Genetic enhancement of cognition: revision 1 to 2

Revision 1

Revision 2

categories: ["enhancement", "genetics"]

tags: ["cognition", "polygenic scores", "gwas", "enhancement", "genomics", "heritability"]

summary: "The proposal to raise human cognitive ability by selecting or editing genetic variants, constrained by the extreme polygenicity of the trait and by pleiotropy."

summary: "The proposal to raise human cognitive ability by selecting or editing genetic variants, constrained by the extreme polygenicity of the trait and by pleiotropy."

updated: "2026-07-27"

updated: "2026-10-29"

humanEvidence: "No human has been genetically enhanced for cognition; the human data are polygenic scores whose accuracy between siblings, the comparison an embryo choice actually faces, is well below their accuracy across strangers."

humanEvidence: "No human has been genetically enhanced for cognition; the human data are polygenic scores whose accuracy between siblings, the comparison an embryo choice actually faces, is well below their accuracy across strangers."

issues: ["The infobox heritability range of 50-80 per cent is given without a citation."]

---

"caption": "Proposed form of human enhancement",

"rows": [

{ "label": "Target trait", "value": "Cognitive ability" },

{ "label": "Heritability in adults", "value": "Roughly 50–80%" },

{ "label": "Heritability in adults", "value": "Roughly 50-80%" },

{ "label": "Architecture", "value": "Thousands of tiny effects" },

{ "label": "Proposed routes", "value": "Selection, editing, gene therapy" },

{ "label": "Demonstrated in humans", "value": "No" },

}

```

**Genetic enhancement of cognition** is the proposal to raise human cognitive ability by choosing or altering genetic variants, either by selecting among embryos or by editing them. Cognitive ability is substantially heritable, which is what makes the proposal seem tractable. It is also among the most polygenic traits ever mapped, which is what makes it intractable in practice. The distance between those two facts accounts for most of the disagreement about whether the idea is a near-term prospect or a category error.

**Genetic enhancement of cognition** is the proposal to raise human cognitive ability by choosing or

altering genetic variants, either by selecting among embryos or by editing them. Cognitive ability

is substantially heritable, which makes the proposal seem tractable, and it is also among the most

polygenic traits ever mapped, which makes it intractable in practice. The distance between those two

facts accounts for most of the disagreement about whether the idea is a near-term prospect or a

category error.

```keyfacts

[

{ "value": "3M+", "label": "Participants in the largest educational-attainment GWAS", "note": "Okbay et al., 2022" },

  { "value": "12–16%", "label": "Variance explained between families", "note": "substantially less within families, where selection operates" },

  { "value": "12-16%", "label": "Variance explained between families", "note": "substantially less within families, where selection operates" },

{ "value": "0", "label": "Cognitive gene therapies attempted in humans", "note": "as of 2026" }

