Contents
58 — Ethics, privacy and society Part 12 — Applications and ethics

58 — Ethics, privacy and society

Before this: Ch 57 — Genomics in practice · Ch 53 — Polygenic scores · Ch 31 — Heritability and response to selection · Time: ~95 min (2 sessions)

This is the last chapter and it is not an appendix. Every problem in it is a real problem with real trade-offs, on which competent and decent people disagree. Where the disagreement is genuine I say so and give both sides their strongest form. Where a claim is simply wrong — and some of the loudest claims here are — I show the defect rather than the disapproval, because the defect is what you can check.

What you'll be able to do

The core idea

Every framework for handling personal data assumes three things: that data can be de-identified, that the person it describes can consent for themselves, and that consent can be withdrawn. Genomic data breaks all three, structurally.

It cannot be de-identified, because a genome is an identifier — permanent, unchangeable, shed onto every glass you touch. It cannot be individually consented to, because half of it is also your sibling's and a quarter your grandchild's, and they were not asked. It cannot be withdrawn, because your relatives' data continues to describe you after you leave.

Your genome is not personal data in the sense the law means. It is a permanent, shared identifier held jointly with everyone you are related to — most of whom you will never meet, some of whom are not yet born, and none of whom you can consent for.

Almost everything difficult below follows from that sentence. The privacy problems follow because a shared permanent identifier cannot be anonymised. The consent problems follow because the unit of decision (the individual) does not match the unit of exposure (the family, and sometimes the community). The equity problems follow because whichever populations are measured first accumulate an interpretive advantage that compounds. And the race arguments go wrong because they mistake a statistical description of continuous ancestry for a taxonomy of discrete kinds.


1. Why a genome cannot be anonymised

Take a SNP with minor q in Hardy–Weinberg proportions (Ch 26). Two unrelated people share a there with probability

P_{\text{match}} = (p^2)^2 + (2pq)^2 + (q^2)^2

At q = 0.5 that is 0.25² + 0.5² + 0.25² = 0.375. Over n independent such SNPs the match probability is 0.375ⁿ, so uniqueness among 8 billion people needs n > ln(1.25 × 10⁻¹⁰)/ln(0.375) = 23.3 — 24 SNPs. At a more typical q = 0.2 the per-SNP match probability is 0.514 and you need 35. That is where the familiar "30–80 SNPs identify a person" figure comes from, and it is worth deriving rather than quoting: the number is small because genotypes are three-valued and common variants are near-maximally informative.

A "de-identified" genotype file is therefore a file of unique keys. Stripping the name removes only the join you would have used.

Identifiability is not a property of a dataset. It is a property of the join between that dataset and the rest of the world. Which is why every mitigation aimed at the file — dropping columns, coarsening ages, k-anonymising postcodes — misses. The attacks work on the join.

AttackMechanismDemonstrated
Quasi-identifier joinGenotype + age + sex + region matched against a public recordRoutine; the same logic that re-identifies "anonymous" health records
Y- surname inferenceThe descends with the surname in patrilineal societies. Profile Y-STRs, query a recreational genealogy database keyed on surname, triangulate with age and stateGymrek et al., Science 2013 — ~12% success for US males, using only free public resources
Long-range familial searchMatch on shared IBD segments; a third cousin is enough to build a and close it with public recordsErlich et al., Science 2018 — searching a 1.28M-record consumer database, "nearly 60% of long-range familial searches return a relative with IBD segments with a total length of 100 cM or more"; the authors project that covering ~2% of a target population gives more than 99% of it a third-cousin match
Membership inference from summariesGiven a target's genotypes and a reference panel, a likelihood-ratio test over many frequencies detects whether the target was in the case cohortHomer et al., PLoS Genetics 2008 — after which NIH withdrew aggregate frequencies from open access
Beacon querying"Is allele X present in this dataset?", repeated over rare allelesShringarpure & Bustamante, AJHG 2015 — 250 queries sufficed against a 65-person beacon, ~5,000 against one holding 1,000 individuals; cost scales with beacon size, and beacons now rate-limit and budget

The Homer result is the one that surprises engineers. Summary statistics are not safe. The perturbation one person makes to an allele frequency is individually invisible and jointly diagnostic across half a million SNPs. Any release pipeline resting on "we only publish aggregates" rests on something disproved in 2008.

Let c be the fraction of a population in a searchable database and N the number of relatives close enough to yield a detectable IBD match. Assuming rough independence,

P(\text{findable}) = 1 - (1 - c)^N

The exponent does the work. Take N ≈ 45, a deliberately conservative effective count:

c1%2%5%10%
Fraction of population findable36%60%90%99%

Two honest caveats, because this model is easier to misuse than to use. The N here is an effective parameter, not a census of your relatives. You have far more third cousins than 45; N absorbs the facts that not every true relationship leaves a detectable segment and that relatives are not independently sampled into a database.

And no single N reproduces the published data. Erlich's two results pull apart: getting the observed ~60% hit rate for a 1.28M-record database — roughly 0.9% coverage of the relevant population — needs N ≈ 107, while getting the projected >99% findability at 2% coverage needs N ≈ 228. No value fits both, which tells you the functional form is wrong and not merely the parameter: relatives are not independent draws into a database, and detectability falls off with relationship. So read the table as an argument about the shape of exposure — it saturates fast, and it saturates because coverage sits in the exponent — and not as a prediction. On Erlich's own figures the exposure at 2% coverage is far higher than 60%.

A database holding one person in fifty indexes most of a population through relatives. The 2% chose to be there. Everyone else made no decision, was never asked, and cannot opt out — because the data that exposes them is not theirs. Individual consent, the mechanism the entire post-Nuremberg apparatus of research ethics is built on, does not have the right shape for this.

