A gene swept through Europe in the time since the pyramids were young
Adults are not supposed to digest milk. Lactase, the enzyme that breaks down the milk sugar lactose, switches off in most mammals after weaning, and it switched off in most humans too until quite recently. A single-nucleotide variant upstream of the lactase gene, called −13910*T in Europeans, keeps the enzyme running into adulthood, and the genomic neighborhood around that variant carries the unmistakable signature of a selective sweep: an unusually long stretch of DNA inherited as one block, with almost none of the shuffling that recombination normally produces over many generations. Bersaglieri and colleagues dated that sweep to somewhere between roughly 2,200 and 20,650 years before the present using two different methods, with a Scandinavian subsample narrowing the estimate to as recently as 1,600 to 3,200 years ago, and calculated a selection coefficient of 0.014 to 0.15 in general European samples — strong enough, they wrote, to be “comparable to that provided by resistance to malaria in malaria-endemic regions” and among the strongest yet documented for any human gene [1]. That is not a metaphor for evolution happening; it is evolution happening, on a timescale shorter than the interval between the founding of Rome and now.
The more interesting fact is that Europe did not invent this trick and did not do it alone. Tishkoff and colleagues sequenced the lactase region in East African pastoralist populations and found three separate variants — none of them the European −13910*T — each independently associated with lactase persistence and each showing its own signature of recent, strong selection: the most common African variant, C-14010, carries an estimated age of roughly 2,700 to 6,800 years and a selection coefficient of 0.035 to 0.097 depending on population and genetic model, with one Tanzanian sample estimated at s = 0.070 [2]. The European and African mutations sit in the same regulatory region, produce the same phenotype, and arose independently in lineages that had already diverged, tracking the independent domestication of dairying cattle in the Fertile Crescent and in Africa. This is convergent evolution in the textbook sense: different starting genotypes, the same environmental pressure — a new, calorie-dense, drought-resistant food source that only worked if the gut kept making lactase — and the same outcome reached twice by different genetic routes. If a reader wants one clean demonstration that human lineages still respond to selection pressures introduced by their own behavior, milk is it.
Selection has not stopped; it has become hard to see
Lactase persistence is easy to detect because the selection was strong and the marker is a single well-understood gene. Most traits are not like that, and most selection pressures on modern humans are not like the loss of an entire food source either — they are faint statistical tendencies buried in the genomes of huge cohorts, and finding them took the arrival of biobank-scale genotyping.
Beauchamp used genome-wide data on more than 8,000 individuals in the US Health and Retirement Study, born mostly between the 1920s and the 1950s, to test whether genetic variants known to predict specific traits were themselves associated with reproductive success — the working definition of ongoing natural selection. He found a statistically significant negative association between a polygenic score for educational attainment and lifetime reproductive success in both sexes: a coefficient of −0.033 in women (p = 0.002) and −0.031 in men (p = 0.013), which converts to roughly one and a half fewer months of expected schooling per generation transmitted through this channel. He found weak, not-robust evidence of selection favoring later age at menarche in women, and no significant selection on genetic scores for body-mass index, glucose, height, or schizophrenia risk [3]. Beauchamp’s own framing of the magnitude is the important part: a month and a half of schooling per generation “pales in comparison” with the roughly six-year rise in average educational attainment that occurred over the same decades through non-genetic channels — nutrition, schooling policy, economic development. Selection is real and measurable; it is also, in his words, being swamped by an “evolutionary override” from everything else that changed faster.
Kong and colleagues found the same signal from a different angle, using genealogical and genotype data on 129,808 Icelanders born between 1910 and 1990. The population-average polygenic score for educational attainment declined by about 0.010 standard units per decade across the full sample, and by about 0.0122 standard units per decade in the subset born after 1940, chosen to reduce a bias from selective survival into the genotyped sample. Because the polygenic score used captured only 3.74% of the total estimated genetic variance in educational attainment, the authors note that extrapolating to the full genetic component implies the true decline could be two to three times faster, on the order of 0.028 standard units per decade [4]. Differential fertility alone — people with lower education-associated scores having more children, weighted by generation time — accounted for an estimated 0.0104 standard units per decade, consistent with the observed total. Two independent designs, two different countries, the same direction and the same order of magnitude: a slow, real, downward drift in one specific polygenic score, running for exactly as long as the twentieth century has been raising actual educational attainment through non-genetic means.
