Every political philosophy is a bet on it — and its width matters more than its centre
May 14, 2026
Anyone designing an institution has to make an assumption about what the people inside it are like. You cannot avoid it: a rule is a prediction about how people will respond to that rule, and predictions require a model.
Call that model $\mathcal{H}$. Not a claim about what any individual wants, but a claim about the range of things a population might turn out to want — a distribution over possible preference distributions. A designer never knows which preference distribution they will actually face, and for a long-lived institution the distribution will change several times before the institution does. $\mathcal{H}$ is what you have instead of that knowledge.
Two claims here.
First: every political philosophy is a bet on $\mathcal{H}$, and the disagreements between them are mostly disagreements about human nature rather than about ethics. Hobbes and Locke do not differ much about what is good. They differ about what people are like, and their institutional recommendations follow from that difference almost mechanically.
Second: the tradition has argued almost entirely about where $\mathcal{H}$ is centred, and the quantity that actually decides the design is how wide it is. If that is right, then the oldest question in political philosophy has been aimed at the wrong parameter — and the parameter that matters is one modern cross-cultural evidence can estimate.
1. The canon, read as a sequence of bets
Written this way the tradition becomes legible. In each case the institutional recommendation follows near-directly from the assumed $\mathcal{H}$.
Hobbes. $\mathcal{H}$ is narrow and pessimistic: humans are roughly equal in capacity, want the same scarce things, and absent a common power will predate on one another. Life is “solitary, poore, nasty, brutish, and short.” A narrow, bad $\mathcal{H}$ licenses a strong, specific prescription — absolute sovereignty. Note the structure carefully: it is confidence about human nature, not pessimism as such, that permits the strong recommendation.
Locke. Same genre, different location. Humans in a state of nature are capable of reason and bound by natural law; the problem is not predation but the absence of an impartial judge. A less pessimistic $\mathcal{H}$ yields a weaker prescription — limited government, revocable by consent.
Rousseau. Inverts the sign. Humans are good and institutions corrupt them. This is not merely an optimistic $\mathcal{H}$; it is the first serious claim that $\mathcal{H}$ is endogenous to the institutions, which §5 argues is the hardest problem in the subject.
Madison. The most sophisticated position in the canon. Federalist 51:
Ambition must be made to counteract ambition. The interest of the man must be connected with the constitutional rights of the place. It may be a reflection on human nature, that such devices should be necessary to control the abuses of government. But what is government itself, but the greatest of all reflections on human nature? If men were angels, no government would be necessary.
Read as design advice this is remarkable, though it should not be over-read. Madison’s premise — that ambition is universal — is itself a confident bet about human nature, not far from Hobbes’s, and “ambition must be made to counteract ambition” is an argument about incentives rather than literally a minimax over preference distributions.
The difference from Hobbes is nonetheless real and worth keeping. Hobbes designs an institution whose success depends on the sovereign being as modelled. Madison designs so the outcome survives whoever happens to hold office — robustness to the thing you cannot predict, arrived at from a premise this post does not need.
Kant, in Berlin’s rendering, makes the same point as a warning rather than a design principle: “Out of the crooked timber of humanity, no straight thing was ever made.”
Marx. $\mathcal{H}$ is not a fixed object at all; consciousness is determined by material conditions and the mode of production. This is the strongest version of the endogeneity claim, and it has a specific consequence that is not often noticed: if $\mathcal{H}$ is fully plastic, hedging against variation in what people want looks pointless — the arrangement will produce whatever human nature it requires. Anyone arguing for hedged, procedural design has to engage this. The reply is that plasticity does not remove uncertainty so much as relocate it: if institutions make preferences, a designer now needs the institution-to-preference map, and §5 argues that map has multiple fixed points and no reliable way to choose among them. Full plasticity is more uncertainty than a fixed human nature, not less — but it is uncertainty of a different kind, and hedging against it is a different problem from the one this post sets up.
Hayek. A claim about the epistemics of $\mathcal{H}$ rather than its content: the relevant knowledge is dispersed, largely tacit, and constitutionally unavailable to any designer. If true, $\mathcal{H}$ cannot be estimated well enough to design against, and deliberate institutional design is misconceived in favour of evolution and selection. He is the opponent this post has to take most seriously, and §6 does not dispose of him.