Forensic genetic genealogy

The 2018 identification of the Golden State Killer suspect — a crime-scene profile uploaded to a public genealogy site, worked back through cousin matches — has since resolved hundreds of cold cases and identified unidentified remains. That is a genuine good, and the argument for it is not weak.

The argument against is not privacy in the abstract. It is that the mechanism conscripts non-participants, and that the databases were built under terms that said nothing about police. When GEDmatch switched every existing user to opt out of law-enforcement matching in May 2019, only a minority opted back in — about 185,000 profiles by late that year, out of a database then holding over a million. New users have to choose at registration, with opt-in pre-selected, and 83% of them stay opted in (Guerrini et al. 2021). Same population, same question, opposite answers: what is "consented" to here is mostly a default.

Governance has partly caught up. A US Department of Justice interim policy (approved September 2019, effective 1 November 2019) restricts federally funded forensic genealogy to violent crime and unidentified remains and requires conventional leads to be exhausted first — but it is policy, not law, and binds only DOJ-funded work. Maryland's 2021 statute goes further: judicial authorisation, specified offences only, laboratory licensing, a private right of action. Most jurisdictions have neither. A 2019 Florida warrant authorising a search of an entire genealogy database over its own opt-out settings is the clearest available statement that platform privacy promises are not enforceable against a court.

What happens when the company dies

23andMe filed for Chapter 11 in March 2025 with over 15 million customers, roughly 80% of whom had consented to research use. A database like that is an asset, and bankruptcy exists to convert assets into creditor payments. After a reopened auction the buyer was TTAM Research Institute, a non-profit formed by the company's co-founder, at $305 million, closing July 2025, under binding commitments to comply with the existing privacy policies and to honour customers' right to delete their data and opt out of research in perpetuity. That was better than the alternatives on offer — and it turned on who won an auction, not on legal structure.

Two lessons. A privacy policy is a contract a company can amend and a court can transfer; consent given to entity A is not consent given to A's acquirer, but the machinery for enforcing that is thin. And a breach of genomic data is permanent: 23andMe's 2023 credential-stuffing incident exposed millions of profiles through the relative-matching feature, which was itself the amplifier. You can reissue a password. You cannot reissue a genome, and you cannot reissue your brother's.

2. What technical mitigations buy, and what they cannot

MechanismProtectsDoes not protect
Controlled access (dbGaP, EGA, trusted research environments)Who gets data, under agreement, with auditNothing mathematically — it is a contract plus a log, and fails silently if ignored
Federated analysis ("code to data")Raw never leave the custodianThe outputs — frequencies, effect sizes, model weights — which are what membership inference consumes
Differential privacyA provable bound on any one participant's influence on a releaseUtility, at genomics scale (below)
Homomorphic encryption / MPCData in transit and in use; the compute party sees no plaintextThe result. If the result is a genotype or a score from one, it identifies as before

Differential privacy is the right formalism, and its difficulty here is specific. A releases per-variant effect sizes at thresholds around 5 × 10⁻⁸ across millions of variants (Ch 51). The privacy budget splits across all of them, and the noise needed at a per-query ε small enough to compose is comparable to the signal. Worse, standard composition assumes independent queries, and genomic queries are the opposite: (Ch 29) means an adversary recovers a noised genotype by averaging over its LD neighbours. Correlation between queries helps the attacker and not the accounting. DP works for coarse releases — cohort counts, aggregated summaries, beacon responses — and poorly for the fine-grained statistics the field wants to share.

Homomorphic encryption is real, deployed for narrow tasks (secure , encrypted variant search), and improving. It also solves a different problem from the one people reach for it to solve.

There is no cryptographic fix for a result that is itself identifying. Encryption protects the pipeline; the question is what comes out of it, and whether that output, joined against the world, names someone. Governance is not a weak substitute for better crypto — for a large class of these problems it is the only layer at which the problem exists.

The design consequence is concrete: what determines privacy outcomes are product decisions — whether relative-matching exists, whether it defaults on, what a beacon answers, what an API rate-limits, whether summaries are released per-variant or binned. They matter more than the cipher suite, and the worked example at the end of this chapter walks one of them through.

3. Discrimination

Two worries: that an insurer or employer uses genetic information against you, and that fear of it stops you being tested or joining a study. The second is measurable and large even where the first is rare.

(US, 2008) bars health insurers from using genetic information in eligibility or premiums and from requiring a test (Title I), and bars employers from using it in hiring, firing or terms of employment (Title II). Its gaps are specific and not marginal:

Not coveredConsequence
Life insuranceAn insurer may ask for and underwrite on a predictive test result
Disability insuranceSame
Long-term-care insuranceSame — the most directly implicated by a late-onset neurodegenerative risk variant
Employers with fewer than 15 employeesTitle II does not apply at all
Manifested diseaseGINA protects genetic information, not people already symptomatic; that is ADA territory, with different tests
US military and some federal contextsSeparate rules

Some US states extend protection to life, disability and long-term-care cover; most do not, so the operative regime depends on where you live.

Other jurisdictions chose differently. The UK runs the Code on Genetic Testing and Insurance (2018) — an agreement between government and the Association of British Insurers, not a statute. Insurers will not require a test and will not ask for a predictive result at all, with one exception: Huntington disease, on life cover above £500,000. Canada's Genetic Non-Discrimination Act (2017) criminalises requiring a test or disclosure as a condition of a contract or service, and was upheld by the Supreme Court in 2020. Australia, judging self-regulation to have failed, announced a ban in September 2024; the enabling legislation passed Parliament in April 2026 and commences in October 2026, prohibiting the use of adverse genetic test results in life-insurance underwriting.