]

```

## What the genetics shows

Twin, adoption and family studies place the heritability of adult cognitive test performance somewhere in the range of one half to four fifths, rising through childhood into adulthood. Heritability estimated from measured common variants is considerably lower, on the order of a fifth to a quarter, and the polygenic scores built from those variants explain less again.

Twin, adoption and family studies place the heritability of adult cognitive test performance

somewhere in the range of one half to four fifths, rising through childhood into adulthood.

Heritability estimated from measured common variants is considerably lower, on the order of a fifth

to a quarter, and the polygenic scores built from those variants explain less again.

The largest relevant studies target educational attainment, a proxy that is cheap to collect at biobank scale. A study of about three million individuals produced a score explaining something in the region of twelve to sixteen per cent of the variance in years of schooling in independent samples.[^okbay2022] Direct genome-wide studies of measured intelligence are smaller and identify hundreds of loci.[^savage2018] In both cases the individual variants are minuscule: the largest common-variant effects correspond to a small fraction of a point on a standardised test.

The largest relevant studies target educational attainment, a proxy that is cheap to collect at

biobank scale. A study of about three million individuals produced a score explaining something in

the region of twelve to sixteen per cent of the variance in years of schooling in independent

samples.[^okbay2022] Direct genome-wide studies of measured intelligence are smaller and identify

hundreds of loci.[^savage2018] In both cases the individual variants are minuscule, with the largest

common-variant effects corresponding to a small fraction of a point on a standardised test.

Two asymmetries follow. First, there is no small set of high-effect "intelligence variants" to install. Second, the variants of large effect that do exist are almost uniformly harmful, because mutations that disrupt neurodevelopment produce intellectual disability. Nature offers a long list of ways to break cognition by changing one gene and essentially no examples of improving it that way.

Two asymmetries follow. First, there is no small set of high-effect "intelligence variants" to

install. Second, the variants of large effect that do exist are almost uniformly harmful, because

mutations that disrupt neurodevelopment produce intellectual disability. Many single genes can break

cognition; essentially none is known to improve it.

> [!key] Polygenicity is the whole problem

> A trait built from ten thousand variants of a hundredth of a point each cannot be enhanced by a technique that acts on one locus at a time. Every proposed route has to solve the many-loci problem before anything else.

> A trait built from ten thousand variants of a hundredth of a point each cannot be enhanced by a

> technique that acts on one locus at a time. Every proposed route has to solve the many-loci

> problem before anything else.

## Proposed routes

### Embryo selection

The nearest-term route is [[polygenic-embryo-screening]], ranking embryos by a cognitive or educational-attainment score. The expected gain is set by how much genetic variation exists among one couple's embryos, how accurate the score is between siblings, and how many embryos are available. Modelling by Karavani and colleagues put the expected shift at a couple of points at most when choosing the best of ten, with wide variance around that expectation, and less again for the number of embryos a real cycle produces.[^karavani2019]

The nearest-term route is [[polygenic-embryo-screening]], ranking embryos by a cognitive or

educational-attainment score. The expected gain is set by how much genetic variation exists among

one couple's embryos, how accurate the score is between siblings, and how many embryos are

available. Modelling by Karavani and colleagues put the expected shift at a couple of points at most

when choosing the best of ten, with wide variance around that expectation and less again for the

number of embryos a real cycle produces.[^karavani2019]

Two proposals attempt to relax the binding constraint. [[in-vitro-gametogenesis]] would generate large numbers of embryos from a couple's cells, raising selection intensity. Iterated embryo selection, described by Carl Shulman and Nick Bostrom, would run several rounds of gamete derivation and selection in vitro to compound gains across simulated generations without waiting for real ones.[^shulman2014] Neither has been demonstrated in humans, and the second requires deriving gametes from embryonic stem cells at a scale nobody has approached.

Two proposals attempt to relax the binding constraint. [[in-vitro-gametogenesis]] would generate

large numbers of embryos from a couple's cells, raising selection intensity. Iterated embryo

selection, described by Carl Shulman and Nick Bostrom, would run several rounds of gamete derivation

and selection in vitro to compound gains across simulated generations without waiting for real

ones.[^shulman2014] Neither has been demonstrated in humans, and the second requires deriving

gametes from embryonic stem cells at a scale nobody has approached.