Byars and colleagues took a third approach entirely, applying formal selection-gradient analysis to five generations of the Framingham Heart Study — a US cohort with unusually dense longitudinal medical and reproductive data — and found significant selection acting on several measured traits in women: total cholesterol under selection to fall (gradient β = −0.743, p = 0.0011, projected to decline from 223.9 to 215.9 mg/dL over ten generations), height under selection to fall slightly (β = −3.999, p = 0.0002, about 0.2 cm per generation), weight under selection to rise slightly (β = 0.861, p < 0.0001), systolic blood pressure under selection to fall (β = −0.963, p = 0.0236), age at first birth under selection to drop by about half a month per generation (β = −1.267, p < 0.0001), and age at menopause under selection to rise by about a month per generation (β = 1.280, p = 0.0035) [5]. None of these numbers describes anything a demographer or a physician would notice happening in real time. All of them describe a population still being sorted by differential reproduction, on traits ranging from blood chemistry to reproductive timing, inside a cohort living under twentieth-century American medicine rather than under any ancestral condition. The honest summary of this whole line of work is symmetrical: selection has not stopped, and its measured effect sizes are consistently tiny relative to the environmental and medical changes operating on the same traits over the same decades.
Fewer children die, so fewer genotypes are judged
Selection needs differential survival or differential reproduction to act on, and the twentieth century removed most of the differential survival component for a huge share of humanity. That is not a rhetorical claim; it has a number. Historically, across most of the world before roughly 1800, close to one in two children born died before the age of fifteen; by 1950 the global figure had fallen to roughly one in four; by 2020 it had fallen to about 4.3%, with several countries — Iceland, Japan, Norway among them — down near 0.4%, ten times below the global average [6]. Every one of those improvements — antibiotics, vaccination, sanitation, obstetric care, neonatal intensive care — worked by decoupling a genotype’s chance of surviving to reproduce from whatever it was that used to kill that genotype’s carriers before puberty. A variant that once carried a meaningful mortality penalty in childhood, from a metabolic defect or an immune weakness or a structural anomaly, now often carries none, because the child survives regardless and grows up to have children of their own.
This is precisely the mechanism behind the most-cited and most-disputed claim in this area: that relaxed selection is allowing deleterious mutations to accumulate in the human gene pool faster than they are being removed. Lynch estimated the human genomic mutation rate at roughly 50 to 100 new mutations per diploid newborn, including about 38 de novo base substitutions, and separately estimated that each newborn carries on average about 0.86 new amino-acid-altering mutations with deleterious fitness effects averaging around 4% or less each. Modeling a scenario of essentially complete relaxation of selection against these mutations in industrialized societies, he projected a fitness decline on the order of 1% to 5% per generation, and warned that over roughly six generations — about two centuries — a doubling of the mutation rate under sustained relaxed selection could compound into a 12% to 60% decline in fitness, adding explicitly that “the net fitness consequences of human mutations remain unclear” and that most of the burden sits in mutations of very small, unpredictable individual effect [7]. Read carefully, this is a conditional projection under a specific, extreme assumption — complete relaxation — stated by its own author as highly uncertain, not a measured trend.
The empirical pushback is worth stating with equal care rather than picking a side by omission. Do and colleagues tested a structurally similar question — whether a population that experienced a much stronger, historically documented relaxation of selection efficiency, non-Africans passing through the out-of-Africa population bottleneck, actually ended up carrying a detectably higher burden of deleterious mutations than Africans, who did not pass through that bottleneck. Comparing thousands of genomes, they found the ratio of deleterious mutation load between West African and European samples statistically indistinguishable from 1.0 across synonymous sites (0.982–1.022), all nonsynonymous mutations (0.994–1.010), and functionally predicted damaging variants — despite the two populations’ genuinely different demographic histories [8]. Their own simulations had predicted the true ratio should be around 0.987 under additive selection, so close to parity that it would be essentially undetectable with the sample sizes available — meaning the empirical non-result is exactly what theory predicted, not evidence that selection efficiency doesn’t matter at all. The authors are explicit that this additive-model result does not rule out effects specific to recessive or strongly epistatic disease architectures, where demographic history shapes allele-frequency distributions in ways an additive load calculation misses.
