3. Faceted EA PGS time trend by dominant ADMIXTURE component, ≤12k BP. Labels report standardized main, interaction, and total betas.

Which Ancestry Predicts Higher EA?

There are two ways to answer this, and they should not be confused.

The first is the better causal-style question: holding time, latitude, coverage, and the other ancestry components constant, which ancestry component independently predicts higher EA?

The four positive ancestry effects are all East Asian farmer-related, but they are not identical. Yellow River / northern farmer captures the Central Plain / Yellow River Neolithic axis, including Yangshao-Longshan-related ancestry. Central-East China farmer captures a central-eastern China axis with both Yangtze and lower Yellow River affinity. Fujia/Dawenkou/Shandong Neolithic is the local Shandong Neolithic farmer component, associated with Dawenkou/Fujia-related eastern coastal farming populations. Southern coastal East Asian captures a more southern/coastal ancestry profile, visible in coastal and later historical southern East Asian samples.

The strongest positive independent EA-associated ancestry effect is Yellow River / northern farmer ancestry. This is the component most closely aligned with the Central Plain / Yellow River agricultural substrate relevant to later northern Han formation.

The second question is descriptive: among individuals whose largest component is X, what is their average EA score? This is less clean because dominant groups differ in time, place, and mixture. But it is still useful.

Dominant-component means are descriptive. Shandong Neolithic farmer-related ancestry has the highest mean EA, while Yellow River / northern farmer ancestry has the strongest positive partial effect in the multivariable model.

So the two results are complementary. Yellow River / northern farmer ancestry is the most robust independent predictor. Shandong Neolithic farmer ancestry has the highest descriptive mean.

Both point back to farmers.

Longshan, Shandong, and the Farming Signal

This is the finding I find most interesting.

The high-EA signal is not attached to random ancestry. It appears in two farmer-related components:

• Yellow River / northern farmer ancestry, which peaks in Yangshao, Longshan, and related Yellow River Neolithic/Bronze Age groups

• Fujia/Dawenkou/Shandong Neolithic ancestry, which represents eastern Neolithic farmer ancestry in Shandong

Dawenkou and Fujia-related groups are generally earlier than Longshan. Dawenkou is often dated roughly 4100-2600 BCE, while Longshan is roughly 3000/2600-1900 BCE. There is overlap at the transition, but the broad sequence is clear: eastern Neolithic farming groups precede and feed into later Longshan-era complexity.

This is where the comparison with Europe becomes hard to ignore. In Europe, Anatolian farmers were not just a demographic wave. They were the population layer that made Neolithic Europe possible. In East Asia, Yellow River and Shandong Neolithic farmers may have played a similar role.

Farming societies created a very different selection environment: sedentism, storage, hierarchy, craft specialization, property, accounting, coordination, and eventually states. Over many generations, these environments could sort or select for cognitive and personality traits useful in dense agrarian societies.

That is the farming-effect hypothesis.

Agriculture creates the niche. The niche changes the people.

The next figure shows standardized EA PGS by metadata population, using groups with at least 20 individuals.

Figure 4. EA PGS boxplot by population, n ≥ 20. Groups are ordered by median EA PGS from highest to lowest.

The ancestry profiles show why raw group means can be misleading. Many archaeological or historical labels are not pure ancestry blocks.

Figure 5. Mean K12 ADMIXTURE profile by population, n ≥ 20.

Did This Help Make the Han?

The genetic formation of the Han is usually told as a story of demography, agriculture, and state formation. Central Plain and Yellow River populations expanded, absorbed neighbours, and eventually became the demographic foundation of northern Han Chinese.

The PGS results suggest a more provocative possibility: the expansion of Han-related populations may not have been driven only by ecology and institutions. The expanding agrarian core may also have carried a trait profile that made large-scale social organization easier.

That would help explain why the Central Plain mattered so much. Millet agriculture gave it density. Geography gave it centrality. Longshan and later Bronze Age societies gave it hierarchy. The state gave it reach. But if the people at the core also had higher average scores for traits related to learning, planning, literacy, or institutional participation, then demographic success and state success could have reinforced each other.

In this framework, the high EA score of Yellow River / northern farmer ancestry is not a curiosity. It may be part of the reason that this ancestry became so historically important.

Share

References

Akbari, A., Perry, A., Barton, A.R. et al. Ancient DNA reveals pervasive directional selection across West Eurasia. Nature (2026). https://doi.org/10.1038/s41586-026-10358-1

Kim, B., Kim, D.S., Shin, J.G. et al. Mapping and annotating genomic loci to prioritize genes and implicate distinct polygenic adaptations for skin color. Nature Communications 15, 4874 (2024). https://doi.org/10.1038/s41467-024-49031-4

Rubinacci, S., Ribeiro, D.M., Hofmeister, R.J. et al. Efficient phasing and imputation of low-coverage sequencing data using large reference panels. Nature Genetics 53, 120-126 (2021). https://doi.org/10.1038/s41588-020-00756-0

Lee, J.J., Wedow, R., Okbay, A. et al. Gene discovery and polygenic prediction from a genome-wide association study of educational attainment in 1.1 million individuals. Nature Genetics 50, 1112-1121 (2018). https://doi.org/10.1038/s41588-018-0147-3

Yengo, L., Vedantam, S., Marouli, E. et al. A saturated map of common genetic variants associated with human height. Nature 610, 704-712 (2022). https://doi.org/10.1038/s41586-022-05275-y

But in ancient DNA, ancestry is not a nuisance variable like a genotyping batch. It is part of the historical process. The populations of West Eurasia did not sit still while time passed over them. Hunter-gatherer ancestry, early farmer ancestry, Steppe ancestry, and later historical mixtures all entered the record at different times and in different places.

That creates a problem for any model that tries to reduce ancient West Eurasia to one corrected time trend.

Akbari et al. ask whether polygenic scores changed through time after controlling for genome-wide relatedness. That is a useful question, but it is not the same as asking whether populations evolved through time. If sampling date is correlated with genetic structure, then a GRM correction can absorb both confounding and genuine historical signal. Worse, if the corrected coefficient is interpreted as individual-level selection, the estimate depends on an exogeneity assumption that may be violated.

TL;DR: Akbari et al.’s method is not useless. It answers a narrow residual question: is there a time trend after genome-wide structure is absorbed? But the broader evolutionary question is different. Much of the signal is between ancestry-like groups, and even within groups the slopes may not be common across ancestry backgrounds. Correcting for ancestry can therefore correct away part of the evolutionary history one is trying to study.

What Akbari et al. did

The core model can be written roughly as:

PGS_i = alpha + gamma * time_i + genetic_relatedness_random_effect_i + error_i

The genetic relatedness term is based on a genome-wide relatedness matrix, or GRM. This matrix summarizes how genetically similar every individual is to every other individual; mathematically, it captures the same broad structure that appears in genetic principal components, but the model uses the GRM as a covariance matrix rather than adding a few PCs as ordinary covariates. Its purpose is to account for the fact that ancient individuals are not independent draws from one homogeneous population. Related individuals and genetically similar groups should not be allowed to manufacture a fake time trend.

That is a reasonable concern. But the resulting `gamma` is not the total temporal change in the trait. It is the residual time trend after genome-wide structure has been absorbed.

Why This Is Not the Same as Total Evolution

Suppose a trait-related PGS rises through time because one ancestry background partly replaces another. Is that confounding, or is that evolution?

The answer depends on the question. If the question is narrowly, “Did the score change within a fixed genetic background?”, then ancestry turnover is something to remove. But if the question is, “Did populations genetically change through time?”, then ancestry turnover is one mechanism by which the change occurred.

That is the first limitation of the Akbari-style correction. It can remove a historically meaningful group component by treating it as structure.

The Stronger Statistical Issue

There is a second issue. Random-effects-style identification depends on the random effect not being correlated with the explanatory variable. Here the explanatory variable is time. In ancient DNA, time is obviously correlated with genetic structure: older samples and younger samples are not drawn from the same ancestry distribution.

The point is not merely philosophical. If the genetic random effect is correlated with sampling date, then the corrected time coefficient may not cleanly estimate the individual-level selection parameter. If there is correlation of this type, then the parameter estimate obtained is biased and inconsistent.

Prof. Gregory Connor pointed out this problem in the Akbari et al. methodology in an earlier tweet. [

The data show the dependence directly. Genetic PCs predict sampling time; the first ten PCs explain about 38 percent of the time variable.

Figure 1 shows the same diagnostic visually. This is the background fact behind the whole problem: date is genetically structured.

If genetic structure predicts date, then a structure-corrected time coefficient is not automatically a clean individual-level selection estimate.

Figure 1. Genetic PCs predict sampling time.

ADMIXTURE components tell the same story. Steppe ancestry rises toward the present, while WHG and EEF are older-enriched. This is not a minor nuisance; it is the central structure of ancient West Eurasian history.

Figure 2. Correlations between sampling time, genetic PCs, and ADMIXTURE components. Time is not independent of genetic structure.

Fixed Effects as a Check

A possible solution is to estimate fixed effects. Fixed effects ask a narrower question:

Within the same group, do later samples have different PGS than earlier samples?

They do not use between-group differences to estimate the time slope. That is the tradeoff. Fixed effects cannot estimate the part of evolution that happened through group turnover, but they can estimate within-group temporal change.

Using ancestry-like PC100 clusters, the fixed-effects estimates remain positive for EA and Height. They are smaller than pooled OLS, but larger than or comparable to the GRM-LMM estimates.

Table 2. Pooled OLS, GRM-LMM, fixed effects, and Mundlak between-group estimates.

For EA, pooled OLS gives a gamma of 0.698. The GRM-LMM estimate is 0.074. The PC100 fixed-effects estimate is 0.216. That means the within-cluster signal is still present, but the GRM correction estimates a much narrower residual trend.

For Height, pooled OLS gives 0.554, GRM-LMM gives 0.065, and fixed effects give 0.114. The same pattern appears, though less dramatically.

How Large Are the Group Effects?

The Mundlak decomposition makes the group component explicit:

PGS_ig = alpha + beta_within * (time_ig - mean_time_g) + beta_between * mean_time_g + error_ig

The within coefficient asks whether samples change through time inside ancestry-like clusters. The between coefficient asks whether clusters with different average dates also differ in PGS.

For EA, the within-cluster gamma is 0.216. The between-cluster gamma is 1.401. The between-group temporal gradient is therefore about 6.5 times larger than the within-group gradient.

For Height, the within-cluster gamma is 0.114 and the between-cluster gamma is 0.439, about 3.8 times larger.

Table 3. Size of the between-group temporal component using PC100 ancestry-like clusters.

Figure 3 summarizes the same point graphically. The answer changes depending on whether the model estimates the raw trend, the residual GRM-corrected trend, or the within/between decomposition.

Figure 3. GRM-LMM and Mundlak-style decomposition of temporal PGS change.

What This Means

Akbari et al.’s method is not useless. It is useful for one narrow question: does a residual average time trend remain after genome-wide structure is absorbed?

But it should not be sold as the full story of evolutionary change. The discarded group component is not automatically confounding. It may be the evolutionary signal.

The stronger problem is that a single corrected time slope may also miss ancestry-specific within-lineage change. If different ancestry backgrounds have different temporal slopes, then a model with one `gamma` compresses multiple trajectories into a single average. That can make real lineage-specific change look smaller than it is.