### Editing

[[germline-editing]] could in principle install alleles that neither parent carries, escaping the sibling-variance limit entirely. Three obstacles stand in front of that, and only one of them concerns cutting DNA.

[[germline-editing]] could in principle install alleles that neither parent carries, escaping the

sibling-variance limit entirely. Three obstacles stand in the way, and only one of them concerns

cutting DNA.

The first is that association studies do not hand over a list of variants to install. A genome-wide hit marks a stretch of chromosome correlated with the trait; the base actually responsible may be any of hundreds nearby. Statistical fine-mapping resolves a minority of loci to a plausible causal variant, and for a trait with thousands of contributing regions the causal set is mostly uncharacterised. Editing a marker rather than a cause changes nothing.

The first is that association studies do not hand over a list of variants to install. A genome-wide

hit marks a stretch of chromosome correlated with the trait, and the base actually responsible may

be any of hundreds nearby. Statistical fine-mapping resolves a minority of loci to a plausible

causal variant, and for a trait with thousands of contributing regions the causal set is mostly

uncharacterised. Editing a marker rather than a cause changes nothing.

The second is scale. Effects measured in hundredths of a point mean that a shift worth having requires very many substitutions in one cell. The highest multiplex edit counts on record were achieved in cultured cell lines, which can be screened, expanded and thrown away; an embryo offers a single attempt, and confirming what happened to it means consuming cells it needs.

The second is scale. Effects measured in hundredths of a point mean that a shift worth having

requires very many substitutions in one cell. The highest multiplex edit counts on record were

achieved in cultured cell lines, which can be screened, expanded and thrown away. An embryo offers a

single attempt, and confirming what happened to it means consuming cells it needs.

The third is damage. Every additional cut raises the probability of the large deletions and rearrangements documented in [[crispr-off-target-effects]], and a one-cell embryo has no spare copies. [[base-editing]] and [[prime-editing]] rewrite bases without severing both strands and are the only plausible chemistry for a high-multiplex attempt, but their behaviour across many loci at once in a human embryo has not been measured.

The third is damage. Every additional cut raises the probability of the large deletions and

rearrangements documented in [[crispr-off-target-effects]], and a one-cell embryo has no spare

copies. [[base-editing]] and [[prime-editing]] rewrite bases without severing both strands and are

the only plausible chemistry for a high-multiplex attempt, but their behaviour across many loci at

once in a human embryo has not been measured.

### Somatic intervention

Editing an adult brain is a different proposition. Most of the relevant biology acts during development, delivery to the central nervous system is the standing obstacle for [[aav-vectors]], and [[somatic-gene-therapy]] for a non-disease indication would face a regulatory bar no sponsor has tried to clear. Interventions aimed at adult cognition are, for now, pharmacological rather than genetic, and the evidence there is covered in [[nootropics]].

Editing an adult brain is a different proposition. Most of the relevant biology acts during

development, delivery to the central nervous system remains the standing obstacle for

[[aav-vectors]], and [[somatic-gene-therapy]] for a non-disease indication would face a regulatory

bar no sponsor has tried to clear. Interventions aimed at adult cognition are, for now,

pharmacological rather than genetic, and the evidence there is covered in [[nootropics]].

## Animal and biological evidence

The clearest single-gene demonstration remains a mouse. Overexpressing the NR2B subunit of the NMDA receptor in the forebrain produced animals that learned faster on several tasks.[^tang1999] The result held up, but so did a subsequent finding that the same animals showed heightened sensitivity to inflammatory pain, an early and instructive illustration of pleiotropy.

The clearest single-gene demonstration remains a mouse. Overexpressing the NR2B subunit of the NMDA

receptor in the forebrain produced animals that learned faster on several tasks.[^tang1999] The

result held up, but so did a subsequent finding that the same animals showed heightened sensitivity

to inflammatory pain, an early illustration of pleiotropy.

A second line concerns klotho, a protein better known in aging research. Human carriers of one copy of the KL-VS variant show slightly better performance on some cognitive measures, and administering klotho to aged rhesus monkeys improved memory performance in a controlled study.[^castner2023] This is a genuine cross-species signal, and it is also narrow: a single, modest, hormone-mediated effect, not a general method for moving a polygenic trait.

A second line concerns klotho, a protein better known in aging research. Human carriers of one copy

of the KL-VS variant show slightly better performance on some cognitive measures, and administering

klotho to aged rhesus monkeys improved memory performance in a controlled study.[^castner2023] This

is a genuine cross-species signal, and also a narrow one: a single modest hormone-mediated effect