Put the two together and the sober position is neither alarm nor dismissal. Lynch’s mechanism is real: modern medicine has removed a filter that used to operate on newborns, and mutation continues to occur regardless of what filters society builds. But the best available empirical test of a comparably sized natural experiment found the resulting difference in total deleterious burden to be too small to detect against a background of ordinary genetic variation, which is consistent with theory that predicts the effect should be genuinely tiny per generation even when the relaxation itself is large and real. A change measured in fractions of a percent per generation, compounding over centuries, is a real long-run phenomenon and a non-event on any timescale a living person will observe.
Editing a person’s cells is not editing the species
Every mechanism discussed so far changes allele frequencies by changing who reproduces and how often. In November 2023, medicine acquired a genuinely different tool: the ability to rewrite a patient’s own DNA directly, without waiting for reproduction to do anything.
On 16 November 2023 the UK’s Medicines and Healthcare products Regulatory Agency authorized Casgevy (exagamglogene autotemcel), jointly developed by Vertex Pharmaceuticals and CRISPR Therapeutics, for patients aged twelve and over with sickle-cell disease or transfusion-dependent beta-thalassemia — the MHRA’s own announcement calling it “the first medicine to be licensed that uses the innovative gene-editing tool CRISPR, for which its inventors were awarded the Nobel Prize in 2020,” and describing it as an alternative to bone-marrow transplant, previously “the only permanent treatment option,” which requires a closely matched donor and carries rejection risk [9]. The European Medicines Agency granted conditional marketing authorization for the same medicine on 9 February 2024, describing the mechanism as taking blood-forming stem cells from a patient’s own blood, editing them outside the body so they produce more fetal hemoglobin — the form of hemoglobin present in babies before birth, unaffected by the mutations that cause sickle-cell disease and beta-thalassemia — and returning the edited cells to the same patient [10].
The manufacturing sequence behind that authorization is what makes the somatic boundary concrete rather than abstract. A patient’s own blood-forming stem cells are collected by apheresis; they leave the body, are edited by CRISPR-Cas9 in a sealed manufacturing suite, and are returned to the same patient after a conditioning regimen clears space in the bone marrow for the edited cells to engraft. At no point in that sequence does a germ cell — a sperm or egg precursor, or the DNA of an embryo — enter the process. The edit lives in blood-forming stem cells, which do not contribute DNA to the next generation. A patient successfully treated with Casgevy has an edited genome in their blood system and an unedited genome in their sperm or eggs; any children they have will inherit the original sickle-cell or thalassemia-causing alleles at the ordinary Mendelian rate, exactly as if the therapy had never happened. This is the single most consequential fact in this entire subject and it is also the most frequently blurred in casual discussion of “gene editing changing humanity”: as of this writing, no approved gene-editing therapy changes the genome that gets passed to a patient’s children, because none of them edit reproductive cells, and doing so intentionally would be a different, and in every jurisdiction that regulates it explicitly, currently prohibited act.
One scientist crossed a line the field had not yet drawn
That prohibition exists because someone already tried the other version, and the field’s own norms were not yet formalized enough to stop him beforehand.
In November 2018, He Jiankui, then a researcher at the Southern University of Science and Technology in Shenzhen, announced — first through reporting by MIT Technology Review’s Antonio Regalado, then in a video released the same week — that he had used CRISPR-Cas9 to edit the CCR5 gene in human embryos, attempting to confer resistance to HIV, and that twin girls had been born from the procedure. Regalado’s original report captured He’s own ambivalence in real time, quoting his acknowledgment that “we should do this slow and cautious, since a single case of failure could kill the whole field,” and a separate public statement from He that he supported “gene editing for the treatment and prevention of disease, but not for enhancement or improving I.Q.” [11]. The scientific community’s reaction to the fait accompli was close to unanimous condemnation, and the procedure preceded the Second International Summit on Human Genome Editing in Hong Kong by only days, catching the summit’s own organizers by surprise.
