Why moral bio-cybernetic/bioenhancement fails at the design document, not the ethics committee.
Proposals to biologically engineer human moral dispositions – to make people more compassionate, more cooperative, less prone to defection – are usually met with ethical objections. Consent. Autonomy. Authenticity. Value lock-in. The long shadow of eugenics.
These objections are serious and several of them are quite decisive. But they are also, in a particular sense, premature. They engage the proposal as though the technical programme were ready and the only remaining question were whether we ought to run it. That framing flatters the proposal. It grants it a maturity it does not have.
There is a more revealing test, and it is the one engineers use on any large proposal before arguing about whether to fund it: **try to write the specification.**
Not a manifesto. Not a research agenda. A design document – the artefact that lets someone else build the thing, and lets a third party check whether it worked. Every field filled in, every acceptance criterion stated, every assumption made explicit enough to be falsified.
When you actually attempt this for moral bioenhancement, something instructive happens. The document does not turn out to be *controversial*. It turns out to be **blank**. And the pattern of which fields are blank is more informative than any of the ethical arguments.
—
The construct problem: you are not turning up a dial
The first field in any specification is: *what, precisely, is being modified?*
“Compassion” is a folk-psychological term. It is not a variable. Before you can build anything you must commit to a decomposition – typically something like an affective component (felt concern at another’s distress), a motivational component (disposition to act at personal cost), a cognitive component (accuracy of your model of the other’s state), a behavioural output (rate and magnitude of costly prosocial acts), and – critically – a **scope function**: to whom does it extend, and how does it decay with social distance?
That last component is where naive versions of the project die.
Human prosociality is not a scalar quantity with a gain knob attached. It is a gradient over social distance: steep and high near kin and in-group, falling off quickly with distance, and effectively flat at the level of statistical strangers. This is why we are simultaneously the species that will run into a burning building for a neighbour and the species that can read a famine death toll over breakfast.
Which means the thing that would actually change outcomes at civilisational scale is not the *amplitude* of caring. It is the **shape of the discount function**. You do not want a higher curve. You want a flatter one.
Nobody has a genetic or neural handle on the shape. And the best-known intervention that raises amplitude appears to make the shape *worse*. Oxytocin spent about a decade as the “moral molecule” before the picture complicated: alongside well-known affiliative effects, a body of work found it strengthening in-group bonding while in some paradigms increasing out-group hostility or defensive aggression. The replication record across this literature is mixed enough that no single result should be leaned on hard. But the direction of the concern is the point. Raising the gain on a parochial system plausibly yields more effective parochialism – more devoted tribalists, better at their tribalism.
There is a second constraint that most versions of the proposal omit entirely: **stability under exploitation**. Any modified disposition must be viable in a mixed population that still contains unmodified defectors. A disposition toward unconditional cooperation is not a stable strategy; it is removed from the population – economically, socially, reproductively – by the people who lack it. Making people kinder inside an unchanged incentive landscape does not produce a kinder world. It produces exploitable people.
So the real target is not “more compassion.” It is something closer to *conditional cooperation, with an unbiased scope function, and with defection-detection and sanctioning capacity fully intact.* That is a much stranger object than the one people imagine. It is also, notably, much closer to what humans already have than to what the proposal would install.
**Status of this field: unresolved – and not primarily as an empirical matter.** It is a conceptual problem that must be settled before measurement is even meaningful.
—
The measurement gate
This is the field that stops the programme, and it stops it completely.
Any intervention specification requires a primary endpoint: a measure with construct validity (it measures the target, not social desirability), test–retest reliability sufficient to detect your expected effect size, sensitivity to within-individual change over the intervention window, resistance to demand characteristics (subjects must not be able to score well by inferring what you want), and ecological validity (it must predict field behaviour, not merely laboratory behaviour).
What actually exists falls into three families, and all three have known, documented problems:
– **Self-report instruments.** Transparently gameable. Correlate substantially with how respondents wish to be seen.