How much discrimination actually occurs? Documented adverse decisions are far rarer than fear of them, and much of what respondents report was triggered by family history — which no genetic-privacy statute regulates — rather than by a test result. Among people at risk for Huntington disease, on the order of 85–90% report concern and roughly 40% report an experience they classify as discrimination, concentrated in insurance and, notably, in family and social settings rather than employment.

A low observed rate is therefore not evidence that protection is unnecessary, because the low rate is partly produced by people declining to be tested: they forgo surveillance that would extend their lives, and decline research participation, degrading the evidence base for everyone.

The actuarial case, at its strongest. Voluntary insurance is priced on symmetric information. If applicants know their and insurers may not ask, high-risk people buy more cover at pooled prices and low-risk people buy less — adverse selection, which raises prices and can unravel a market. On this view a ban is not neutral fairness; it is an unstated transfer from low-risk to high-risk purchasers.

The reply, also at its strongest. Adverse selection has been measured, and the result is two things at once. In relative terms it is real and large among the few people who have actually tested: Huntington carriers are up to five times as likely as the general population to hold long-term-care cover, and Oster and colleagues (2010) argue that even a modest expansion of genetic information could threaten that market's viability — the same coverage the table above as the one most directly implicated. In aggregate market terms it stays small, and for the reason built into that sentence: carriers are a vanishing fraction of applicants, because predictive tests are uncommon and mostly weakly predictive. Against that sits the large cost of deterring clinically indicated testing — and note the loop, because the thing keeping aggregate selection small is precisely the testing suppression this section condemns. Beneath the empirics is a normative choice no dataset settles: whether insurance pools risk or sorts it. A society that thinks pooling is the point restricts underwriting on unchosen characteristics; one that thinks sorting is the point does not. Both are coherent, and the same unresolved disagreement sits under every debate about pre-existing conditions.

The consent form assumes you can describe the research at enrolment. Biobanking assumes you cannot.

ModelParticipant agrees toCost
SpecificThis study, this questionRe-consent for every new use; attrition and bias as the cohort ages; often impossible at scale
BroadA described domain of future research, governed by a named oversight bodyConsent to a process, not a project — legitimacy transfers to the governance
DynamicAn ongoing digital relationship: granular, revisable preferences, notification of new usesReal autonomy gain; real participant burden, and it selects for the engaged, skewing the cohort
Opt-out / waiverResearch proceeds unless the participant objectsDefensible for low-risk record research; corrosive for anything identifiable or sensitive

is the workhorse, and its honest description is uncomfortable: you are not consenting to a study, you are consenting to a committee. That is how consent to any long-lived institution works — but the ethical weight has moved from the form to the data access committee, and committees vary enormously in rigour and in whose interests they represent.

The right not to know is not irrationality to be talked out of. A 30-year-old declining a predictive Huntington test faces a result that is unactionable, irreversible, and interacts with insurance exactly as §3 describes; uptake of predictive HD testing has stayed low for decades.

It collides with secondary findings. Sequence an exome for one indication and you incidentally observe everything else. ACMG recommends actively examining a curated list of medically actionable genes — grown from 56 in 2013 to more than 80 across revisions — and reporting what is found, with an opt-out. The defence: declining to look at a pathogenic BRCA1 variant you have already sequenced is a strange kind of ignorance to protect. The objection: it converts a diagnostic test into an unrequested screening programme, on a committee's list, in people with no family history — where is far lower than a literature ascertained in affected families suggests.

Return of results is a versioning problem. Variants move between , likely pathogenic and benign as evidence accumulates (Ch 55), so a report is a snapshot of a moving evidence base. Whether a laboratory owes a duty to recontact years later is unsettled in most jurisdictions — and it is a systems problem before it is an ethical one: nobody built the infrastructure to push an update to a PDF issued in 2019.

Children. Professional societies default to deferring predictive testing for adult-onset conditions, to preserve the future adult's right not to know. The counter is real: parents may need the information for reproductive planning or cascade testing, and where childhood surveillance exists deferral is harmful. Newborn screening is a genuine exception — a population programme essentially without individualised consent, justified by treatability — and genomic newborn screening (BabySeq; the UK Generation Study) stresses that justification hard, because sequencing returns far more than the treatable conditions the programme rests on. Adults who cannot consent rely on proxies with a best-interests standard plus assent, and here §5 bites: a proxy can consent for a person, and nobody can consent for a lineage.

Between 1990 and 1994, researchers collected blood from about 400 members of the Havasupai Tribe for a study of the community's very high diabetes prevalence. The samples were subsequently used for research on schizophrenia, on inbreeding, and on population migration — including a thesis concluding the tribe's ancestors crossed the Bering Strait, contradicting the tribe's own account of its origins. A tribal member discovered this in 2003 at a university presentation. Litigation settled in 2010: $700,000 to 41 members, return of the samples, support for a clinic and school. Because it settled, it set no precedent.

The structure of the harm is the transferable lesson. No individual was re-identified. No individual suffered a medical consequence. The harm fell on the group — its standing, its origin narrative, its ability to control what is said about it — and was inflicted by research a conventional consent form, read literally, arguably permitted. A framework whose only unit is the individual cannot represent this kind of injury, let alone prevent it.

Henrietta Lacks makes a point about time. Cells taken during her 1951 cervical cancer treatment without consent (not then required) became HeLa, the most widely used human cell line in biology. The HeLa genome was published in 2013 and withdrawn after the family objected, followed by an NIH agreement giving family members a seat on the committee controlling access; a settlement with Thermo Fisher over commercial use was reached in August 2023. Seventy years is not an unusual interval here. Data you generate today will outlive the consent regime under which you generated it.

The constructive response has a name. FAIR — Findable, Accessible, Interoperable, Reusable — optimises data for machine reuse and is why public genomics works at all. It is silent on who benefits and who decides. CARE — Collective benefit, Authority to control, Responsibility, Ethics — articulated by the Global Indigenous Data Alliance in 2019, supplies that missing axis and is built to sit alongside FAIR, not replace it.