rather than a general method for moving a polygenic trait.

Comparative and evolutionary arguments are sometimes offered as evidence that large gains are available, on the grounds that human cognition changed rapidly in evolutionary time. That change involved coordinated modification of developmental programmes over hundreds of thousands of years, not the substitution of a few alleles.

Comparative and evolutionary arguments are sometimes offered as evidence that large gains are

available, on the grounds that human cognition changed rapidly in evolutionary time. That change

involved coordinated modification of developmental programmes over hundreds of thousands of years,

not the substitution of a few alleles.

## Why this is harder than it sounds

**Within-family attenuation.** A score's advertised accuracy is measured across unrelated strangers, and a substantial part of what it measures is not the child's own DNA. Some of it is the parents' behaviour, which correlates with their genotype and shapes the household. Some of it is ancestry and mating pattern, which align genotype with social position. Sibling comparisons strip those out, because siblings share parents, household and ancestry; what survives is the direct effect, and only the direct effect is available to an embryo choice. For educational attainment the surviving fraction is well below the headline figure, which matters because educational attainment is where nearly all the discovery power sits.

**Within-family attenuation.** A score's advertised accuracy is measured across unrelated strangers,

and a substantial part of what it measures is not the child's own DNA. Some of it is the parents'

behaviour, which correlates with their genotype and shapes the household. Some of it is ancestry and

mating pattern, which align genotype with social position. Sibling comparisons strip those out,

because siblings share parents, household and ancestry. What survives is the direct effect, and only

the direct effect is available to an embryo choice. For educational attainment the surviving

fraction is well below the headline figure, which matters because educational attainment is where

nearly all the discovery power sits.

**Pleiotropy.** Genome-wide genetic correlations link educational attainment to a range of psychiatric outcomes, positively for some conditions and negatively for others.[^bulik2015] Selecting or editing on a cognitive score therefore moves risk for other traits in ways that are only partly characterised. The direction of some of these correlations is itself contested, since they are estimated from the same confounded population data.

**Pleiotropy.** Genome-wide genetic correlations link educational attainment to a range of

psychiatric outcomes, positively for some conditions and negatively for others.[^bulik2015]

Selecting or editing on a cognitive score therefore moves risk for other traits in ways that are

only partly characterised. The direction of some of these correlations is itself contested, since

they are estimated from the same confounded population data.

**Gene–environment interplay.** Population mean scores on cognitive tests rose substantially across the twentieth century in many countries, a change no genetic mechanism explains. Whatever produced that shift was environmental, and it was larger than anything genetic selection currently offers.

**Gene and environment together.** Population mean scores on cognitive tests rose substantially

across the twentieth century in many countries, a change no genetic mechanism explains. Whatever

produced that shift was environmental, and it was larger than anything genetic selection currently

offers.

**Measurement.** Test performance, educational attainment and cognition in any richer sense are not the same variable. Optimising the measurable proxy is not obviously optimising the thing of interest, a problem that recurs throughout [[human-enhancement]].

**Measurement.** Test performance, educational attainment and cognition in any richer sense are not

the same variable. Optimising the measurable proxy is not obviously optimising the thing of

interest, a problem that recurs throughout [[human-enhancement]].

> [!caution] Contested claims

> Claims that cognitive enhancement by embryo selection is now practical rest on population-level score accuracy applied to a within-family decision. Mainstream statistical genetics regards that substitution as invalid, and no company offering the service has published outcome data on selected children.

> Claims that cognitive enhancement by embryo selection is now practical rest on population-level

> score accuracy applied to a within-family decision. Mainstream statistical genetics regards that

> substitution as invalid, and no company offering the service has published outcome data on

> selected children.

## Ethics and politics

The arguments split along familiar lines. [[procreative-beneficence]] supplies the case that parents have reason to select for cognitive advantage as for any other expected benefit. Against it, [[disability-rights-and-enhancement]] raises the expressivist objection, and [[access-and-inequality]] raises the prospect that a cognitive advantage available only to the wealthy compounds existing stratification faster than any other enhancement would. Because cognitive advantage is substantially positional, the dynamics analysed in [[enhancement-arms-race]] apply with particular force: if the gain is competitive rather than absolute, universal adoption leaves everyone where they started while consuming real resources.