– **Economic games** (dictator, ultimatum, public goods, trust). Behaviour in these correlates weakly-to-modestly with real-world prosocial behaviour. The lab-to-field transfer problem here is one of the more uncomfortable open sores in the literature.
– **Confederate-based laboratory paradigms.** Better ecological validity, but poor scalability and severe single-use problems – you cannot re-run them on the same subject.
The psychometric reliability of these instruments is nowhere near what would be required to detect the modest effect sizes any realistic intervention would produce.
The comparison that makes this vivid: a cardiovascular intervention has LDL cholesterol as a validated surrogate endpoint, blood pressure as a second, and hard endpoints – infarction, mortality – ascertained at registry scale with near-perfect reliability.
**There is no LDL of compassion. There is no mortality-equivalent hard endpoint.** There is nothing you could enter in the “primary outcome measure” field of a trial registration that a competent reviewer would not reject.
This is not a difficulty. It is a category failure. Without a validated endpoint there is no dose-finding, no efficacy claim, no safety signal, and no way to distinguish a working intervention from a broken one. You would be optimising against a function you cannot evaluate.
—
The causal chain has no established arrows
A specification requires a causal model with every link established and quantified:
`intervention → molecular change → circuit change → systems-level change → psychological change → behavioural change`
What exists is a correlational sketch of one link. The empathy/compassion dissociation work – Singer, Klimecki and colleagues – implicates anterior insula and anterior cingulate cortex in empathic distress, and medial orbitofrontal cortex, ventral striatum and affiliation-associated regions in compassion. This is genuinely interesting, and the finding that compassion training increases positive affect while empathy training increases distress and burnout is one of the more useful results in the area.
But it should be read as suggestive, not as a wiring diagram. It is correlational. It is spatially coarse – a functional imaging voxel contains on the order of a million neurons. Sample sizes are typically small, and this subfield has documented reproducibility problems for precisely this class of finding.
What is missing is any **causal** manipulation that reliably, durably, and selectively increases the construct. Oxytocin was the strongest candidate and its literature partially collapsed under replication pressure; even the question of whether intranasal administration achieves meaningful central nervous system delivery remains contested. Contemplative training produces real effects, but modest ones, requiring ongoing practice – a behavioural intervention, not a lever a biological one could be built on.
—
Genetic architecture: no editable targets
For any germline proposal, the specification requires target loci with established causal effect, characterised effect sizes, a complete pleiotropy map, and characterised epistasis and gene–environment interaction.
For prosociality-adjacent traits – agreeableness, self-reported empathy – SNP-based heritability is modest and polygenic scores explain a low single-digit percentage of variance in independent samples. (Treat specific figures as approximate and check current sources; the direction is not in doubt.) The architecture is massively polygenic – thousands of variants of individually negligible effect – heavily pleiotropic, and poorly transferable across ancestries and environments.
Then there is a recursion problem that is rarely acknowledged: **a genome-wide association study is only as good as its phenotype.** Run against the invalid instruments described above, what you recover is the genetic architecture of *scoring highly on a questionnaire*. That is not the target. It may not even be adjacent to the target.
Multiplex editing at the scale of thousands of loci, with uncharacterised epistasis and an unmapped pleiotropy burden, is not a hard engineering problem awaiting effort. It sits outside the space of things currently attemptable.
—
The safety instrument is inside the system it monitors
This is the field I find genuinely novel, and it has no analogue in ordinary medicine.
Post-market drug safety rests on adverse event reporting. Patients notice something has gone wrong and report it. The system assumes the patient’s evaluative faculty is intact and independent of the intervention.
For a values-modifying intervention, that assumption fails by construction. The adverse event class *includes changes to the faculty that generates the report*. If the intervention shifts what a person values, then self-report is compromised as a safety instrument in exactly the failure mode you most need to detect. A population successfully modified toward a particular specification of compassion may no longer contain anyone disposed to recognise the modification as a harm.