The tension is real and should not be smoothed over: open data is a genuine public good, and every restriction costs discoveries not made; indigenous data sovereignty is also a genuine good, grounded in a documented history where openness was the mechanism of harm. The workable syntheses are procedural — community review boards with authority to refuse, benefit-sharing negotiated before collection, and tiered access where some analyses are open and others need community approval. Do not reach for the Nagoya Protocol here, however often you see it cited in this context: its definition of a genetic resource is "any plant, animal, microbial or material of other origin", and human genetic material is explicitly outside its scope. It is the nearest existing benefit-sharing model, which is why it gets invoked, and it does not reach this case at all. H3Africa's model, African-led with access governed on the continent, is the most developed large-scale example, and it builds capacity as well as data. Shipping samples out of a community and publishing on them is the Havasupai structure with better paperwork.

6. Equity: the ancestry bias and its concrete cost

As of August 2026 the Diversity Monitor records 88% of discovery-stage GWAS participants as being of European — a group comprising about 16% of the world's population (Martin et al. 2019). Participants recorded as African ancestry account for 0.3%, with a further 2.8% recorded as African American or Afro-Caribbean, against a sub-Saharan African population that is roughly 15% of the world's.

Watch the denominator in that first figure, because it is easy to get wrong in the direction that flatters your argument. 16% is the world share of people of European ancestry, which includes the European-descended populations of the Americas and Oceania that most of the 88% was recruited from. Europe's resident population is about 9%, and quoting that against an ancestry numerator nearly doubles the apparent over-representation. The real ratio is bad enough. This is not an abstraction, and it has four mechanical consequences.

1. Variant interpretation degrades. Clinical classification leans on population : a variant common in healthy people is not causing a severe rare disease (Ch 55). That filter needs a frequency estimate in the patient's own ancestry. Manrai et al. (2016) showed that several variants reported as pathogenic for hypertrophic cardiomyopathy were benign polymorphisms common in African-ancestry populations and rare in the European controls then available; patients — all of African or unspecified ancestry — received positive reports for a disease they did not have, and simulations showed that modest numbers of African-ancestry controls would have prevented it. The failure was not in the assay or the reasoning. It was in the reference panel.

2. More variants of uncertain significance. Under-represented ancestries yield higher rates for the same test, because evidence density is lower. Say this precisely: a VUS is a statement about the evidence, not about the variant. A higher VUS rate in a group means we know less about that group, not that its genomes are more ambiguous.

3. do not transfer. PRS accuracy decays with genetic distance from the training cohort, enough that a score developed in European-ancestry samples can lose most of its predictive value in African-ancestry individuals (Ch 53). The causes are mundane:

  GWAS finds a tag SNP, not the causal variant  ────┐
  LD between tag and causal variant differs         │
    between populations (Ch 29)              ───────┼──►  the weight is wrong
  allele frequencies differ, so variance            │      in the new sample
    explained differs                        ───────┤
  effect estimates absorb stratification and        │
    indirect genetic effects (Ch 51, Ch 53)  ───────┤
  environments and their interactions differ ───────┘

failure is caused by biased sampling and by LD structure. It is not evidence that the biology differs between groups. Train a score in an African-ancestry cohort of equivalent size and it works there and transfers poorly the other way. The asymmetry is in the data, not in the people.

4. Discoveries are missed. Variants common in one population and absent in another are only findable where they are common. PCSK9 is the standard exemplar, and it is worth stating exactly, because the usual telling overstates it. The gene's link to cholesterol was found first in European families with autosomal hypercholesterolaemia, through gain-of-function (Abifadel et al., Nature Genetics 2003). What the African-ancestry participants supplied was the opposite direction — , and the outcome data that made them a drug target. In one analysis of the Atherosclerosis Risk in Communities cohort, nonsense carried by 2.6% of Black participants and essentially absent in Europeans gave a 28% LDL reduction and an 88% reduction in coronary heart disease, while the one common European loss-of-function variant (R46L, 3.2% of white participants) gave 15% and 47% (HR 0.50, 95% CI 0.32–0.79, P = 0.003; Cohen et al., NEJM 2006). A 47% reduction at P = 0.003 is not a signal a European-only design would have missed — so the honest claim is not "far later" but that a European-only design would have seen a far weaker effect and, crucially, would have lacked the near-complete-knockout carriers who made the safety and therapeutic case unambiguous.

The risk of widening rather than narrowing. Deploying PRS clinically today distributes a benefit largest for the already best-served group. Even with impeccable intent a uniform rollout amplifies an existing disparity, because the tool's accuracy is itself unequally distributed. That is a reason to sequence deployment carefully, not a reason against the technology.

What is being done. All of Us has enrolled a cohort in which about 80% of participants come from communities historically under-represented in biomedical research and 46% from under-represented racial and ethnic groups, with whole genomes for hundreds of thousands; v4 aggregates 807,162 individuals and HPRC Release 2 provides 460 from 200+ people (reference/verified-facts.md). Recruiting from historically exploited communities, though, requires §5's governance first.

The global access gap is the sharpest inequity in the field, and sickle cell disease is the cleanest case. The best-developed genome-editing therapy in existence treats it (Ch 38), at a US list price around $2.2M, requiring apheresis, myeloablative conditioning and a transplant unit. The burden is concentrated overwhelmingly in sub-Saharan Africa and India. A therapy whose delivery requirements exceed the health-system capacity of the places the disease is concentrated is not, operationally, a cure for that disease yet.

Suggested stopping point (Session 1 of 2). §§1–6 cover consent, privacy, clinical reporting, and access. §§7–end shift to ancestry, eugenics, editing, and governance — a natural break if you are reading in two sittings.