He was tried in China and, in December 2019, convicted of illegal medical practice under Article 336 of China’s Criminal Law and sentenced to three years in prison [12]. That case has since reshaped the surrounding legal landscape rather than settling into a historical footnote. Zou, Li, and Tao’s 2025 comparative analysis of germline-editing regulation in China, the US, and the UK states plainly that “no country currently allows human heritable genome editing” as a matter of approved clinical practice, and that China’s own post-2018 legal reforms — amendments to its Criminal Law and Civil Code, followed by 2024 ethical guidelines explicitly stating that “any clinical research on heritable genome editing is irresponsible and not allowed” — were direct responses to the case. In the US, the FDA and NIH regulate only somatic gene therapy for clinical use, while heritable human genome editing research remains legally unapprovable for clinical application; in the UK, the Human Fertilisation and Embryology Authority permits embryo research only up to fourteen days of development and “has not approved any such clinical studies” of heritable editing [12]. The World Health Organization’s 2021 governance framework, produced in direct response to the same episode, recommends a global registry of human genome-editing research, a confidential reporting channel for “illegal, unregistered, unethical or unsafe” work, and an expert committee to oversee international clinical-trial monitoring specifically for genome-editing technologies “of concern” [13]. The 2018 event did not just draw scientific condemnation; it produced a durable, still-active piece of international governance infrastructure.
The technical case against embryo editing has gotten stronger, not weaker, since 2018, because researchers went and looked for what could go wrong at the molecular level. Zuccaro, Egli, and colleagues used Cas9 to attempt to correct a disease-causing mutation in human embryos and sequenced the results carefully enough to see the repair process in real time. They found that roughly half of the Cas9-induced double-strand breaks were never repaired by the mechanism researchers had hoped for — precise homology-directed repair copying the correct sequence from the other parental chromosome — and instead persisted unrepaired through the embryo’s first cell division, at which point mitosis simply lost the damaged chromosome arm entirely, producing embryos missing large stretches of a chromosome rather than carrying the intended single-letter correction [14]. Off-target Cas9 activity produced further chromosomal losses and small insertions or deletions elsewhere in the genome. This is exactly the class of failure mode — mosaicism, off-target edits, and now large-scale chromosome loss that a standard genetic test could miss if it sampled the wrong cell — that critics raised against He’s 2018 experiment in the abstract; here it is documented directly, in a controlled experiment on embryos never intended for implantation. A germline edit that goes wrong this way does not stay contained in one patient the way a failed somatic edit does; it propagates into every cell of a person who did not consent to the intervention and, if they reproduce, into their own children as well. That asymmetry of consequence, not merely institutional caution, is the technical core of why the germline line is drawn where it is.
The quiet channel is already legal and already scaling
None of this means the boundary between “genome as inherited” and “genome as chosen” is standing still. It means the channel through which choice is actually entering the human gene pool at scale is not gene editing at all — it is embryo selection, which has been legal for decades and asks no one for a new regulatory framework because it edits nothing; it only chooses among embryos that already exist.
Preimplantation genetic testing, performed on embryos created through IVF before one is selected for transfer, has gone from a niche procedure to a default option inside a single decade. Roche, Racowsky, and Harper’s analysis of the SART CORS registry, which the authors state covers more than 95% of US clinics offering fertility services, found that PGT use rose from 13% of all assisted-reproduction cycles in 2014 (18,805 of 140,392 cycles) to 32% in 2017 (54,442 of 171,381 cycles), climbing every single year in between — 13%, 21%, 28%, 32% — a statistically monotonic increase [15]. The registry could not separate PGT for aneuploidy from PGT for single-gene disorders or structural rearrangements, but the authors note that aneuploidy screening was already understood to make up the large majority of PGT cycles by that point. This is adoption on the shape of a standard technology S-curve, not a rare specialist procedure, inside a legal framework that has not needed to change because nothing about it edits a genome — it selects among the genomic combinations meiosis and fertilization already produced.