So the limitation is twofold:

1. The GRM correction can wash away between-group historical change.

2. The single-slope design can miss within-group or lineage-specific change when slopes differ by ancestry. Estimating a single-slope design is not even statistically consistent if the reality is that there are multiple-slope trends.

In ancient DNA, ancestry is not just something to control for. It is one of the ways evolution entered the data.

References

Akbari, A., Perry, A., Barton, A. R. et al. Ancient DNA reveals pervasive directional selection across West Eurasia. Nature (2026). https://doi.org/10.1038/s41586-026-10358-1

Allen Ancient DNA Resource, AADR v66.2, Reich Lab.

Share PifferPilfer

Ancient DNA has made one old temptation harder to resist. Once we can line up skeletons by date, plot their polygenic scores, and watch some traits move across thousands of years, it is natural to talk about a single trend through time.

That is also how the recent Akbari et al. selection scan is framed statistically. Their core model estimates one time coefficient for each trait after correcting for genome-wide relatedness. That is a clear and useful estimand, but it raises a question: what if ancient West Eurasia does not have one temporal trend?

Europe was not a single biological conveyor belt. The people sampled at 10,000 BP, 5,000 BP, and 2,000 BP were not just earlier and later versions of the same population. They carried different proportions of ancestry related to western hunter-gatherers, early European farmers, Steppe pastoralists, and later historical mixtures. If those ancestry backgrounds also had different trajectories, then a single time coefficient will compress too much history into one number.

That is what I wanted to test here. The question is not only whether polygenic scores changed through time. The sharper question is whether the single-trend assumption is adequate: did PGS trajectories move in the same way across ancestry backgrounds, or did different ancestries carry different slopes through time?

Very Short Methods Note

I used ancient West Eurasian individuals from AADR v66, restricted to samples from the last 18,000 years with K7 ADMIXTURE components available. The analysis sample contains 11,102 ancient individuals. Time is coded so that higher values mean more recent samples. The three ancestry components highlighted here are WHG (Western Hunter Gatherers), EEF (Anatolian/Early European Farmers), and Steppe_Yamn_CW (Steppe/Yamnaya/Corded Ware). I modelled each component as a function of time, latitude, longitude, and their interactions. I then tested whether the time slope for polygenic scores differs by ancestry background using models of the form:

PGS ~ time + ancestry component + time x ancestry component + latitude + longitude + coverage

The polygenic scores are standardized. EA is educational attainment, Height is height, and DD is delay discounting. DD is exploratory because the underlying score is much sparser than EA or Height.

The Ancestry Background Changes Sharply Through Time

The ancestry composition of the sample is strongly time-structured. WHG is high in the oldest bin, EEF rises after the spread of farming, and Steppe ancestry rises sharply after 5,000 BP.

Table 1 gives the basic time-window summary. The values are mean ADMIXTURE component proportions, so they should be read as broad ancestry signals rather than literal population labels.

The same pattern is easier to see visually. Figure 1 shows the turnover directly: the oldest samples are WHG-heavy, the middle Holocene samples are much more farmer-related, and the later samples carry much more Steppe-related ancestry.

Figure 1. Mean WHG, EEF, and Steppe ancestry components by time window.

The Steppe Increase Is Not Just a Date Effect

The Steppe component increases toward the present even after controlling for latitude and longitude. The adjusted slope is positive and large. EEF and WHG move in the opposite direction.

The spatial interactions are important. Steppe ancestry increases more strongly at higher latitude and toward the west. EEF and WHG decline toward the present, but their decline is structured more by longitude than by latitude.

Table 2 reports the core regression results. The first numeric column is the time slope after latitude and longitude are included. The interaction columns ask whether that time slope changes at different latitudes or longitudes.

Figure 2 then turns those interactions into predicted ancestry trajectories. It is useful because it shows that the Steppe rise is not evenly distributed across space.

Figure 2. Predicted ancestry components through time at different latitudes.

The ancestry turnover alone already makes the single-trend assumption look fragile. But the more important question is whether this matters for the polygenic scores themselves.

Below I test that directly: first by asking whether ancestry components change the PGS slopes through time, and then by decomposing the signal into within-group and between-group components. This is where the Akbari-style correction becomes most interesting.

Read more

When early farmers moved into Europe from Anatolia and the Aegean world, they did not enter an empty continent. They met western hunter-gatherers whose ancestors had lived there for thousands of years. In most of Europe, this was not just the spread of a new subsistence package. Olalde et al. (2026) summarize the broader Neolithic transition as 70-100% ancestry turnover across much of Europe between 6500 and 4000 BCE, with northwest Europe as a striking exception where forager ancestry persisted much longer.

The later Steppe-related expansions were just as consequential in some places. Olalde et al. (2018) estimated that Beaker-associated ancestry replaced about 90% of Britain’s gene pool after about 2450 BCE, and Olalde et al. (2026) summarize the process as roughly 90-100% replacement of local Neolithic ancestry within a few hundred years. In other words, some of these prehistoric migrations were close to total population replacement at the genetic level, even when archaeology shows complicated mixtures of people, practices, and material culture.

But how genetically different were these incoming populations from the people they partly replaced?

One way to answer that question is to use Fst, a standard measure of genetic distance between populations. An Fst of zero means two groups are genetically indistinguishable at the measured markers. Larger values mean greater allele-frequency differentiation.

I compare the ancient results with distances among modern superpopulations. In genetic datasets, a superpopulation is a broad continental-scale grouping, such as Europe, Africa, East Asia, South Asia, West Asia, Oceania, or the Americas. It is not identical to the everyday word race, but it is one of the technical categories closest to it: a coarse grouping of populations that share relatively more ancestry with each other than with populations outside the group.

Methods note

I used the AADR v66 2M dataset. The ancient groups are broad source-like sets defined from the AADR metadata and K=7 ADMIXTURE components, then Hudson Fst was computed directly from the genotype data. The ADMIXTURE components were used to select samples, not as the final object of comparison.

For the modern benchmark, I used the pairwise Hudson Fst estimates in fst/modern_groups_v66, restricted to modern population pairs with at least five individuals in both groups. That gives 5,253 modern population-pair comparisons. Where useful, I also use the existing broad superpopulation labels: Africa, Europe, West Asia, South Asia, East Asia, Oceania, and the Americas.

These labels are intentionally broad. They are not meant to imply that the Anatolian farmers or the Steppe people were one uniform tribe. They are source-like clusters built to estimate the size of two historical genetic gaps.

Two ancient replacement fronts

Before looking at the numbers, it helps to put the movements back on the map. The first arrow represents the spread of farmer-related ancestry from Anatolia and the Aegean into a Europe still partly occupied by western hunter-gatherers. The second represents the later movement of Steppe-related ancestry into farming Europe.

The first contrast is the larger one. The Fst between early farmers and western hunter-gatherers is 0.077. The later Steppe-vs-farmer contrast is 0.035.

That difference already says something historically plausible. Early farmers and western hunter-gatherers were separated by a deeper Near Eastern versus European hunter-gatherer ancestry divide. Steppe groups, by contrast, were not a completely alien source relative to Europe. They were themselves a mixture involving eastern hunter-gatherer-related and Near Eastern/Caucasus-related ancestry, and they entered a Europe whose farmer populations already had some hunter-gatherer ancestry.

The map shows where the two replacement fronts were. But the more interesting question is not just where these populations moved. It is how genetically far apart they were, and whether those distances were small by modern standards, large by modern standards, or somewhere in between.

Below the paywall I compare the ancient farmer-WHG and Steppe-farmer distances to thousands of modern population-pair Fst estimates. This is where the result becomes interpretable: one ancient migration looks closer to ordinary within-region variation, while the other sits much higher, approaching the lower range of modern continental-scale differences.

Read more

A standard GWAS compares unrelated people. That gives you a lot of power, but it also means the estimated allelic effects can absorb more than just direct biology: ancestry structure, geography, family environment, assortative mating, and indirect parental effects can all creep in if they are not perfectly controlled. Those problems become especially uncomfortable when the goal is to compare scores across populations, because even a subtle residual bias can masquerade as a striking genetic pattern.

That is one reason critics of cross-population PGS comparisons, including Gusev, have argued that family-based results showing weak or non-significant population differences should make us skeptical of using these scores to predict phenotypic differences across populations. Put more bluntly, his argument is that once you move from standard population GWAS to family or direct-effect GWAS, the dramatic between-ancestry differences largely evaporate or even reorder, which suggests that much of the original signal was stratification rather than real genetic causation. He states the position very starkly: “cross-population PGS comparisons have no theoretical basis whatsoever,” and he also argues that family-GWAS analyses of IQ produce very different group rankings from population-GWAS-based scores. In a Substack essay he similarly warns that the score mean in an external population can become an “arbitrary sum of predictive signal, noise, and bias.” He made the family-GWAS objection more explicit in a comment to one my earlier posts.

But what is a family GWAS? It asks a cleaner question: within families, when siblings differ in the alleles they happened to inherit, do they also differ in the trait? That within-family comparison strips away much of the shared background and gets us closer to direct genetic effects, but the price is lower power.

I started from the recent family GWAS paper by Guan and colleagues, which is interesting not just because it is family-based, but because it proposes a more general design for family GWAS that aims to be both more powerful and more robust than the usual within-sib setup. In other words, the Guan et al. framework is trying to keep the causal cleanliness of family designs while recovering some of the information that simpler family models leave on the table.

Once I took the direct-effect summary statistics and applied a conventional clump-and-threshold screen at p < 5e-8, almost nothing was left. Height retained 46 clumped SNPs, cognitive performance retained just 1, and educational attainment, hourly income, and household income retained 0. So the question became: if the signal is real but too diffuse to survive a brutally strict threshold, is there another way to pull it out?

The signal I am trying to recover here is not ordinary within-population predictive accuracy. It is a cross-population polygenic selection or adaptation signal: the possibility that populations differ, at least in part, because trait-increasing and trait-decreasing alleles have been pushed systematically up or down in frequency over time. That is why the filter is built from cross-population frequency structure. Cross-population LD is not a new term here; I am using it in the established sense of coordinated allele-frequency differences across populations, so that alleles at different loci do not move independently. If there is a real directional adaptation signal in the family-GWAS effects, then alleles with the same sign of phenotypic effect should not move independently across populations; they should show some coordinated frequency movement in the same broad direction.

The schematic below illustrates the intuition. If there is no coordinated cross-population signal, different SNP subsets will point in different directions. If there is a real adaptation-like signal, trait-increasing alleles should cluster together across populations, and trait-decreasing alleles should cluster together in the opposite direction.

That kind of coordinated movement ought to show up as higher split-half reliability across SNP subsets, which is closely related in spirit to the cross-population logic behind Berg and Coop’s Qx framework. Separately, I also use a method-of-correlated-vectors style check: if the scores that look more internally coherent are also the ones that line up better with the external phenotype, that is at least suggestive evidence that reliability is tracking real signal rather than noise. The hope is that reliability can help separate coherent polygenic signal from low-power noise.

Main Results

The validation step here follows a familiar logic from the polygenic-selection literature: if a polygenic score is really capturing a cross-population genetic signal, one way to probe that is to ask whether it correlates with average population phenotypes, as in earlier work by Piffer and by Berg and Coop.