7. Race, ancestry, and what genetics does not license

describes your actual genealogical relationships — which populations your ancestors came from, in what proportions, when. It is continuous, multidimensional, estimable with quantifiable uncertainty, and it varies within an individual across the genome, because different segments have different histories.

Race is a social classification: categories whose boundaries were drawn by historical and political processes, which have changed repeatedly within living memory and differ between countries. It is real in its effects — exposure, discrimination, access to care — and those effects are frequently what a "racial difference" in a health outcome is measuring.

They correlate imperfectly. They are not interchangeable, and using one where you mean the other is the commonest technical error in this area.

The apportionment result and the objection to it

Lewontin (1972) partitioned human genetic variation: roughly 85% within populations, ~8% among populations within continental groupings, ~6% among continental groupings. Any two people from the same village differ in most of the ways any two people on Earth differ.

Edwards (2003) raised a real objection, and it should be stated at full strength because on its own terms it is correct. Single-locus apportionment discards the correlation structure across loci. Individual variants classify people poorly; many variants jointly classify them almost perfectly — not a philosophical claim but the daily experience of anyone who has run on a matrix (Ch 28) and watched continental groups separate on the first two components. That structure is real, and it is what makes ancestry inference work and stratification correction necessary.

Both results are correct; they answer different questions.

QuestionAnswerWhat it licenses
Can individuals be assigned to ancestry clusters from genotypes?Yes, accurately, given enough markersAncestry inference, stratification control, mapping
Is most variation within rather than between those clusters?Yes, ~85% withinOnly that most human genetic variation is shared. It places no bound on the size of any particular between-group trait difference — see (a)–(e) below for the arguments that do that work
Do the clusters constitute discrete natural kinds?NoNothing

The third row is where the argument turns. Unsupervised methods return clusters because clustering is what you asked for: in ADMIXTURE and relatives, K is a parameter you choose, not a quantity the data discovers. Sample three continents and you recover three clusters; sample densely along a transect and you recover a , because human variation is overwhelmingly clinal — frequencies change gradually with distance under migration and isolation by distance. The apparent discreteness of the output is a property of the sampling design, not of the species.

Hence continental labels are poor operational proxies. Admixture proportions vary enormously within any self-identified group; for most specific quantities you would want to predict, the within-label variance exceeds the between-label difference. If your model needs ancestry, estimate ancestry. It is measurable, so measure it.

What is wrong with the genetic argument for between-group differences

The argument, in the form it is always made:

  1. Groups G₁ and G₂ differ on average in trait T.
  2. T is substantially heritable within G₁ and within G₂.
  3. Therefore the difference between them is substantially genetic.

Step 3 does not follow. (a) and (b) are each independently fatal to the inference; (c), (d) and (e) close off the routes usually taken around them, and are worth grading honestly rather than piling on — (c) blocks a proposed rescue, (d) is the closest thing to a direct test and is not a clean one, (e) is a proof that something environmental can move population means by more than the gaps at issue.

(a) is a within-group statistic carrying no between-group information. This is the two-pots argument of Ch 31: genetically variable seed split between rich and poor soil gives h² ≈ 1 within each pot and a between-pot difference that is 100% environmental by construction. Heritability is computed from within-group variance; the between-group mean does not appear in its definition. No value of h², however large or well estimated, constrains the cause of a between-group difference.

(b) The groups are not randomised into their environments. In the pots, the experimenter randomised the seed. Socially defined human groups differ systematically in nutrition, schooling, income, pollution, discrimination and medical care — which is largely what the categories track. The confound is not a residual to adjust away; it is comprehensive and correlated with the grouping by construction.

(c) The measurement the argument needs cannot currently be made. "Then compare means across groups" is invalid with present methods, for reasons established elsewhere in this book rather than invented here: PRS means are shifted by residual stratification, because stratification and trait can share the same geographic structure; the LD and frequency differences of §6 systematically bias the weights; and between-family estimates carry indirect genetic effects and assortative-mating components that do not transfer (Ch 53). A between-group PRS difference is confounded with the failure itself, and no current method separates the two.

(d) The nearest thing to a direct test exists, has been run, and does not settle it. In an admixed population, individual ancestry proportions vary continuously and are measurable (Ch 28). If a group difference in T were genetic, T should track individual ancestry proportion within that population — a comparison that at least avoids contrasting two socially separated groups.

The classic studies found nothing. Scarr, Pakstis, Katz and Barker (1977), "Absence of a relationship between degree of white ancestry and intellectual skills within a black population", and Loehlin, Vandenberg and Osborne (1973) both report no association. Both, however, estimated ancestry from a handful of blood-group markers — a weak proxy for genome-wide ancestry — and both were underpowered. Replications using genome-wide ancestry estimates do report non-null slopes; they are also published almost entirely outside mainstream genetics venues, in journals whose review standards, and in some cases whose editorial history, a reader has to weigh as part of the evidence. That is information about the finding's reception, and it is not by itself a refutation.

So do not settle this by counting papers on either side. Settle it by seeing what the design can and cannot do. Ancestry proportion inside an admixed population is itself correlated with appearance, with socioeconomic position, and with how a person is treated — so objection (b) applies here too, and no version of this study separates a genetic effect from ancestry-correlated social treatment. Regressing a trait on ancestry proportion is not a randomised experiment, and it was never going to be one. The honest statement is that (d) is contested and structurally confounded, which is why nothing above depends on it: (a) and (b) carry the refutation on their own, and would carry it whichever way (d) came out.

(e) The trait moves faster than allele frequencies can. Population mean scores on cognitive tests rose substantially across the twentieth century — the Flynn effect — by margins exceeding the group gaps at issue, over intervals far too short for allele frequencies to have changed. Whatever produces those shifts is environmental, large, and demonstrably able to move population means.