It is worth working through what that channel can and cannot do to a population’s allele frequencies, because the arithmetic is genuinely calculable and genuinely modest. Consider a recessive disease allele at population frequency
Set
Genuine germline modification, as opposed to selection among existing embryos, already has one narrow legal foothold worth naming precisely because it complicates any claim that “no country allows heritable editing” is the whole story. The UK legalized mitochondrial donation treatment in 2015, becoming, in the regulator’s own words, “the first country in the world to regulate mitochondrial donation” [16]. The technique replaces a small quantity of a prospective mother’s own mitochondrial DNA — under 1% of a child’s total genetic material, by the regulator’s characterization — with mitochondria from a donor, to prevent the transmission of serious mitochondrial disease; because mitochondrial DNA is inherited maternally, this constitutes a genuine, if extremely limited, heritable genetic change passed on to a child and onward to that child’s own daughters. It is not CRISPR editing and it does not touch nuclear DNA, but it is a real precedent for exactly the category the WHO framework and the Zou et al. comparative analysis describe as universally prohibited: a legislature can, in fact, carve out a narrow heritable exception for a severe disease indication when the intervention is judged safe enough and the alternative is judged bad enough.
Three scenarios for 2100, each with a way to prove it wrong
Extrapolating any of the preceding threads to the end of the century requires naming, for each one, who is making the claim, what has to hold true for it to happen, what a person checking in 2100 should actually look at, and what result would prove it wrong. Three scenarios below are not mutually exclusive; elements of more than one are plausible in different countries and different disease categories simultaneously.
Scenario one: drift-dominated continuity. This is the default extrapolation of the evidence in the first half of this article, and it is my own baseline reading rather than any cited author’s explicit forecast. It holds that by 2100, human allele frequencies at most loci have continued to shift under the same small, hard-to-detect forces documented by Beauchamp, Kong, and Byars — continued mild selection on education-associated and reproductive-timing variants, continued mild relaxation of selection against deleterious variants roughly in line with the small effect sizes Lynch modeled and Do et al.'s empirical test bounded — with no jurisdiction having approved heritable genome editing for clinical use at population scale, and with embryo selection remaining the dominant deliberate channel, expanding in reach but constrained by the recessive-allele arithmetic above to modest, slow effects on disease-allele frequencies. Assumptions: no major discontinuity in the regulatory consensus documented by Zou et al., no runaway reduction in the cost or difficulty of embryo editing that changes its risk-benefit calculus, and continued reliance on IVF as a minority pathway to conception. Indicator: the WHO registry and equivalent national trial registries continuing to show zero approved heritable-editing clinical programs, alongside continued biobank-based detection of selection gradients of the same small order of magnitude documented in this article’s cohorts. Disconfirmed by: any national regulator approving a heritable genome-editing clinical trial before 2100, or a biobank study reporting a selection-gradient magnitude an order of magnitude larger than those in the 2010s-era Framingham, Icelandic, or US cohort studies cited above, which would indicate some new, much stronger selective force has emerged.
Scenario two: screening-driven allele-frequency compression at named loci. This scenario claims that for a specific, named set of severe recessive-disease loci — cystic fibrosis, spinal muscular atrophy, and comparably severe, well-characterized single-gene disorders already targeted by expanded carrier screening and PGT-M — population allele frequencies in countries with high IVF and carrier-screening uptake measurably decline relative to countries without comparable screening infrastructure, over a horizon to 2075. This is my own extrapolation of the Roche et al. adoption curve combined with the recursion above, not a claim made by Roche and colleagues themselves, who studied only historical utilization, not projected allele-frequency consequences. Assumptions: PGT and expanded carrier screening continue the S-curve trajectory documented from 2014 to 2017 rather than plateauing, IVF’s share of total conception continues rising in high-income countries, and no offsetting force — such as increased carrier fitness from better disease management extending reproductive lifespan — pushes frequencies back up. Indicator: national carrier-frequency registries or large biobank studies in high-screening countries (the UK, several Nordic countries, parts of East Asia) showing a measurable decline in named severe-recessive-allele frequencies relative to matched low-screening-uptake populations. Disconfirmed by: carrier frequencies for these named conditions remaining flat or rising in high-screening countries over multiple decades, which the arithmetic above suggests is the more likely outcome precisely because the effect is structurally small even under optimistic uptake assumptions.