For each trait, I computed average PGS across a broad 50-population panel assembled from 1000 Genomes and gnomAD together with additional public reference datasets, including Denmark MAC5, EgyptWGS, TaiwanBB, SWEGEN, 1KG Poland, and HGDP Papuan. The phenotype correlations then use the overlapping populations for which the relevant external phenotype is available, so the IQ-linked analyses and factor analyses are based on the matched phenotype subset rather than all 50 populations, and the height regressions use the smaller subset with Height plus HDI or protein data.

When I used split-half reliability to prune the scores, the cognitive, educational, and socioeconomic PGS all became much more predictive across populations.

Educational attainment rose from r = 0.264 to r = 0.806 against IQ, household income from 0.378 to 0.701, hourly income from essentially zero to 0.579, and cognitive performance from 0.413 to 0.623. Across the ten trait-method combinations, the meta-correlation between phenotype association strength and split-half reliability was 0.680. And when I factor-analysed the four IQ-linked scores, the raw common factor was basically uninformative, whereas the optimized factor correlated r = 0.700 with IQ.

Height was the exception, and that turns out to be informative too. On its own, the height score did not line up well with observed height across populations. But once I treated height as the environmentally sensitive trait it obviously is, the pattern changed: the optimized height PGS became predictive after adjusting separately for HDI or protein availability, whereas the raw height PGS remained unconvincing. So the broad picture is not that reliability magically fixes every trait in the same way. It is that reliability seems to sharpen the latent signal, while the final interpretation still depends on the biology and the confounding structure of the phenotype.

The optimization itself was simple in spirit: start with the liberal score, then keep throwing away the SNPs that make the score less internally coherent across populations. Here split-half reliability just means the following. Take the SNPs in a score, split them into two random halves, compute the population score from the first half and from the second half, and then ask whether populations that rank high on one half also rank high on the other. If the two halves keep telling a similar story, the score is internally coherent; if they disagree, the score is mostly noise. I repeated that random split many times and averaged the result, so the reliability estimate is not driven by one lucky partition.

Below the paywall, I show how the method works in practice: the correlation heatmaps, the height sensitivity analyses, the raw-vs-optimized factor structure, and the exact greedy SNP-selection procedure used to recover signal from a low-powered family GWAS.

Read more

That opens the door to a harder question than who came from where. We can now ask how much ancient populations differed from us, in what direction they differed, and whether those differences stayed within the range of ordinary regional variation or climbed into something larger.

That question matters even more now that ancient genomics has moved beyond ancestry in the narrow sense. The first explicit test of directional selection on cognitive ability using ancient DNA was published in 2017 by Michael Woodley and myself (Woodley et al., 2017). My later work (Piffer and Kirkegaard, 2024; Piffer, 2025) extended that approach and reported directional selection across multiple polygenic traits over the last twelve thousand years. Akbari et al. (2026) then replicated and greatly expanded that general result at much larger scale in West Eurasia.

Once you accept that ancient populations were being reshaped by migration, drift, replacement, and selection, another question becomes hard to avoid. Were ancient Europeans merely earlier versions of modern Europeans, or were some of them separated from us at a scale closer to what ordinary people would recognize as a major human division?

Geneticists usually avoid the word race and use broader labels such as superpopulation instead. In practice, superpopulations are the large continental-scale clusters used in modern population-genetic work: Europe, West Asia, East Asia, South Asia, Africa, the Americas, and Oceania. That makes them the closest technical counterpart to what people usually mean when they ask whether one population is as different as another race.

The obvious problem is confounding. Small ancient groups can look artificially extreme, and mixed-region comparisons can produce misleading nearest-population matches. So the analysis here uses a stricter benchmark: ancient groups are restricted to Europe, the modern side keeps all present-day superpopulations for comparison, and the main analysis keeps only ancient groups with at least twenty-five individuals.

Set the word itself aside for a moment. If you compare ancient Europeans to modern Europeans, do they stay within the range of ordinary modern European variation, or do they climb into the range now seen between present-day superpopulations?

To answer that, I use Hudson’s Fst, a standard population-genetic measure of genetic distance. Very loosely, higher Fst means two populations are more genetically differentiated from one another.

It is worth pausing to note that this kind of analysis depends on an enormous collective effort. Ancient DNA became historically informative not just because of new methods, but because laboratories, archaeologists, and database builders created datasets large enough to make temporal comparisons credible. The Harvard ancient-DNA lab and its many collaborators around the world have been central to that process, as have data infrastructures such as the Allen Ancient DNA Resource and the National Genomics Data Center in China, which make ancient genomes available in forms researchers can actually use.

Methods note. The benchmark uses Hudson’s Fst from the AADR v66 2M panel (Mallick et al., 2024). The main working set is the Europe-only ancient benchmark with ancient groups restricted to N >= 25. That leaves 105 ancient European groups for the same-region comparison: 47 in 0-3k BP, 31 in 3-5k BP, and 27 in 5-10k BP.

Results

Start with the raw Europe-only benchmark, because that is what makes the whole question interesting in the first place. Before any size filter is applied, the median ancient Europe versus modern Europe Fst is 0.129. That is not just above ordinary within-Europe variation. It is slightly above the median distance between present-day superpopulations. If you stopped there, the result would look stark: ancient Europeans would seem to be separated from modern Europeans by distances roughly on the scale of present-day superpopulation differences.

But that raw version is not the one to trust. Small ancient groups inflate the upper tail, and once they are removed the picture changes. In the main N >= 25 analysis, the median ancient Europe versus modern Europe Fst falls to 0.051, while the median modern within-superpopulation Fst is 0.027 and the median modern between-superpopulation Fst is 0.118 . So the cleaned result still places the typical ancient-modern comparison well above routine within-superpopulation variation, but no longer at the level of the typical gap between present-day superpopulations.

Figure 1. Cleaned Benchmark Distributions

Table 1. Headline benchmark numbers for the cleaned Europe-only N >= 25 analysis.

Space is the obvious source of genetic distance. Populations that live far apart usually differ more than populations that live close together. The more interesting question is what time does on top of that. Once the ancient side is restricted to Europe and the small groups are cleaned out, the answer is no longer muddled: older Europeans are systematically farther from the present than more recent Europeans are. The median same-region Fst is 0.030 for 0-3k BP, rises to 0.060 for 3-5k BP, and to 0.062 for 5-10k BP. In other words, the genetic distance between ancient and modern Europeans does not just reflect geography. It grows as you move backward in time.

Figure 2. Ancient-Modern Fst Through Time

What matters now is not just the median, but the overlap: how often ancient-modern distances are actually larger than the genetic distances separating present-day human populations, and which ancient groups still sit in that upper tail.

Read more

Introduction

Ancient DNA contains a surprise. When researchers extract polygenic scores linked to educational attainment from prehistoric skeletons, the scores rise through time (Piffer & Kirkegaard, 2024; Piffer, 2025). Hunter-gatherers score lower than early farmers, early farmers lower than Bronze Age populations, and so on. The standard interpretation is that this reflects something about the passage of time itself such as selection, migration, demographic turnover, or some mixture of all three.

But time is not the only thing changing across that span. Social and technological complexity are changing too. A Paleolithic forager band and an Iron Age kingdom are not just separated by millennia. They are separated by farming, writing, cities, specialization, storage, hierarchy, and the entire accumulated weight of what we loosely call civilization. So here is the question worth asking: when those polygenic scores rise through ancient history, are we really just watching a clock tick or are we watching something track the emergence of more complex ways of organizing human life?

In an earlier post I used ancient individuals from the AADR dataset to assign each archaeological period a civilization-stage score and ask whether that score predicts educational-attainment polygenic scores even after controlling for absolute date. In practice, that meant coding Paleo-Mesolithic groups as 1, Neolithic groups as 2, Copper Age groups as 3, Bronze Age groups as 4, and Iron Age groups as 5.

This follow-up revisits that question using corrected archaeological labels, the much larger AADR v66 2M dataset, and a stronger emphasis on the subset of samples whose period labels can be assigned directly from metadata rather than reconstructed from looser fallback rules. The point is not to abandon the original question. It is to ask whether the same broad pattern survives when the labels are cleaner and the sample is larger.

The leverage comes from a basic fact of world prehistory: civilization and chronology do not move in lockstep. Greece had farming millennia before Britain. Some populations at the same date occupied very different social worlds. That mismatch is what makes it possible to pull the two apart and ask which one is actually doing the work.

A time trend by itself could reflect selection, migration, demography, drift, or some mixture of all of them. The harder question is whether the relevant axis is not just time, but civilization.

If you take two ancient populations from different social worlds, but do not let chronology do all the explanatory work, does their place on the civilizational ladder still matter?

The modern world has made one thing hard to miss. Cognitive performance, schooling, institutional complexity, and economic development cluster together. But once we move back into deep history, the picture becomes much hazier. We still know remarkably little about whether the genetic variants associated with educational attainment track not just the passage of time, but the emergence of more complex forms of social organization.

Gregory Clark argued that preindustrial societies may have undergone differential reproduction in ways that slowly shifted the distribution of traits favorable to economic success. Galor and Moav built a related idea into a formal evolutionary growth framework. In their model, the Malthusian world did not simply hold humanity down. It also created selection pressures that gradually favored traits complementary to human capital, technology, and later economic takeoff.

I am not testing their mechanism directly, but the question points in the same direction. If traits linked to educational attainment were historically relevant to the capacity of populations to sustain more complex forms of life, then archaeological stage should retain some predictive power even after absolute time is held constant. In that sense, the civilization-stage score is crude, but it is also useful. It is a rough proxy for where a population stood in the long transition from simple subsistence regimes to more knowledge-intensive and organizationally demanding societies.

At any given date, some populations were still living in relatively simple social worlds while others had already crossed into much more complex ones. And populations grouped into the same broad archaeological stage could be separated by very large spans of time. That mismatch is what makes the whole exercise possible. Once we exploit the fact that date and stage are misaligned, we can ask the real question: when educational-attainment polygenic scores rise through ancient time, are we just looking at chronology, or are we also looking at civilization?

The key identification fact is simple: archaeological stage and absolute date do not move in lockstep, as shown in Fig. 1.

Figure 1. Chronology and Civilization Stage Do Not Move in Lockstep

Brief Methods Note

This follow-up uses the AADR v66 2M panel. After excluding modern individuals, archaic and non-human samples, and groups lacking the variables needed for the model, the full ancient analysis set contains 12,221 individuals. The stricter metadata-only subset, in which archaeological periods come directly from metadata rather than heuristic assignment, contains 8,069 individuals. The main specification is a weighted mixed model with ancestry PCs, coverage, latitude, absolute date, and a random intercept for merged archaeological group (`Group_base`). The headline robustness checks focus on the metadata-only subset and on the <=12k BP window.

The focal trait is educational attainment because it is the best-powered available polygenic index for the broad cluster of abilities and dispositions that modern societies reward as human capital. Delay discounting was included for a different reason: as an exploratory proxy for the non-cognitive side of human capital, especially future orientation and willingness to defer immediate reward. Height was included as a comparison trait rather than a substantive target. If Height were flat while EA remained strongly structured, that would make the EA result easier to interpret as trait-specific rather than background structure.