Note what has been argued, and what has not. This is not "the science is silent, so choose the kinder answer", and it is not "the data have settled it". The inference is invalid on its own terms: the step from within-group heritability to a between-group cause has no support in the definition of heritability, and the groups were never randomised into their environments. Every proposed rescue — polygenic score means, admixture regression — inherits the same confound rather than escaping it. The refutation is more useful to you than the condemnation, because you can check every step of it, including the parts where the honest answer is that the evidence is contested.

Race-based versus ancestry-informed versus genotype-based medicine

The same confusion has done concrete clinical harm through a recognisable failure mode: race used inside a clinical algorithm as a proxy for an unmeasured variable.

The eGFR equations carried a coefficient raising estimated kidney function for patients identified as Black, derived from an average creatinine difference whose causes were never established. The effect was to make measured kidney disease look milder, delaying referral and transplant listing; the 2021 CKD-EPI race-free equations removed it. Spirometry carried an analogous race correction for decades; the ATS moved to race-neutral interpretation in 2023.

If the real predictor is…Measure…Example
A genotypeThe genotypeAPOL1 G1/G2 alleles and kidney disease — present in some people of West African ancestry and absent in others
Genetic ancestry (for allele-frequency priors)Estimated ancestry, continuouslyVariant-interpretation reference panels; pharmacogenomic allele frequencies
Environment, exposure or accessThe exposureAir quality, occupational history, insurance status, experienced discrimination

Race-as-proxy fails in all three, and in the third it does something worse than fail: it converts a disparity produced by unequal conditions into a fixed characteristic of the patient and then adjusts for it — embedding the disparity in the tool meant to detect it.

8. Reproductive decisions

Prenatal screening. Cell-free DNA screening detects fetal from maternal blood with high sensitivity and specificity (Ch 57). You know what happens next: at a prevalence near 1 in 1,000, a 99.7%-sensitive test with a 0.04% false-positive rate has a near 71% — nearly three positives in ten are wrong. That is a calculation you can do in your head, and it is nonetheless routinely conveyed to patients as a diagnosis rather than a , with irreversible decisions taken on the misconception. The problem is not statistical illiteracy in the abstract; it is that these tests are marketed directly and reported without the prior.

The expressivist objection, seriously. Disability-rights scholars (Asch, Parens and others) argue that a societal programme of screening for condition X sends a message about the value of existing lives with X; that prenatal testing selects against people, not conditions; and that prospective parents' picture of life with the condition comes almost entirely from clinicians describing pathology.

The standard replies: a decision about which child to bring into existence is not a judgment about any existing person; reproductive autonomy is a strong and widely shared commitment; and the conditions screened for span such a severity range that no single argument covers both a lethal neonatal condition and one compatible with a long life.

What should survive the exchange, even if you reject its conclusion, is its strongest empirical point: people living with many of these conditions report quality of life far higher than clinicians and prospective parents predict — the disability paradox is among the most replicated findings in health psychology. The information these decisions rest on is therefore biased in a known direction, and fixing that is compatible with any position on the underlying ethics. Meanwhile the aggregate outcome — in several countries, very high termination rates after a prenatal diagnosis of Down syndrome and a shrinking population living with it — is a population-level consequence no individual decision aimed at, which is exactly why individual autonomy does not fully answer the objection.

Selection for non-disease traits. Preimplantation genetic testing for polygenic risk (PGT-P) is commercially offered and, on current evidence, weak — for reasons that follow from the statistics, not from any ethical premise:

  1. The variance available is within-family. Siblings from the same two parents vary far less than the population, and you choose among a handful of viable embryos. Published estimates of expected gain are small — on the order of a few centimetres of height, or a few IQ points under favourable assumptions.
  2. The weights are estimated between families and applied within one. Population effects include stratification, indirect genetic effects and assortative-mating inflation, all of which shrink or cancel within a sibship — which is where the selection happens (Ch 53).
  3. . Selecting on one score moves every genetically correlated trait, in directions not measured for that embryo.
  4. . Every problem in §6 applies to every embryo whose ancestry differs from the training cohort's.

Professional societies have advised against clinical use. The market exists anyway, which is a governance observation rather than a scientific one.

The therapy/enhancement line is invoked constantly and does not survive contact. Deafness and achondroplasia are regarded by many within those communities as identities rather than diseases; treating short stature within the normal range is enhancement in a lab coat; vaccination is uncontroversially enhancement of the immune system. A fuzzy line does not make the ends the same — repairing a lethal metabolic defect and selecting for height are genuinely different — but it does mean the line cannot bear the weight of a policy, and any policy resting on it will be argued about at the boundary forever.

9. Heritable genome editing

Somatic editing changes the treated person and dies with them. editing changes every descendant. The reagents are identical; the difference is which cell you put them in (Ch 38).

He Jiankui's 2018 editing of CCR5 in human embryos is the case everyone cites; Chapter 38 gives the molecular detail. The governance outcome belongs here. A Shenzhen court convicted him in December 2019 of illegal medical practice — that was the charge available, because editing embryos was not itself a crime at the time — and sentenced him to three years and a ¥3M fine. China then closed the gap: the 2020 amendment to the Criminal Law makes implantation of a gene-edited human embryo an offence carrying three to seven years. Note the sequence, because it is the general pattern of §11 in miniature — the law arrived after the experiment, and was written by the country the experiment embarrassed. The ethical analysis is separable from the technical failure and sharper for it:

ObjectionWhy better technique does not answer it
No unmet needSperm washing already prevents paternal HIV transmission. With no numerator in the risk/benefit ratio, no level of precision justifies the risk
Consent is structurally impossibleThe subject cannot consent and neither can any descendant. There is no withdrawal mechanism, ever
Novel have no evidence baseThe natural Δ32 allele has population-scale data behind it; a novel in the same gene has none. A data problem, not a technique problem
and verificationThe editor acts while the zygote divides. You cells you discard and infer about cells you keep — the inference gap is intrinsic
Irreversibility across generationsAn error propagates into a lineage; somatic error is bounded by one lifespan

The strongest case in favour deserves stating: a small set of couples cannot obtain an unaffected embryo by preimplantation testing — for instance where both partners are homozygous for the same pathogenic variant. For them editing is the only route to an unaffected genetically related child. The 2020 International Commission convened by the US National Academies and the UK Royal Society examined exactly this and found the eligible group very small. The counter is that gamete donation and adoption already address it, and that a technology with unbounded downstream consequences is poorly justified by an edge case with existing alternatives. Reasonable people land differently, and more data will not resolve it.

The enhancement worry has an honest technical answer. Complex traits are polygenic with tiny per-variant effects (Ch 30). Editing for intelligence or height would need hundreds to thousands of edits, each carrying off-target and structural risk, using between-family effect estimates that shrink within one, in a background of unknown interactions. The architecture that makes embryo selection weak makes editing weaker. The realistic near-term worry is not enhancement; it is a rogue clinic repeating 2018 with one edit and a marketing claim.

Governance as of August 2026: heritable editing is prohibited by law in more than 70 countries and by the Oviedo Convention; the 2020 Commission specified narrow preconditions and a translational pathway if clinical use were ever contemplated; the WHO issued a governance framework in 2021; the Third International Summit (London, March 2023) concluded heritable editing remains unacceptable at this time. The structural gap is obvious: prohibition is national and the technology is portable. Medical tourism is the failure mode, and one prosecution in one jurisdiction — the He case above — is the whole of the deterrence record. That is not nothing, and it is not much: it worked because the work was done at home and announced to the world.

10. Gene drives, dual use, and synthesis screening

invert Mendelian transmission so a costly can spread from low frequency (Ch 38), breaking the arithmetic all of Part 5 rests on. Against roughly 600,000 malaria deaths a year, mostly children, the case for suppressing a vector species is not marginal. The governance problem is that consent here is geographic and the technology is not: a drive does not stop at a border, can introgress into hybridising species, and is not recallable. Who is the consenting party for an ecosystem, and what does one state's refusal mean when the mosquitoes do not check? The Convention on Biological Diversity requires case-by-case risk assessment with engagement of affected communities, and the leading programmes have spent years on engagement before any release. No gene-drive organism has been released into the wild as of August 2026.

Dual use has an uncomfortable structure: the norms that make this a science — publish the sequence, publish the method, make it reproducible — are the norms that make dangerous work reproducible. The canonical case remains the 2011–12 transmissible H5N1 experiments and the debate over redacting the methods, which established that no institution had a mechanism for the decision and that redaction is largely futile once the result is known to exist.

Synthesis screening is the field's clearest chokepoint control. Ordering synthetic DNA runs through a modest number of commercial providers, so screening orders against sequences of concern and verifying customers is tractable at a real bottleneck — which is why the International Gene Synthesis Consortium has operated a voluntary framework for years. US policy has been unstable: a 2023 executive order produced an OSTP Framework for Nucleic Acid Synthesis Screening in 2024; a May 2025 executive order directed it be revised or replaced within 90 days; as of early 2026 no replacement has issued, leaving providers in limbo, some applying the 2024 version and others pausing. The deeper erosion is technological: benchtop synthesisers move synthesis inside the customer's building and dissolve the chokepoint, and design tools proposing novel sequences weaken screening lists that match against known agents.

11. Governance and the pacing gap

There is no single regulator, only a patchwork with holes in predictable places.

InstrumentCoversNotable hole
FDA device pathwayTests marketed as diagnostics; a handful of authorised DTC health-risk and pharmacogenomic reportsMost DTC output, including ancestry and "wellness", sits outside it
CLIALaboratory analytical qualitySilent on whether the test means anything clinically
HIPAACovered entities: providers, plans, clearinghousesA DTC genomics company is not a covered entity. Your clinical genome is protected; the one you bought is governed by a terms-of-service document
and national equivalentsHealth insurance and employment§3's gaps
GDPRGenetic data as a special category needing a legal basisJurisdictional, with broad research exemptions
Common Rule / research ethics committeesFunded human-subjects researchPrivate research outside funded institutions; and fit poorly in a framework built on individual risk
US state genetic-privacy statutesA growing patchwork (Utah's Genetic Information Privacy Act, effective May 2021, was first; Montana, Indiana, South Dakota and others have followed through 2026)Fifty answers, none reaching a company incorporated elsewhere

The Myriad decision is the most instructive case law here. Association for Molecular Pathology v. Myriad Genetics (2013) held that isolated naturally occurring genomic DNA is a product of nature and not patent-eligible, while cDNA — which does not occur in nature — is. Exclusive rights over BRCA1/BRCA2 testing ended, prices fell sharply, multi-gene panels became legal to offer.

What did not change is the instructive part. Myriad had accumulated, under exclusivity, the largest collection of BRCA variant observations paired with clinical outcomes, and did not share it. The patent went away; the interpretive moat did not. The response was collective — Sharing Clinical Reports, then and the ClinGen expert panels (Ch 55) — rebuilding a public evidence base from laboratories willing to deposit.

In genomics the durable monopoly is not on sequence. It is on interpretation. Sequence is cheap and commoditising; a curated, longitudinal variant–outcome database is expensive, defensible and cumulative. Anyone reasoning about competition, access or public-good provision here should watch the interpretation layer, not the assay.