Scenario three: narrow-indication heritable editing gains a second and third jurisdiction. This scenario claims that by 2100, at least two more countries follow the UK’s 2015 mitochondrial-donation precedent and legalize a narrowly defined heritable genetic intervention — plausibly mitochondrial replacement itself, spreading beyond the UK, or a nuclear-genome heritable edit restricted to a small number of well-characterized, highly penetrant, severe monogenic diseases with no other reproductive option, under the kind of registry and international-oversight regime the WHO framework already outlines. This is explicitly a scenario, not a prediction I am confident in, precisely because it depends on political and ethical judgments this article cannot forecast. Assumptions: the WHO registry framework or a successor body becomes the accepted international coordination mechanism referenced in national legislation, at least one additional country’s medical regulator judges a specific heritable nuclear edit’s risk profile acceptable for a narrow indication after further resolution of the mosaicism and off-target risks Zuccaro and colleagues documented, and no comparable scandal to the 2018 case recurs to reset the field’s caution. Indicator: any national legislature or medical regulator, beyond the UK’s mitochondrial-donation framework, formally authorizing a heritable nuclear-genome edit for clinical use, reported through the WHO registry or equivalent national registries. Disconfirmed by: 2100 arriving with the UK’s 2015 mitochondrial-donation law still standing as the sole legalized heritable-genetic-modification framework worldwide, and every national position on heritable nuclear editing unchanged from the prohibition Zou, Li, and Tao document as universal in 2025.
The first species whose selection pressures are increasingly self-authored
None of the three scenarios above requires believing that gene editing is about to remake human biology at the population level, and none of the evidence earlier in this article supports the opposite, equally common claim that medicine has switched off evolution entirely. Both claims are wrong for the same reason: they treat evolution as a single dial, when what the evidence actually shows is several distinct, independently moving mechanisms operating on wildly different scales at once. Lactase persistence swept two continents under selection coefficients strong enough to notice by eye, in populations that had no idea what a gene was. Modern biobank cohorts detect selection gradients on education, cholesterol, and reproductive timing that are real, directionally consistent across independent studies, and roughly a hundred to a thousand times smaller in effect than the environmental and behavioral changes moving the same traits over the same decades. Relaxed selection is quietly loosening the grip on deleterious mutations, at a rate theory and the best available natural experiment both suggest is too small to detect within a human lifetime. Somatic gene editing already rewrites individual genomes at scale, entirely without touching the gene pool those genomes will contribute to. Germline editing remains legally and technically walled off everywhere except one narrowly drawn UK statute governing a fraction of one organelle’s genome — a wall built, in no small part, because one scientist tested where it stood in 2018 and the field decided, after the fact, to make the answer explicit.
What makes the present moment different from every prior chapter in this same long story is not that humans are evolving faster, or that evolution is about to be switched off, or on, by decree. It is that for the first time, the mechanisms determining which genotypes get to persist include institutions that write their decisions down, publish their reasoning, and can be asked, in advance, what evidence would change their minds. Lactase persistence was never voted on. The MHRA’s approval of Casgevy was. The WHO’s registry, the UK’s Human Fertilisation and Embryology Act, and the criminal court in Shenzhen are, each in its own limited domain, doing something no prior selective force in this species’ history has done: exercising a veto over a specific genetic outcome, in public, with a paper trail. That is a narrow and unglamorous kind of novelty compared to the popular imagination of designer humans, and it is also the actual, verifiable one. Humans did not stop evolving. They started keeping minutes.