Results

The first result is the one that motivated this entire line of inquiry: educational-attainment polygenic scores still rise through time in the new dataset. More recent ancient samples tend to have higher EA scores than older ones. That part of the story survives the larger AADR release, corrected labels, and score-specific weighting.

But the more important result is that chronology does not seem to exhaust the signal. The cleanest way to show that is to treat archaeological stage as an ordered outcome and ask whether higher EA scores predict assignment to later stages once absolute date, ancestry PCs, coverage, and latitude are controlled. In those ordered-probit models, the answer is yes. The association survives not just in the metadata-only subset, but also in the stricter metadata-plus-direct-date and metadata-plus-low-uncertainty subsets.

The main result is not simply that later samples score higher on EA. It is that, net of date and ancestry controls, higher EA scores are associated with later archaeological stage, and that relationship is more robust than the corresponding Height association in the stricter subsets.

The headline specification treats archaeological stage as an ordered outcome rather than a continuous score.

Table 1. Trait effects and Years BP controls across dating-quality subsets.

The real question is whether these results hold up once you start tightening the screws: stricter dating, metadata-only classifications, comparison traits, and alternative model specifications. That is where this either becomes interesting or falls apart.

Read more

Akbari’s explanation is that my studies were not cited because, in his view, they lacked proper calibration, did not adequately deal with inflation or multiple testing, and therefore did not count as relevant prior evidence for directional selection. That argument conflates two different claims. It treats “I think this paper is wrong” as equivalent to “this paper is irrelevant.” A paper can be disputable and still be prior work. A paper can be flawed and still be the earlier attempt to answer the same question. In science, criticism is a reason to cite and distinguish. It is not a reason to pretend the earlier paper does not exist. Nature’s own rule is not “cite only papers you approve of.” It is to acknowledge earlier contributions accurately.

The overlap here is not loose. In March 2024, Emil Kirkegaard and I published Evolutionary Trends of Polygenic Scores in European Populations From the Paleolithic to Modern Times in Twin Research and Human Genetics. That paper reported positive temporal trends for educational attainment, IQ, and SES-related scores in ancient Europe, alongside declines over time in schizophrenia, depression- and neuroticism-related scores. Akbari et al.’s West Eurasia preprint appeared in September 2024, and the Nature version was published on 15 April 2026. Their paper reports decreases in schizophrenia risk and increases in measures related to cognitive performance across West Eurasia. The overlap doesn't require the papers to be identical. Same broad region, same time-series logic, same cluster of cognitive and psychiatric traits, same directional story. They also report effects of deep ancestry on polygenic scores that broadly line up with what I had already found, including, for example, higher educational-attainment polygenic scores in Early European Farmers.

The overlap goes well beyond a few matching results. In my 2017 paper and again in my 2024 paper, I explicitly tested the broader Holocene accelerated-evolution model with ancient DNA, focusing on whether recent directional selection had pushed cognition-related variants upward through time. Akbari et al. later addressed that same basic model in the same broad West Eurasian context, with many of the same phenotypes and the same directional logic. The statistical machinery differs, but the core question does not.

The East Eurasian side is even harder to dismiss. On 30 January 2025, I published Directional Selection and Evolution of Polygenic Traits in Eastern Eurasia: Insights from Ancient DNA. I said explicitly that the paper was meant to test whether the earlier West Eurasian pattern could be replicated in the East. The answer, in that paper, was yes: education- and intelligence-related scores rose through time, while anxiety- and depression-related scores moved in the opposite direction. Then, on 4 April 2026, Barton et al. posted their preprint on convergent natural selection at both ends of Eurasia, with Akbari as one of the authors. By then I had already published the West Eurasian pattern and its East Eurasian extension.

That is why this is not just a dispute about “models.” Science tracks findings as well as frameworks. If one paper reports a directional increase in cognition-related polygenic scores and a directional decline in psychiatric-risk related scores over ancient time series, and a later paper reports the same broad pattern with more data or a more elaborate pipeline, the later paper is not starting from zero. It is confirming, refining, narrowing, or disputing an earlier result. Methodological superiority can justify a critical citation. It does not justify silence. Nature’s own guidance to reviewers only makes sense on that assumption: if the conclusions are not original, relevant references should be supplied.

This line of work also did not begin in 2024. Back in 2017, Woodley, Younuskunju, Balan, and I published Holocene Selection for Variants Associated With General Cognitive Ability: Comparing Ancient and Modern Genomes in Twin Research and Human Genetics. That paper compared ancient and modern genomes in order to test a Holocene selection hypothesis for cognitive-associated variants. It was not the same design as the later time-series papers, but it was obviously part of the same broader program: using ancient DNA to ask whether cognition-related alleles changed over the holocene.

The timeline alone makes the oversight defense implausible. Akbari et al.’s Nature paper shows “received” on 14 September 2024, “accepted” on 4 March 2026, and “published” on 15 April 2026. So this was not a manuscript that flashed by in a week. It sat in peer review for roughly a year and a half. I had already contacted the authors long before publication. Even if one imagines an email being overlooked, the papers themselves were sitting in the literature while this manuscript was being reviewed, revised, and defended. Akbari’s reply makes clear what happened in the end: not ignorance, but refusal.

The public peer review file makes his position look weaker, not stronger. Reviewers raised serious concerns about the statistical machinery, confounding, inflation, significance thresholds, and whether the polygenic claims were genuine or artifacts. One reviewer wrote that there were “major concerns about the statistical machinery and interpretations.” Another objected to the authors’ claim that they were exploiting something “not utilized in previous scans,” saying “This is not the case.” Another wrote that the title “rather overpromises.” The paper was not waved through on the ground that its authors had finally reached a uniquely rigorous standard that rendered all previous work irrelevant. It went through exactly what papers are supposed to go through: criticism, revision, qualification, and rebuttal.

By Akbari’s logic, any later author could erase prior work simply by declaring it methodologically insufficient, without any external scrutiny. But science cannot function that way. If that principle were accepted, citation would become a private reward system controlled by whoever publishes later from the stronger institution, with the bigger dataset, in the more prestigious journal. Smaller researchers would be acknowledged only when elite groups decided they were respectable enough. That is not a standard of rigor. It is a standard of power.

There are older examples of this problem. Mendel’s work was rediscovered in 1900 by later researchers who had obtained similar results, but rediscovery did not erase Mendel’s priority. Wegener’s continental-drift hypothesis was attacked for decades, yet plate tectonics did not make Wegener vanish from the history of the idea. Avery, MacLeod, and McCarty identified DNA as the transforming principle in 1944, and later work strengthened that case, but it did not make their paper irrelevant. In each case, later confirmation or methodological improvement changed the standing of the claim. It did not cancel the earlier contribution.

I am not claiming that my papers are beyond criticism. Of course they are open to criticism. They can be challenged, reanalyzed, corrected, or superseded. That is normal. What is not normal is the idea that directly relevant prior studies cease to count because a later author dislikes their methods. If anything, disagreement makes citation more necessary. Readers should be able to see the sequence of claims, the overlap in findings, and the points of dispute for themselves. Otherwise replication gets laundered into novelty, and the literature starts telling a false story about where an idea began.

Akbari is not just saying that he thinks my methods were weak. He is saying that this judgment frees him from the obligation to cite prior work that addressed the same question and reported earlier, overlapping results. That is not a defense of scientific standards. It is a rejection of one of the basic ways science keeps its own record honest.

You cite the earlier paper. You say where you think it went wrong. You explain what your method does better. You let readers see the continuity and the disagreement in the open. You do not reproduce the same broad empirical pattern, on the same class of traits, in the same historical frame, and then behave as though the earlier literature never existed.

Because this is no longer a matter of oversight, I am submitting a formal complaint to Nature about the failure to acknowledge directly relevant prior work. My aim is not to relitigate every methodological disagreement in public, but to correct the scientific record. If later authors believe an earlier paper is flawed, they are free to say so. What they are not free to do is treat it as though it never existed.

References

Piffer, D., & Kirkegaard, E. O. W. (2024). Evolutionary trends of polygenic scores in European populations from the Paleolithic to modern times. Twin Research and Human Genetics, 27(1), 30-49. https://doi.org/10.1017/thg.2024.8

Piffer, D. (2025). Directional selection and evolution of polygenic traits in Eastern Eurasia: Insights from ancient DNA. Twin Research and Human Genetics, 28(1). https://doi.org/10.1017/thg.2024.49

Piffer, D., & Kirkegaard, E. O. W. (2024). Predictive accuracy of polygenic scores from European GWAS among Chinese provinces. Mankind Quarterly, 65(1), 58-71. https://doi.org/10.46469/mq.2024.65.1.6

Woodley, M. A., Younuskunju, S., Balan, B., & Piffer, D. (2017). Holocene selection for variants associated with general cognitive ability: Comparing ancient and modern genomes. Twin Research and Human Genetics, 20(4), 271-280. https://doi.org/10.1017/thg.2017.37

Akbari, A., Barton, A. R., Gazal, S., Li, Z., Mah, M., Meyer, M., Mallick, S., Pinhasi, R., Rohland, N., Price, A. L., & Reich, D. (2024, September 15). Pervasive findings of directional selection realize the promise of ancient DNA to elucidate human adaptation [Preprint]. bioRxiv. https://doi.org/10.1101/2024.09.14.613021

Akbari, A., Perry, A., Barton, A. R., et al. (2026). Ancient DNA reveals pervasive directional selection across West Eurasia. Nature. https://doi.org/10.1038/s41586-026-10358-1

Barton, A. R., Rohland, N., Mallick, S., Pinhasi, R., Akbari, A., & Reich, D. (2026, April 4). Convergent natural selection at both ends of Eurasia during parallel radical lifestyle shifts in the last ten millennia [Preprint]. bioRxiv. https://doi.org/10.64898/2026.04.03.716344

Iceland has a population smaller than many cities, yet it has produced a remarkable number of strongmen, and the Nordic region more broadly has a long record of overperforming in sports where raw strength matters. Hafþór Júlíus Björnsson is the most famous modern example.

Let’s start with the sport where the northern pattern is hardest to dismiss.

I combined four men’s strongman title sets: World’s Strongest Man, the Arnold Strongman Classic, the IFSA World Championship, and Europe’s Strongest Man. For each country, I summed historical gold, silver, and bronze finishes across those competitions, scored them 3, 2, and 1, and then divided by 2024 population to get per-capita rates. One caveat is obvious: because Europe’s Strongest Man is a continental event, this table is a little more favorable to European countries than a pure world-only measure would be.

Combined major strongman performance per capita

Iceland comes first by a huge margin. Lithuania is second. Then comes a Nordic cluster: Finland third, Norway fourth, Sweden fifth, and Denmark ninth.

Strongman is also the sport in which a muscle-related advantage should show up most clearly. There are no weight classes, and success depends heavily on absolute strength and body mass. At the same time, the Nordic countries have a strong cultural investment in this kind of competition, so genes are unlikely to be the whole explanation. Culture and training almost certainly play a role too.

The strongman numbers are strong enough that they demand an explanation. Culture is part of it. But culture is probably not all of it. A recent genetics study adds a new piece to the puzzle.