The pacing gap is the general condition. Law is jurisdictional, slow and reactive; sequencing capacity, editing reagents and analysis pipelines are portable, fast and available by post. Two consequences. Technical design choices are policy — defaults, retention windows, rate limits, whether relative-matching exists — and they are made years before any statute reaches them. And professional norms plus data-access governance do the work statutes do not, which makes the seriousness of an access committee not a bureaucratic detail but, for most of this data, the actual control.


Common misconceptions

What people thinkWhat's actually true
Genomic data can be de-identified by removing names~30–40 common SNPs make a person unique. Identifiability comes from the join with external data, which stripping columns does not touch
Sharing only aggregate statistics is safe frequencies leak membership (Homer 2008), which is why NIH withdrew open aggregate data
Encryption solves genomic privacyIt protects data in transit and in use. If the output is a or a score from one, the output identifies. There is no crypto fix for an identifying result
Consenting to share your genome is your decision aloneIt exposes parents, siblings and children who were not asked. Erlich et al. (2018) project that a database covering 2% of a population yields a third-cousin-or-closer match for more than 99% of it
protects you from genetic discriminationNot in life, disability or long-term-care insurance, not at employers with fewer than 15 staff, and not once disease has manifested
Race and are the same thing measured differentlyAncestry is continuous, genealogical and measurable; race is a social classification whose boundaries have repeatedly changed. They correlate imperfectly and are never interchangeable
High within groups implies group differences are geneticHeritability is computed from within-group variance and contains no between-group information (Ch 31)
Comparing across ancestries measures genetic differenceIt measures the score's own failure, confounded with residual stratification, and no current method separates them. Portability failure reflects biased training data and differing LD — train in the other population and the asymmetry reverses
A means the variant is probably harmless — or probably harmfulIt means the evidence is insufficient. VUS rates are higher in under-represented groups because we have less data
Race-adjusted clinical equations correct for biologyThey usually proxy for an unmeasured variable. The eGFR and spirometry corrections were removed (2021, 2023) because the adjustment embedded the disparity it claimed to handle
Embryo selection on polygenic scores can meaningfully raise a traitWithin-sibship variance is a fraction of population variance, weights are estimated between families, and moves everything else
editing is banned everywhere, so the issue is settledProhibited in 70+ countries — but prohibition is national and the technology is portable. Enforcement, not consensus, is the open problem
Ethics is what you do after the analysisDefaults, retention, query limits and matching features are decided at design time and are the operative policy for most of this data

Worked example: you are asked to ship a relative-matching feature

A product manager wants the feature every consumer-genomics company has: show each customer their genetic relatives in the database. You have 500,000 customers in a population of 25 million, and an array assaying 600,000 SNPs. Work out what you are actually shipping.

Step 1 — how identifying is the data? From §1, at MAF 0.2 the -match probability between two unrelated people is 0.64² + 0.32² + 0.04² = 0.4096 + 0.1024 + 0.0016 = 0.5136 per SNP. Uniqueness among 8 × 10⁹ people needs

n  >  ln(1.25e-10) / ln(0.5136)  =  (-22.80) / (-0.6663)  =  34.2   →   35 SNPs

You hold, per customer, a key with roughly 17,000 times the SNPs required to be globally unique. No transformation short of destroying the data makes it non-identifying, and hashing does not help — an adversary with a candidate genotype hashes it the same way.

Step 2 — how many non-customers does the feature expose? c = 500,000 / 25,000,000 = 2%. Using §1's conservative effective N ≈ 45 — remembering that it is an effective parameter and that Erlich's own figures for a 2%-covered population are considerably higher:

P(≥1 detectable relative in the database)
   = 1 - (1 - 0.02)^45
   = 1 - exp(45 × (-0.020203))
   = 1 - exp(-0.9091)
   = 1 - 0.4028  =  0.597     →  about 60%

Roughly 15 million people are now discoverable through your product, and that is the optimistic number. Half a million consented; the other 14.5 million were never asked and cannot opt out, because the exposing data is not theirs to withdraw. Note the shape: coverage sits in the exponent, so exposure saturates fast — doubling to 4% gives 1 − 0.96⁴⁵ = 84%.

Step 3 — what it does to breach severity. Without matching, one compromised account yields one profile. With matching, it yields that customer's relative list — names, degrees of relationship, sometimes locations. The blast radius of a stolen credential is multiplied by the match-list length, which is precisely the mechanism of the 2023 consumer-genomics breach.

Step 4 — what the mitigations actually cover.

ProposalCoversDoes not cover
Encrypt at rest and in transitStorage compromise, interceptionThe threat here — authorised queries and legitimate outputs
Default the feature offCustomers who never enable itThe exposed non-customers, whose exposure comes from relatives enabling it
Rate-limit and anomaly-detect queriesBulk enumeration, credential stuffing at scaleA targeted search for one person
Coarsen output ("3rd–5th cousin", no segment map)Precision of reconstructionExistence of the match, which is the identifying part
Law-enforcement policy requiring legal processCasual access; sets a normA warrant, which overrides your settings (§1)
Commitments drafted to survive acquisitionSome post-sale riskBankruptcy court, a hard forum for promises with nothing behind them

Step 5 — the decision. There is no purely technical answer, and that is the lesson. The residual exposure of ~15 million non-consenting people is inherent in the feature, not a bug in its implementation. What remains is a judgment about whether that exposure is worth what the feature delivers — made, in this industry, by an engineer and a product manager, with no ethics review, years before the first statute arrived.

You are the person who will be asked to build this. That is why this chapter sits in a book about algorithms and file formats. Every consequential question in it — what to retain, what to release, what defaults to set, whose reference panel to use, which ancestry label to hard-code into a schema, whether a score should be shown to a clinician at all — arrives disguised as a technical decision, and gets made by whoever is closest to the code.

Connections