The arguments split along familiar lines. [[procreative-beneficence]] supplies the case that parents

have reason to select for cognitive advantage as for any other expected benefit. Against it,

[[disability-rights-and-enhancement]] raises the expressivist objection, and

[[access-and-inequality]] raises the prospect that a cognitive advantage available only to the

wealthy compounds existing stratification faster than any other enhancement would. Because cognitive

advantage is substantially positional, the dynamics analysed in [[enhancement-arms-race]] apply with

particular force. If the gain is competitive rather than absolute, universal adoption leaves

everyone where they started while consuming real resources.

There is also a research-conduct problem. Cognitive genomics has a history entangled with eugenics, and much of the field's caution about enhancement framing reflects that history rather than the statistics alone. Several of the companies marketing cognitive scores for embryos have been criticised by the same researchers whose summary statistics they use.

There is also a research-conduct problem. Cognitive genomics has a history entangled with eugenics,

and much of the field's caution about enhancement framing reflects that history rather than the

statistics alone. Several of the companies marketing cognitive scores for embryos have been

criticised by the same researchers whose summary statistics they use.

## Outlook

The most likely near-term development is not a working enhancement but better evidence about how little the current tools deliver. Within-family genome-wide studies at biobank scale are the technically decisive experiment: they estimate direct genetic effects, and the size of those estimates sets a ceiling on what any selection or editing scheme could achieve. If within-family effects on cognition prove as attenuated as they are for educational attainment, the enhancement case rests on multiplex editing, which is not close.

The most likely near-term development is not a working enhancement but better evidence about how

little the current tools deliver. Within-family genome-wide studies at biobank scale are the

decisive experiment, because they estimate direct genetic effects and the size of those estimates

sets a ceiling on what any selection or editing scheme could achieve. If within-family effects on

cognition prove as attenuated as they are for educational attainment, the enhancement case rests on

multiplex editing, which is not close.

The harder question is what a positive result would mean. A validated within-family score capable of shifting cognition by a few points would sit alongside interventions in nutrition, schooling and air quality that are known to move the same outcome by comparable amounts at a fraction of the cost, and would still be argued about far more.

The harder question is what a positive result would mean. A validated within-family score capable of

shifting cognition by a few points would sit alongside interventions in nutrition, schooling and air

quality that are known to move the same outcome by comparable amounts at a fraction of the cost, and

would still be argued about far more.

Two things would change that calculation. The first is a mechanistic result rather than a statistical one: a variant whose effect on neural development is understood well enough to be engineered rather than merely selected. Nothing on the current maps looks like that, and the pleiotropy problem suggests few candidates will. The second is a shift in what is being enhanced. Most of the argument assumes general cognitive ability as measured by existing tests, which is the trait with the largest and least tractable polygenic architecture. Narrower targets — working-memory capacity, resistance to age-related cognitive decline, the specific deficits of a diagnosed condition — are smaller problems and land closer to therapy than to enhancement.

Two things would change that calculation. The first is a mechanistic result rather than a

statistical one, meaning a variant whose effect on neural development is understood well enough to

be engineered rather than merely selected. Nothing on the current maps looks like that, and the

pleiotropy problem suggests few candidates will. The second is a shift in what is being enhanced.

Most of the argument assumes general cognitive ability as measured by existing tests, which is the

trait with the largest and least tractable polygenic architecture. Narrower targets, such as

working-memory capacity, resistance to age-related cognitive decline, or the specific deficits of a

diagnosed condition, are smaller problems and land closer to therapy than to enhancement.

That last point is where the debate is most likely to actually go. The line between preventing decline and raising a baseline is not sharp, and interventions developed for [[proteostasis|neurodegeneration]] will be the first to test how societies police it.

That last point is where the debate is most likely to go. The line between preventing decline and

raising a baseline is not sharp, and interventions developed for [[proteostasis|neurodegeneration]]

will be the first to test how societies police it.

## See also