Read more

When Answers Become Cheap

For a long time, doing research meant being unusually good at producing answers. You had to know the literature, track down obscure references, summarize complex debates, run the analysis, write the paper, and connect scattered facts into something coherent. The glamour was in the breakthrough, but the daily reality was answer-production.

AI is changing that.

Not because it has solved research, and not because human expertise is suddenly obsolete. The change is subtler. AI is making many kinds of answers cheap. It can already generate competent summaries, first-draft explanations, code, critiques, outlines, and literature reviews in seconds. Some of these are shallow, some are flawed, and almost all need checking. But that is beside the main point. The cost of producing plausible answers is collapsing.

And whenever something becomes cheap, something else becomes scarce.

In research, that scarce thing is increasingly the question.

Einstein’s Beam of Light

There is a reason Einstein is still the canonical example here. Before he had a theory, he had a question. He later recalled asking himself, at sixteen, what would happen if he could chase a beam of light. In his own words: “If I pursue a beam of light with the velocity c...”

What matters is not just the romance of the anecdote. It is the sequence. The question came first. The answer came much later.

That is often how real intellectual change begins. Not with a polished result, but with a strange question that makes the existing picture suddenly look unstable. The greatest researchers are often not the people who merely answer the most questions. They are the people who ask the one question that makes an entire field look different.

What AI Is Actually Changing

I do not mean the vague classroom advice that one should be “curious.” I mean the much harder intellectual act of asking a question that is actually worth answering. A question that is not banal, not malformed, not conceptually confused, not impossible to test, and not already answered in slightly different language three hundred times before. A question that reveals something genuinely new instead of merely generating text.

That part is becoming more important because large language models are built to answer prompts, not to originate the deepest ones. They can help brainstorm questions, of course. But even then they are operating inside a frame someone else has already chosen. They do not decide, in the human sense, what is missing from a field. They do not feel the irritation of an unresolved contradiction. They do not know which puzzle is fundamental and which one is just ornamental.

That means the human role in research moves one level up. Less effort goes into manufacturing prose, code, and synthesis from scratch. More effort goes into deciding what deserves to be investigated in the first place.

From Possession to Selection

That is a profound shift. In the old research economy, the premium was often on possession. Who had the knowledge, the methods, the technical skills, the hours, the access?

In the new research economy, the premium shifts toward selection.

Which question matters? Which framing actually holds up? Which comparison would be illuminating rather than decorative? Which hypothesis would survive contact with evidence rather than merely sounding impressive?

This does not mean answers stop mattering. It means first-pass answers matter less, because they are no longer rare. The real bottleneck becomes judgment. Judgment about where to look, what counts as evidence, and whether a polished answer is actually responsive to the problem or simply a clever extrapolation from the prompt.

The New Danger: Synthetic Intellectual Sludge

In fact, AI may make bad questions more dangerous than before.

When answers were expensive, weak questions often died quietly because they were too costly to pursue. Now a weak question can generate ten pages of a confident, worthless answer (the classic “AI slop”) in a minute. We may be entering an era where answers outrun the questions that warranted them, where the problem is not lack of output but lack of discrimination.

The challenge will not be getting answers. It will be stopping the flood of answers to questions that never mattered.

That is why the ability to ask a good question is not becoming a soft skill. It is becoming the hardest part of research.

Why the Best Researchers May Pull Further Ahead

AI will probably raise the floor for research. Many more people will be able to produce decent summaries, passable code, competent literature reviews, and respectable drafts. That democratization is real.

But it may also raise the ceiling for a smaller group who have genuine taste in problem selection. Because once everyone has access to decent answer-generation, the differentiator becomes the quality of the initial intellectual move. Seeing the hidden variable. Spotting the neglected comparison. Asking the one question that makes ten others unnecessary.

So AI may flatten some kinds of expertise while making others more decisive. Average work gets easier to produce. That's exactly why unusual judgment becomes harder to fake.

Research Becomes More About Judgment

The deepest shift is this: research is moving from being primarily a test of memory, synthesis, and execution to being increasingly a test of judgment.

Judgment about what matters. Judgment about what is missing. Judgment about how to phrase the problem. Judgment about what evidence would actually settle it. Judgment about when an answer is fake, premature, trivial, or contaminated by the wording of the question itself.

That is why asking the right questions should not be treated as a motivational cliché. In the age of AI, it becomes the central intellectual skill.

Conclusion

Einstein did not begin with relativity. He began with a teenager’s strange question about what it would feel like to chase a beam of light. The answer took years. The question took a moment, and it took everything.

That is where research is heading. Not toward less intelligence, but toward intelligence applied earlier, before the work begins, at the point where someone has to decide what is actually worth knowing. That decision is becoming harder to fake and harder to delegate. It may turn out to be the whole game.

That is strange, because laziness is one of the most familiar facts of ordinary life. Everyone knows the phenomenon. A person keeps postponing a task that matters, not because he does not understand the reward, not because the reward is too small, and not even because the reward is too far in the future, but because he simply recoils from the effort itself. He does not want to begin. He does not want to persist. He does not want to bear the friction.

Yet once you try to formalize that intuition, the concept becomes morally radioactive. Laziness is not a neutral modern label. It carries the residue of sloth, vice, and blame. It is one of the seven deadly sins. And that may be one reason the scientific literature has mostly approached the territory indirectly, through safer and cleaner constructs like conscientiousness, self-control, and delay discounting, rather than asking head-on whether some people are simply more averse to exertion than others.

That caution is understandable. No serious psychology should smuggle theology in through the back door. But the opposite error is also possible. Once a trait becomes morally charged, researchers become unusually hesitant to treat it as a real dimension of human variation at all. They prefer to rename it, dissolve it into environment, or absorb it into broader constructs that sound less accusatory. In that sense, laziness may represent a kind of scientific blind spot: not because the phenomenon is unreal, but because the word is too morally dangerous to touch directly.

The result is that psychology has become much better at measuring impatience than at measuring unwillingness to exert effort.

The classic example is the marshmallow task. A child is offered one treat now or two if he waits. For decades, that experiment served as the popular emblem of delayed gratification and future-oriented discipline. But later work greatly complicated the legend. The predictive power of the task turned out to be much smaller than the original mythology implied once background variables were taken more seriously, and more recent work has continued to treat the task more cautiously than its pop-psych afterlife would suggest. The marshmallow task still captures something real, but what it captures is mainly willingness to wait. It does not cleanly capture willingness to work.

That distinction is important, because willingness to wait and willingness to work are not the same thing.

A person may fully accept that a future reward is valuable, may be perfectly capable of waiting for it in principle, and yet still avoid the actions required to obtain it. Not because he discounts the future steeply, but because the path itself feels aversive. The task is tedious, draining, monotonous, frustrating, attentionally costly, or simply more unpleasant than its alternatives. The problem is not time preference. The problem is effort cost.

This is where I think the concept of laziness should be sharpened.

Laziness is not just low conscientiousness. Conscientiousness is too broad. It bundles together reliability, orderliness, dutifulness, self-discipline, and persistence. Those things overlap with laziness, but they are not identical to it. Nor is laziness just procrastination, because procrastination can arise from anxiety, self-doubt, fear of failure, or perfectionism. Nor is laziness simply impatience, because impatience is about devaluing rewards that arrive later.

Laziness, in the stricter sense, is better understood as effort aversion net of impatience.

That raises the next question. What exactly is effort?

The wrong answer is to treat it as something mystical or merely introspective. The overly narrow answer, on the other hand, is to reduce it to calories alone. The calorie-based view has an enormous advantage: it grounds effort in a real, measurable physical quantity. That should not be dismissed prematurely. In physical tasks, effort clearly does involve energy expenditure, and the idea that humans evolved to economize scarce energy has obvious plausibility. At the same time, the literature on cognitive effort strongly suggests that the subjective cost of effort is not exhausted by immediate metabolic burn. Researchers often define cognitive effort in terms of the way rewards lose subjective value when they require more mental work, and some influential theories interpret the feeling of effort as partly an opportunity-cost signal, not just raw fuel consumption.

So effort should be defined more carefully as the subjective cost of mobilizing and sustaining resources toward a goal. Sometimes that cost is physical exertion in the narrow energetic sense. But it can also include sustained attention, inhibition of distractions, monotony, frustration, cognitive strain, and the sacrifice of more pleasant alternatives. Pain avoidance belongs in this territory, but it does not exhaust it. Effort is broader than pain, because even painless tasks can feel effortful when they demand concentration, persistence, or self-control. And effort is broader than calories, because two people can face the same objective demand while experiencing very different subjective costs.

Once you define effort that way, the scientific blind spot becomes easier to see. We already have decent tools for measuring time preference. We have much less satisfactory tools for measuring the sheer unwillingness to incur effort costs after time preference has been taken into account.

That is the missing variable.

A person may avoid a task for at least four different reasons. He may not value the outcome. He may value it but discount it because it is delayed. He may doubt his own ability and therefore expect failure. Or he may understand the reward perfectly well, believe he can achieve it, and still recoil from the effort required. Only the last case is really close to what ordinary language has long meant by laziness.

If that is right, then laziness should not be measured by one familiar proxy like conscientiousness or one iconic task like the marshmallow test. It should be measured by a battery designed to separate effort aversion from its nearest confounds.

The first part of that battery would be a straightforward delay-discounting task. This is the easy part. Offer choices between smaller-sooner and larger-later rewards and estimate how steeply the person discounts delayed outcomes. That gives you a measure of impatience.

The second part would be an effort-discounting task. Here the delay is held constant or explicitly modeled, while the required effort varies. Rewards lose value as they become more effortful to obtain. This can be done with physical effort, cognitive effort, or both. In the cognitive literature, paradigms like COG-ED were built precisely to quantify the subjective value of cognitive effort, and a growing literature now treats cognitive and physical effort-based decision making as related but not identical domains.

The third part would control for pain sensitivity and general aversiveness. Otherwise critics could always argue that what you are measuring is not laziness but a broader tendency to avoid unpleasant experiences of any kind. The fourth part would control for expected success and ability. Someone who does not start because he believes he will fail is not lazy in the same sense as someone who could succeed but hates the exertion. The fifth part, which matters most for ordinary life, would be a real-world initiation and persistence measure: how quickly people begin low-glamour tasks, how often they abandon them, how often they choose lower-effort routes when the payoff is otherwise similar, and how consistently they follow through across days or weeks.

Only after all that would you have something that deserves to be called a measure of laziness.

In formal terms, the construct would look something like this: laziness is the residual tendency to avoid action after controlling for delay discounting, expected success, ability, and generalized aversiveness. In plainer English, it is the reluctance to bear the felt cost of exertion even when the future payoff is clear and worth having.

What makes this especially striking is how little standardized work has been done under the actual name. Only recently has the literature even started to produce dedicated laziness scales, including a 2025 Laziness Assessment Scale whose authors explicitly present it as addressing the absence of a standardized measure. That alone tells you that this is still conceptually underdeveloped territory.

The larger point is not moralistic but analytical. Human beings differ not only in patience, intelligence, and conscientiousness, but plausibly also in their tolerance for exertion itself. Some people are willing to pay effort costs over and over again in pursuit of remote goals. Others are far more likely to stall, drift, postpone, or choose the path of least resistance even when they understand perfectly well what they are sacrificing. A science that measures the first pattern but treats the second as conceptually embarrassing is leaving out something important.