You would therefore need an external, non-self-report harm criterion, specified in advance, held by someone outside the modified population. Nobody has (yet…) proposed a workable one.
Note that this is not a philosophical objection dressed up as an engineering one. It is a missing section in the safety file. And it generalises: irreversibility is not merely one cost to be weighed against others, because it removes the mechanism by which anything gets weighed later. Ordinary bad policy is reversed because those harmed by it object. This is the one class of intervention that can eliminate the constituency capable of identifying the error.
—
What the blanks tell us:
Lay the fields out and the completion state is stark. Delivery technology has partial content and active research behind it. Nearly everything else is empty – and the two most upstream fields, construct definition and outcome measurement, are empty in ways that no amount of funding or intelligence resolves from a single location. They are filled by cohorts, instruments, longitudinal data, and decades.
Two things follow.
**First: the ethical objections and the technical emptiness point the same way.** This is worth noticing rather than treating as coincidence. The consent problem, the value lock-in problem, and the pharmacovigilance problem are the same structural fact appearing in three registers – an intervention that alters the evaluator cannot be evaluated by the altered. That the technical specification is blank at precisely the points where the ethics is most troubling is not an accident. It reflects that we do not understand the object well enough to specify it *or* to consent to it.
**Second: the causal premise is probably wrong anyway.** The proposal assumes that destructive collective behaviour is primarily a psychological trait being expressed. But humans are already extraordinarily cooperative by primate standards – we punish unfairness at cost to ourselves, we cooperate with strangers we will never meet again. What competitive systems do is *select* for defection at the level of firms and institutions, largely independent of the dispositions of the people inside them.
The evidence for this is not subtle. When emergency conditions suspend normal procurement controls – competitive tender, due diligence, published contracts, audit trails – fraud losses jump by orders of magnitude. Same population, same dispositions, different controls. Removing the checking is what changes the behaviour.
None of which means dispositions are irrelevant. Some people are cruel, some enjoy it, and the variance is real. But what institutions and norms determine is how much *scope* those dispositions get – whether cruelty is costly or licensed, marginal or ambient. Both halves are true, and the tractable half is the second one. Ostrom’s work on commons governance showed groups solving defection problems through monitoring, graduated sanctions, and local rule-making, with nobody’s psyche altered at all.
And where genuinely catastrophic risk is the concern, it concentrates in a very small number of people with access to weapons systems, engineered pathogens, or critical infrastructure. Screening and constraining that population is orders of magnitude more tractable than modifying a species. It has real problems – who screens the screeners, capture risk – but they are the ordinary problems of institutional design rather than the irreversible rewriting of a lineage.
—
Where the real problem is
If you take the specification exercise seriously, the interesting frontier turns out not to be where the proposal points.
The measurement field is a live, unsolved, genuinely deep problem: how to construct a valid and reliable instrument for a latent construct that resists direct observation, where the act of measurement perturbs the thing measured and the subject has incentive to game the readout.
That is a problem in **measurement theory** more than in biology. Psychology has been notably bad at it – partly because the field’s training does not emphasise what a physicist’s or metrologist’s does: error propagation, calibration, sensitivity limits, distinguishing signal from instrument artefact, and knowing when your resolution cannot support your claim.
Solving it would be valuable regardless of what anyone concluded about enhancement. It is upstream of clinical trials in psychiatry, of policy evaluation, of most of behavioural science. It is where someone with quantitative training would have a genuine edge.
The specification exercise is not, in the end, an argument for despair about the underlying goal. It is a redirection. The document is blank at the top, and the top is where the work is.
So let that work begin.
*Further reading and citations: Persson & Savulescu, ***Unfit for the Future*** (the strongest case for the affirmative); John Harris’s reply on the freedom to fall; Paul Bloom, ***Against Empathy***; Elinor Ostrom, ***Governing the Commons***; Singer & Klimecki on the empathy/compassion dissociation; Habermas, ***The Future of Human Nature***.*
Leave a Reply