So perhaps the right way to think about laziness is not as a sin, and not as a loose insult, but as a trait-like disposition toward effort avoidance once impatience has been subtracted away.

The marshmallow task measures who hates waiting. A real laziness science would measure who hates exertion.

And until psychology does that more cleanly, it will keep confusing two very different human problems: the inability to resist the present, and the unwillingness to bear the cost of effort.

Share

But time is not the only thing changing across that span. Social and technological complexity are changing too. A Paleolithic forager band and an Iron Age kingdom are not just separated by millennia. They are separated by farming, writing, cities, specialization, storage, hierarchy, and the entire accumulated weight of what we loosely call civilization. So here is the question worth asking: when those polygenic scores rise through ancient history, are we really just watching a clock tick or are we watching something track the emergence of more complex ways of organizing human life?

Using ancient individuals from the AADR dataset, I assigned each archaeological period a civilization-stage score and asked whether that score predicts educational-attainment polygenic scores even after controlling for absolute date. In practice, that meant coding Paleo-Mesolithic groups as 1, Neolithic groups as 2, Copper Age groups as 3, Bronze Age groups as 4, and Iron Age groups as 5.

The leverage comes from a basic fact of world prehistory: civilization and chronology do not move in lockstep. Greece had farming millennia before Britain. Some populations at the same date occupied very different social worlds. That mismatch is what makes it possible to pull the two apart and ask which one is actually doing the work.

A time trend by itself could reflect selection, migration, demography, drift, or some mixture of all of them. The harder question is whether the relevant axis is not just time, but civilization.

If you take two ancient populations from different social worlds, but do not let chronology do all the explanatory work, does their place on the civilizational ladder still matter?

The modern world has made one thing hard to miss. Cognitive performance, schooling, institutional complexity, and economic development cluster together. But once we move back into deep history, the picture becomes much hazier. We still know remarkably little about whether the genetic variants associated with educational attainment track not just the passage of time, but the emergence of more complex forms of social organization.

Gregory Clark (2007) argued that preindustrial societies may have undergone differential reproduction in ways that slowly shifted the distribution of traits favorable to economic success. Galor and Moav (2002) built a related idea into a formal evolutionary growth framework. In their model, the Malthusian world did not simply hold humanity down. It also created selection pressures that gradually favored traits complementary to human capital, technology, and later economic takeoff. On that view, the Industrial Revolution was not a bolt from the blue. It was the late visible phase of a much longer process.

I'm not testing their mechanism directly but the question points in the same direction. If traits linked to educational attainment were historically relevant to the capacity of populations to sustain more complex forms of life, then archaeological stage should retain some predictive power even after absolute time is held constant. In that sense, the civilization-stage score is crude, but it is also useful. It is a rough proxy for where a population stood in the long transition from simple subsistence regimes to more knowledge-intensive and organizationally demanding societies.

At any given date, some populations were still living in relatively simple social worlds while others had already crossed into much more complex ones. And populations grouped into the same broad archaeological stage could be separated by very large spans of time.

A map of the spread of agriculture makes the point immediately. The transition to farming did not unfold at a uniform pace across Europe. Some regions crossed into the Neolithic much earlier than others, while large parts of the continent remained outside it for centuries or millennia longer. That staggered transition is exactly what makes the present analysis possible. If different populations occupied different civilizational stages at the same broad date, then archaeological stage and chronology can, at least in principle, be pulled apart.

The sample is global in scope, not limited to Europe, which provides the statistical power needed to detect meaningful mismatches.

That mismatch is what makes the whole exercise possible. Once we exploit the fact that date and stage are misaligned, we can ask the real question: when educational-attainment polygenic scores rise through ancient time, are we just looking at chronology, or are we also looking at civilization?

How I tested it

I used ancient individuals from the public AADR 1240k dataset and grouped them into five broad archaeological stages, from Paleo-Mesolithic to Iron Age. These were coded from 1 to 5 as a rough ordinal measure of civilizational complexity.

The key outcome was each individual’s educational-attainment polygenic score, derived from ancient DNA. I then asked a simple question: does civilization stage still predict that score once absolute date is held constant?

To test this, I ran models on both the full 45,000-year sample and the Holocene alone, controlling for date, genomic quality, geography, and, in the main specifications, ancestry. That last step matters because it reduces the risk of mistaking large-scale population movements for a civilization effect.

I also reversed the analysis by treating civilization stage itself as the outcome in ordered probit models. That made it possible to ask the same question from the other direction: are individuals with higher educational-attainment polygenic scores more likely to belong to later archaeological stages, even after the same controls? If the pattern holds both ways, it is less likely to be a quirk of one model.

Time and civilization are related, but not the same thing

Archaeological stage and calendar date are correlated, but far from perfectly. Populations living at similar dates can occupy very different civilizational positions, and populations assigned to the same broad stage can be separated by long stretches of time. This can be seen in the figure below.

Mismatch Between Time and Civilization Stage

If civilization stage predicts EA-linked polygenic scores after time is held constant, then the result cannot be reduced to a simple temporal climb.

What happens in the full sample

The civilization-stage score is a clear positive predictor of educational-attainment polygenic scores even after controlling for date, coverage, and latitude.

Paid subscribers get the full coefficient plots, robustness checks, ancestry-controlled models, and my detailed interpretation of what this result does and does not mean

Read more

That question came up recently when Remitly published a ranking of the world’s “most generous nations.” Their index is based on the Interpersonal Generosity Scale, a psychological measure built around self-descriptions of kindness, empathy, attentiveness, and willingness to put others first. Remitly says it surveyed 4,500+ adults in 25 countries through Prolific, using 10 first-person statements scored on a 6-point agree-disagree scale, and then averaged those scores by country. In other words, this is not mainly a measure of charity or volunteering. It is a measure of how generous people say they are in everyday interpersonal life.

That makes it interesting, but also raises a deeper question. If a country scores highly on a self-report generosity scale, does it also score highly on a more behavioral measure of giving?

To check that, I compared the Remitly ranking to the CAF World Giving Index. CAF’s measure is still survey-based, so it is not “objective” in the strict sense. But it is closer to behavior than to self-description. It asks whether people, in the past month, have donated money to charity, volunteered their time, or helped a stranger. CAF then averages the positive responses into a country score. In the latest edition, the index covers 142 countries, draws on Gallup World Poll data, and uses the same three core questions that have anchored the series for years.

So this is really a comparison between two different ideas of generosity.

The first is interpersonal generosity: being caring, supportive, kind, available, and willing to put others first. The second is reported prosocial behavior: giving money, giving time, and helping strangers. These are not identical. But they should overlap if both are tapping into a common underlying trait or social norm.

When I matched the 25 countries in the Remitly ranking to their CAF World Giving Index scores, the relationship was positive but only moderate. The Pearson correlation was about 0.45. That is enough to say the two measures are related, but nowhere near enough to say they are measuring the same thing. Some countries are high on both. Others diverge sharply. Japan and Poland score low on both measures, which pushes the relationship upward. But South Africa, for example, looks much stronger on the interpersonal self-report scale than on the CAF behavior-based index.

The first thing to ask is whether countries that describe themselves as generous also look generous on a more behavior-based index.

Read more

I do not think that conclusion follows from their analysis.

The disagreement is not over whether polygenic scores should be used. Barton et al. used them, and so did I. The real difference lies in how they were analyzed. In my 2025 Twin Research and Human Genetics paper, I also got null results with a European-based pigmentation score in East Asians. However, once an East Asian-based pigmentation score (Kim et al., 2024) is used, my direct regression approach detects a temporal increase toward the present, whereas Barton et al.’s framework still does not.

Consequently, their negative result is more plausibly read as a limitation of method than as proof that East Asian depigmentation was largely over before the Holocene.

Their method helps explain why. The core of their paper is a GLMM designed to detect consistent directional allele-frequency change over time. In their own description, it tests for “consistent trends in allele frequency change over time” that can be explained by directional selection. They also say quite openly that heterogeneous samples help when selection pushes in a consistent direction across space and time, but lose power when the pattern fluctuates. This gives us a hint to why pigmentation is a difficult trait for this framework in East Eurasia, where geography, ancestry, and latitude all structure the phenotype.

Barton et al. then extend the analysis to polygenic traits, but still within that same general logic: they ask whether pigmentation emerges as a strong, coherent signal in a very broad and heterogeneous ancient sample. For pigmentation, even after switching from mainly European GWAS weights to East Asian-based scores, they still report no significant result after multiple-testing correction. But that is not the same question I asked in my paper. I modeled the East Asian-based pigmentation score directly at the individual level, regressed it on date, and then added ancestry, geography, and coverage as covariates.

The temporal trend is not a fragile bivariate pattern either. It remains in the multivariable models after ancestry and geographic controls are added.

Two details are worth keeping in mind when reading this plot. First, Southern China is the reference group, so the ancestry coefficients are interpreted relative to it. Second, Date is measured in years BP, which means a negative coefficient for Date indicates lower light-skin PGS in older samples, i.e. darker predicted pigmentation deeper in the past.

Barton et al.’s GLMM is built to detect consistent allele-frequency change across a heterogeneous macro-regional sample, and they explicitly note that such samples gain power when selection pushes in the same direction over space and time but lose power when the pattern fluctuates. Their trait-level statistics, g and gsign, inherit the same logic: they ask whether many trait-associated variants move in a coordinated way strongly enough to produce a significant aggregate signal. My regression asks a different question. Using an East Asian-based pigmentation score, it tests whether the net individual-level pigmentation score rises toward the present after ancestry, geography, and coverage are controlled. That makes it better suited to a trait whose history is spatially structured and ancestry-dependent rather than continent-wide and uniform.

The divergence becomes obvious when you look directly at the fitted temporal trend from the East Asian-based pigmentation score.

Unlock the analysis of why the historical signal reappears.

Read more

At first I told myself it was probably just an oversight. The literature is big, and these things happen. But after seeing my work replicated without citation more than once, I got tired of pretending that was a satisfying explanation. This latest case was the last straw.

That is why I am writing this post.

I said it more bluntly on X: “I love how Akbari and Reich keep replicating my results without bothering to cite my papers!” This post is my attempt to document it carefully, because too many readers, and some hostile commenters, seem unaware that I had already published in peer-reviewed journals on these questions before these later papers appeared.

This isn't just about my ego. There's a real scientific problem here. Credit and replication are two pillars of science, but they belong to different sides of it. Giving credit is part of the ethical structure of science: it is how the record stays honest about who contributed what. Replication is part of the methodological structure of science: it is how findings become more reliable. Both are damaged when later researchers reproduce an earlier result while presenting it as if it emerged for the first time in their hands.

Replication should strengthen science. But it only does so when it is acknowledged as replication, confirmation, or extension of earlier work. When prior work is omitted, replication is turned into apparent novelty, and the scientific record becomes distorted.

The Timeline

The sequence of events matters for the record:

• In 2024, I published a paper on temporal trends in polygenic scores in ancient European populations. The abstract reported “positive directional selection for EA, IQ, and SES traits,” but the paper’s main findings went beyond that. In the discussion, I wrote that my results “indicated a decrease in PGS for traits such as neuroticism, depression and schizophrenia,” and I added that I had “observed an elevated schizophrenia PGS in ancient samples,” implying a decline toward the present.

• Later that same year, Akbari et al. posted their West Eurasia preprint. In their abstract, they wrote that they identify combinations of alleles associated with “lower risk for schizophrenia and bipolar disease” and “increased measures related to cognitive performance,” including “scores on intelligence tests, household income, and years of schooling.”

• Then, in early 2025, I published a second paper extending this same framework to Eastern Eurasia. I was explicit about what I was doing: “This article aims to replicate these findings,” referring back to my earlier European paper. The results were clear. In the abstract I reported “positive selection for cognitive-related traits such as IQ, EA and autism spectrum disorder (ASD), alongside negative selection for anxiety and depression.” In the discussion I made the pattern even more concrete: “EA, IQ, and ASD displayed negative correlations with sample age, indicating increases over time likely driven by positive directional selection,” while “anxiety, depression, and (dark) skin color PGSs showed positive correlations with Years BP, consistent with decreases over time.” I then stated plainly: “These results replicate findings from prior research on ancient West Eurasian genomes.” In other words, before Barton et al. published their East Eurasia paper, I had already put into print the basic claim that Eastern Eurasia showed the same kind of long-term polygenic shifts I had previously reported for Europe. Barton et al. later framed their paper around “highly correlated signals of adaptation” in East and West Eurasia and “convergent evolution in response to parallel transitions to food producing economies.” This is the same broad comparative argument, presented later and without citation.

• Skin pigmentation makes the same point in a slightly different way. In my East Eurasia paper, I did not just discuss cognition and psychiatric traits; I also reported that “anxiety, depression, and (dark) skin color PGSs showed positive correlations with Years BP, consistent with decreases over time,” and added that the East Asian light-skin score “confirmed evidence of negative selection for darker skin (or positive selection for lighter skin).” Barton et al. also treated skin pigmentation as a major East Eurasian result, but as an exception: in their abstract they wrote that “most skin lightening in East Asians arose prior to the Holocene,” and later said they “failed to observe strong selection at skin pigmentation in the East Eurasian cohort.” We were addressing the same trait in the same region. The difference is that they presented a different conclusion. But that only strengthens the case for citation, not weakens it. When a later paper revisits the same trait and reaches a different or narrower conclusion, the earlier paper should be part of the conversation, not erased from it.

I had even written a Substack essay asking whether the same SNPs move through time in similar directions in Europe and East Asia. But the buildup to that comparative angle did not begin on Substack. Before that, I had already published a paper on Chinese provinces showing that European-derived polygenic scores could meaningfully track variation inside an East Asian population. In my Chinese provinces paper, I explicitly cited that study and wrote that it showed European-derived PGSs “predict differences in mean phenotypic traits across contemporary Chinese provinces,” and that these findings “support the utility of European-derived PGSs for predicting variation within East Asian populations.” So by the time Barton et al. framed their East Eurasia paper around “highly correlated signals of adaptation” and “convergent evolution” across East and West Eurasia, I had already laid out that broader Europe–East Asia comparative logic in both a published paper and a public essay.

Prestige is not an excuse

Sure, two groups can land on the same question independently. That happens. But when that happens, you cite the earlier work instead of pretending the conversation began with you. Science is not weakened by replication; it depends on it. If another group revisits a question I have already explored and finds similar results, that should be a good thing. Replication is one of the ways science corrects itself and gains confidence in a finding.

The problem begins when replication is framed as novelty. A later paper may have a larger sample, more funding, and a more prestigious author list, but that does not erase the work that came before it. On the contrary, it creates a greater responsibility to distinguish clearly between what is genuinely new and what is confirming, extending, or repackaging earlier results. Otherwise the scientific record is distorted. Readers come away with the impression that the idea originated with the most powerful lab rather than with the person who first published it.

The Choice of Silence

The authors were contacted over a year ago and given the opportunity to respond or amend the record.

They did not reply.

I have now contacted the authors again by email and am waiting for their response. If they reply in a substantive way, I will update this post accordingly.

Conclusion

I'm posting this because I have no other option. I'm not at a major institution. I can't call in a favor or apply quiet pressure. The usual advice is to stay polite and let it go — which is easy advice to give when your name isn't the one getting dropped. We contacted them. They didn't respond. So here we are.

The record should be clear:

• I published on temporal polygenic trends in ancient Europe before the Akbari West Eurasia paper appeared.

• I published on these same questions in Eastern Eurasia before their East Eurasia paper appeared.

• I used the same methodologies to address the East-West comparative angle before their convergence paper appeared.

A healthy scientific culture shouldn’t allow “Big Science” to quietly overwrite the work of smaller researchers simply because they have more resources. Citation is the bare minimum we should expect.

References

Akbari, A., Barton, A. R., Gazal, S., Li, Z., Kariminejad, M., Perry, A., Zeng, Y., Mittnik, A., Patterson, N., Mah, M., Zhou, X., Price, A. L., Lander, E. S., Pinhasi, R., Rohland, N., Mallick, S., & Reich, D. (2024, September 15). Pervasive findings of directional selection realize the promise of ancient DNA to elucidate human adaptation [Preprint]. bioRxiv. https://doi.org/10.1101/2024.09.14.613021

Barton, A. R., Rohland, N., Mallick, S., Pinhasi, R., Akbari, A., & Reich, D. (2026, April 4). Convergent natural selection at both ends of Eurasia during parallel radical lifestyle shifts in the last ten millennia [Preprint]. bioRxiv. https://doi.org/10.64898/2026.04.03.716344

Piffer, D. (2025). Directional selection and evolution of polygenic traits in Eastern Eurasia: Insights from ancient DNA. Twin Research and Human Genetics, 28, 1–20. https://doi.org/10.1017/thg.2024.49

Piffer, D., & Kirkegaard, E. O. W. (2024). Evolutionary trends of polygenic scores in European populations from the Paleolithic to modern times. Twin Research and Human Genetics, 27(1), 30–49. https://doi.org/10.1017/thg.2024.8

Piffer, D., & Kirkegaard, E. O. W. (2024). Predictive accuracy of polygenic scores from European GWAS among Chinese provinces. Mankind Quarterly, 65(1), 58–71. https://doi.org/10.46469/mq.2024.65.1.6

The paper was small and I would not oversell it today. It involved 96 participants, most of them students, and used a battery that included the Big Five, temperament measures from the short TEMPS-A, schizotypy-related scales, divergent thinking tasks, a very short Remote Associates Test (RAT), and the Creative Achievement Questionnaire. Still, I think it asked the right question. I was not only interested in whether “creative people are high in openness,” which was already well known. I wanted to know whether artistic and scientific creativity might rest on somewhat different psychological foundations.

Openness to experience was the broadest and most consistent predictor. It was positively associated with ideational fluency, with artistic achievement, and with overall creative achievement. That part fit the existing literature nicely. However, when I extracted broader personality factors, the clearest extra predictive signal beyond the Big Five appeared for CAQ Science, not for CAQ Art. In the discussion, I interpreted one of those factors as closer to impulsivity or dominance, and it drew together low conscientiousness, hypomania, unconventionality, risk taking, aggression, and impulsivity. Scientific creative achievement also related more strongly to RAT performance than to simple fluency, whereas artistic achievement related more to ideational fluency. In other words, artistic and scientific creativity did not seem to live in exactly the same psychological neighborhood.

Subscribe now

That, to me, was always the cool part of the paper. The point was not just that creative people are curious, imaginative, and open. The point was that scientific creativity seemed to sit a little closer to an activated, mildly hypomanic-spectrum temperament profile. Not full-blown pathology. Not romantic nonsense about suffering as the source of genius. Something more specific: energy, approach motivation, reduced inhibition, nonconformity, and the tendency to push into novelty rather than shrink from it.

What has happened since then is that the literature has become both more cautious and, in a way, more supportive. The most important paper for framing this is Baas, Nijstad, Boot, and De Dreu’s 2016 Psychological Bulletin review, Mad Genius Revisited. Their core argument was that the literature looks contradictory only because people lump together very different forms of psychopathology. They distinguish approach-based vulnerabilities, such as positive schizotypy and bipolar risk, from avoidance-based vulnerabilities, such as anxiety, depressive mood, and negative schizotypy. Their meta-analytic results fit that distinction: risk of bipolar disorder, including hypomanic and manic liability, was positively associated with creativity, whereas depressive mood showed a weak negative association.

Baas et al. suggest that some forms of creativity benefit from a temperamental style that increases exploration, broadens associative search, and lowers the threshold for trying unusual ideas. That is much closer to what I thought I was picking up in 2014 with the combination of hypomanic-spectrum temperament, impulsive nonconformity, risk taking, and scientific creative achievement.

Later work on bipolarity and creativity points in the same general direction, though in a more modest tone. A 2023 review and meta-analysis by Forthmann and colleagues focusing specifically on bipolar disorder and creative cognitive potential, usually measured by divergent thinking, reported a small overall positive relationship and found that results varied by bipolar status, with euthymic and subclinical presentations looking more creativity-compatible than depressive states. That is exactly the sort of nuance the old “mad genius” cliché usually misses: the link, when it exists, seems to involve activated or subclinical states much more than debilitating ones.

A 2024 review by Mathias Benedek adds another useful correction. It argues that creative potential, including divergent thinking, explains only limited variance in actual creative achievement and that the path from potential to achievement is mediated by behavior, expertise, environmental support, and domain specificity. That fits my own old results rather well. In the 2014 paper, ideational fluency mattered, but it did not simply turn into real-world achievement across all domains. Scientific achievement related more to RAT-like associative problem solving, artistic achievement more to fluency, and the personality profile mattered differently across domains. Creativity in the lab and creativity in the world are related, but they are not the same thing.

So far, so good. One can tell a reasonably coherent story. Openness is the broad base trait. Scientific and artistic creativity differ. Mildly hypomanic or approach-oriented temperamental traits may sometimes help, particularly in forms of creativity that reward drive, novelty seeking, and boundary pushing. The literature is not tidy, but it is no longer as confused as it once seemed.

And this is where things become frustrating from a genomics perspective.

The GWAS Blind Spot

This is where the story becomes genuinely frustrating from a genomics perspective. It is no longer literally true that creativity has been ignored altogether by molecular genetics. There is at least one GWAS based on psychometric creativity measures, but it analyzed only about 4,664 Han Chinese participants after exclusions—far too small for a trait that is almost certainly highly polygenic.

There is also now a much larger 2024 study, but its solution to the phenotype problem is revealing. It does not measure creativity directly with divergent-thinking tasks, the Creative Achievement Questionnaire, or anything comparably rich. Instead, it uses occupational creativity: people are classified into artistic, scientific, or managerial “creative occupations” based on job codes, and a continuous “creative achievement” variable is built from three O*NET items about how much originality, idea fluency, and creative thinking the occupation requires. That is clever as a scaling strategy, but it is also a weak substitute for the thing psychologists usually mean by creativity. It is closer to the genetics of sorting into jobs that are labeled creative than to the genetics of creativity itself.

That is why I still think there is a real blind spot here. The small study is closer to the phenotype but badly underpowered. The large study has the power, but the phenotype is poor. Even that 2024 paper acknowledges both the domain-specificity problem and the statistical limits of the current literature, and notes that creativity phenotypes are highly polygenic.

What makes this harder to excuse is that the phenotype problem is real, but not insurmountable. We are not dealing with some trait so elusive that it cannot be measured outside a laboratory. The Creative Achievement Questionnaire was explicitly designed to be objective, empirically valid, and easy to administer and score. It is imperfect, of course, but it is a far more direct measure of real-world creative output than classifying someone as creative because of their occupation code.

So the real issue is not that no one has ever tried. It is that modern genomics has still not produced what one would actually want: a large, well-powered GWAS of direct creativity phenotypes—creative achievement, divergent thinking, scientific originality, artistic production—rather than occupational proxies. That absence is odd given the enormous cultural interest in genius and originality, and given how easy it would be to add at least a brief self-report creative-achievement instrument to large cohorts.

And the stakes are not trivial. If the genetics of creativity were understood even modestly well, the consequences would extend beyond psychology into education, talent identification, and eventually reproductive decision-making. That is precisely why this blind spot matters. Creativity is one of the traits most tightly bound up with human progress, yet at the molecular level we still understand it much more poorly than many adjacent traits.

Share PifferPilfer

Refer a friend

The mistake is to push that idea too far and treat R1b as if it were an ancestry percentage. A Y-chromosome haplogroup can track a real historical process without measuring genome-wide ancestry in any simple one-to-one way.

The Basques are a good test case. They have very high frequencies of R1b, comparable to those seen in several populations farther north in Europe, as shown in the map below.

Figure 1. R1b Y- haplogroup frequency

Yet their genome-wide ancestry profile is not northern European. In plots of ancestry components, Basques typically show much less Steppe-related ancestry and much more ancestry ultimately deriving from Europe’s earlier Neolithic populations. The contrast is clear in the figure below, reproduced from Evolution: On the origin of Basques.

Figure 2. Deep ancestries in modern European populations.

If R1b were a straightforward ancestry meter, populations with similarly high frequencies of R1b should also look broadly similar in genome-wide ancestry. But they do not. Basques and more northerly Europeans can overlap in paternal lineage frequency while differing substantially in their autosomal ancestry profiles.

The key point is that the same haplogroup can be carried by populations with different autosomal backgrounds. What needs explaining, then, is not why Basques have R1b at all, but why a population with such high R1b can remain much less Steppe-shifted than northern Europeans with similar haplogroup frequencies. This is what I call the “Basque paradox”.

The answer, I will argue, is that the same total frequency of R1b does not imply the same demographic history. Ancient DNA shows that R1b carriers in Iberia were not simply genetic equivalents of R1b carriers farther north. Once that is recognized, the Basque paradox becomes much easier to understand.

Ancient DNA resolves the puzzle

First, using the whole ancient AADR dataset restricted to males younger than 7500 BP, I tested whether R1b carriers are more Steppe-shifted than non-R1b individuals after accounting for time. After removing the linear time trend, the difference remains very large: Welch t = -19.14, p = 1.13e-74, Cohen’s d = 0.77, Hedges’ g = 0.77. In plain language, R1b does track Steppe-related ancestry. It just does not track it perfectly or in a one-to-one way.

Figure 3 shows that general association. It plots residual Steppe admixture after removing the time trend, separately for R1b and non-R1b individuals in the whole ancient dataset. The two distributions overlap, but they are clearly shifted. R1b is associated with more Steppe ancestry.

Figure 3. Residual Steppe admixture in the whole ancient AADR dataset after removing the linear time trend, shown separately for R1b and non-R1b males.

We are getting closer to the solution of the Basque paradox. The crucial question is whether the ancient R1b carriers of Iberia were themselves less Steppe-rich than the ancient R1b carriers of northern and central Europe.

Read more

In the standard archaeological and linguistic model, the story begins in the Proto-Indo-European world of the Pontic-Caspian steppe, often associated with Yamnaya-related populations. Later, from that broader steppe background, emerged eastern Bronze Age groups such as Sintashta and related populations, which are much more plausibly connected to Proto-Indo-Iranian. So the usual model is not that Yamnaya themselves were Proto-Indo-Iranian. Rather, Yamnaya-related peoples belong to an earlier Proto-Indo-European horizon, while Sintashta and related eastern steppe groups represent a later development, more directly tied to the branch that eventually gave rise to Indo-Aryan and Iranian languages.

The map below summarizes the main populations used as geographic anchors in the analysis. It is not meant as a claim about exact homelands or migration routes in any narrow sense. Rather, it situates the broad ancestral zones discussed in the text: the western steppe world associated with Yamnaya, the eastern steppe zone associated with Sintashta and related Middle to Late Bronze Age groups, the Zagros-related pre-steppe populations of Iran, and the Central Asian BMAC horizon that later became crucial in shaping Iranian ancestry.

Figure 1. Map of the main ancestral source regions discussed in the Iranian analysis

If Iranian languages ultimately derive from a Proto-Indo-Iranian world connected to later steppe populations, then one should indeed expect some steppe-related ancestry in later Iranian-speaking populations.

The steppe connection is real. But how large was it, and what exactly did it enter? That is where the nationalist simplification starts to break down. Once we look at the ancient Iranian sequence, and then at formal models of modern Iranian groups, the picture becomes much more layered than the usual “Aryan” story suggests.

That, though, is not the main question. The more interesting question is how much of Iran’s ancestry came from that steppe-linked source, and how much was already there before it arrived.

Read more

Not in the literal sense of skin colour, but in the genetic sense of which populations are the most “typical”, or the least affected by external admixture.

The intuitive way to answer this is to look at a PCA plot. Populations that cluster tightly and sit at the “core” of Europe might seem more “pure”, while those drifting toward other regions might appear more admixed.

But this approach has a flaw.

PCA plots are low-dimensional summaries of a much higher-dimensional structure. Most of the genetic variation is compressed into just two axes, and distances in that space can be misleading. Two populations that look close in a 2D plot may differ substantially once additional components are considered.

Even if we tried to formalize this approach — for example, by measuring how far individuals sit from their population centroid — we would still run into arbitrary choices: how many components to include, how to weight them, and what exactly counts as “central”.

PCA is useful for visualization, but it is not a reliable way to quantify how “pure” a population is.

A better starting point: genetic distance

A more direct approach is to use genetic distances.

Here I use FST, computed on the modern subset of the AADR dataset, to measure how different populations are from each other. For each European population, I calculate two quantities: its average genetic distance to other Europeans, and its average distance to a set of non-European outgroups (African, East Asian, South Asian, Central Asian, and Siberian).

At first glance, this seems like exactly what we need. If a population is far from all non-European groups, we might be tempted to call it “pure”.

And indeed, if you rank populations by their distance to non-Europeans, the familiar names appear at the top: Sardinians, Basques, and island populations like Orcadians.

Isolation inflates everything

Populations that are genetically isolated within Europe are also further from the rest of the world. Isolation inflates genetic distance in all directions. A group that has drifted away from other Europeans will automatically look more distant from Africans, East Asians, or South Asians, even if it has not experienced less external admixture.

This becomes clear when you plot the two quantities against each other:

There’s a very tight relationship between the two. Populations that are far from the rest of Europe are also far from non-European groups.

Sardinians, Basques, Orcadians all end up in the same corner. Not just far from the outside world, but far from other Europeans too. Mainland populations sit much closer together.

At that point the ranking stops being very informative. If a population is more isolated, it will look more distant from everything.

So just taking distance to non-Europeans and calling that “purity” doesn’t really work. It’s mixing two different things.

The question then becomes how much of that distance is just isolation, and how much reflects actual connections outside Europe.

Read more

Anthropologists have long been interested in differences in body proportions across human populations. One reason is that these differences often fit a broader zoological pattern known as Allen’s rule. In mammals and birds, populations and species from colder environments tend to have shorter extremities, while those from warmer environments tend to have longer ones. The usual explanation is thermoregulatory. Heat is lost through the body surface, whereas heat production depends more on body mass. Longer limbs increase surface area relative to volume and can facilitate heat dissipation; shorter limbs do the opposite and help conserve heat.

That makes Allen’s rule a plausible framework for thinking about human variation in limb proportions, but it is not a complete explanation. Body proportions are influenced by many factors, including nutrition, disease burden, developmental conditions, and the history of population mixture and drift. Even if climate mattered, it is not obvious whether the resulting differences are mainly genetic, mainly developmental, or some combination of the two. In animals more generally, both routes are possible: some differences reflect evolved genetic divergence, while others can arise through plastic responses during growth.

This was the unresolved issue in humans. The older anthropological literature documented recurrent population differences in relative limb length and trunk length, but it was much less able to say what produced them. Were these differences primarily developmental responses to local conditions, or did they also reflect genetic differences between populations?

Bartell et al. revisit this question with samples large enough to study body proportions at GWAS scale. Using very large samples from the UK Biobank and China Kadoorie Biobank, they study sitting height ratio and related traits at GWAS scale. The key measure is the sitting height ratio, or SHR, the fraction of total height accounted for by the torso. Higher SHR means relatively longer trunks and shorter legs; lower SHR means the reverse.

The resulting pattern is not limited to a single comparison. East Asian ancestry individuals show higher SHR than Europeans, and the broader multi-ancestry comparison places African, South Asian, European, and East Asian groups along the same general axis, with lower SHR corresponding to relatively longer limbs. In the main comparison reported in the paper, East Asian individuals have an average SHR of about 0.539, compared with 0.530 in Europeans. Taken together, the ordering is clear: African ancestry groups are lowest, followed by South Asians, then Europeans, with East Asians highest.

Figure 1. Sitting height ratio differs systematically across ancestry groups in the UK Biobank.

The ordering is consistent with the older climatic interpretation, but the main point is simpler. This is not an isolated contrast between two groups. The same broad gradient appears across multiple ancestry groups.

Earlier anthropometric work pointed in this direction, but the evidence was sparse. Here the pattern emerges clearly in very large samples. Populations differ not just in stature, but in body proportions.

Height, however, collapses multiple biological processes into a single measure. Sitting height ratio, once adjusted for height, is almost genetically independent from height itself. Two individuals can have the same genetic propensity for height and still differ in how that height is distributed between trunk and legs. The variants that make you taller are often not the same variants that determine your proportions.

The study identifies roughly five hundred loci associated with SHR. Many lie in regions already associated with height, but often involve different variants within those regions. The same locus can contain one variant that affects overall height and another that shifts proportions. In some cases, variants decrease sitting height while increasing leg length by a larger amount, showing that these are not just general growth effects but region-specific effects on the skeleton.

There are clear phenotypic differences across populations, but that does not imply different genetic systems. At the genome-wide level, the opposite is true. Genetic correlations across ancestries are close to one, and effect sizes at associated loci are very similar.

Figure 2. The genetic architecture of body proportions is largely shared across European and East Asian populations

This does not mean every locus is identical across populations. A small number of fine-mapped signals do show heterogeneity. But they are a minority. The overall architecture is overwhelmingly shared. Population differences in body proportions are therefore not explained by different sets of genes, but by differences in allele frequencies across many of the same loci.

What the paper does not show, however, is the most direct test of that interpretation. If the population differences are really genetic, then populations should differ not just in phenotype, but in their average polygenic propensity for higher sitting height ratio. I computed those scores by population using the GWAS summary statistics, and the pattern is exactly what you would expect.

Read more