Rawls. Deliberately widens $\mathcal{H}$ as a procedural device. The veil of ignorance instructs you to design as though you did not know which preference distribution you would face or which position you would occupy — artificially inflating uncertainty so only robust arrangements survive. Rawls understood that width drives the design; he installed it by stipulation rather than measuring it.
Fukuyama. The Origins of Political Order is, more explicitly than anything else in the modern canon, an attempt to write $\mathcal{H}$ down from biology and derive political development from it. Two load-bearing primitives: kin selection, the disposition to favour family, which he identifies with patrimonialism; and reciprocal altruism, the disposition and capacity to cooperate with non-kin. Add norm-following and the desire for recognition, and political development becomes a story about institutions fighting a persistent attractor — repatrimonialisation, the tendency of impersonal rule-following states to decay back into favour-trading among kin and clients whenever enforcement slackens, because that is the human default.
That last point is the most useful thing in the book for institutional design, and it deserves stating in general terms: $\mathcal{H}$ has an attractor, and institutional decay is human nature reasserting itself. A mechanism that requires people not to favour their relatives is not neutral with respect to human nature. It is spending energy against a gradient, permanently. Any evaluation of institutions that ignores this will systematically overrate the ones that look good on paper and rot in practice.
2. What we actually know
$\mathcal{H}$ is not purely a matter of philosophical taste. Substantial parts of it have been measured. Roughly in order of how well established:
Reciprocal altruism and strong reciprocity. Humans cooperate with non-kin well beyond what self-interest models predict and — critically — will pay personal costs to punish defectors who have not harmed them (Fehr and Gächter; Bowles and Gintis). Altruistic punishment is what makes large-scale cooperation stable, and it is a design resource: mechanisms can recruit it.
Norm psychology. Humans acquire, internalise and enforce arbitrary local norms with unusual facility (Henrich, Boyd, Richerson). This is why institutions work at all — and it also means that much of the content of what people want is supplied by the institutional environment, which is the endogeneity problem again.
Coalitional psychology. Rapid, low-cost formation of in-group/out-group distinctions, with associated shifts in generosity and hostility. Minimal-group experiments show it takes almost nothing to trigger. Building institutions on an $\mathcal{H}$ with tribalism assumed away is the characteristic error of liberal design, and it fails suddenly rather than gradually.
Recognition and status. Ubiquitous, cross-culturally robust, and awkward for preference-satisfaction models because it is positional — not everyone can have more of it. Any objective built on individual goal-achievement has trouble here: positional goals are zero-sum by construction, so a population holding many of them has a ceiling no institution can raise.
And the finding that constrains all of the above: WEIRD sampling. The behavioural corpus was collected overwhelmingly from Western, Educated, Industrialised, Rich, Democratic subjects, who are unrepresentative outliers on many measured dimensions (Henrich, Heine and Norenzayan, 2010). Henrich et al. (2001) ran ultimatum games across fifteen small-scale societies — with dictator and public-goods games in subsets — and found offers and rejection behaviour varying enormously with local market integration. In some societies offers well above half were rejected at much the same rate as low ones, apparently because accepting a large gift incurs an unwanted reciprocal obligation. There is no universal homo economicus. There is also no universal homo reciprocans.
Ostrom belongs here as the empirical rebuttal to the Hobbesian reading. Governing the Commons documents long-lived, self-organised institutions managing shared resources without either privatisation or Leviathan, and extracts design principles from them. The result is not that people are nice; it is that the space of workable arrangements is far larger than a narrow $\mathcal{H}$ predicts.
3. Width, not centre
Now put §2 next to the actual design problem.
The cross-cultural evidence does not tell us where $\mathcal{H}$ is centred. It tells us $\mathcal{H}$ is wide — that the range of preference distributions human populations actually exhibit, across different economic and institutional arrangements, is large. Henrich’s fifteen societies are not measurement noise around a true human nature. They are draws.
There is an objection here that has to be met rather than waved at, because it turns the evidence around. Henrich’s offers tracked local market integration: the variation is largely explained by economic and institutional structure. By §5’s own argument that $\mathcal{H}$ is endogenous, the observed spread is then not irreducible width in human nature but width induced by institutions — and a narrow-but-institution-conditional $\mathcal{H}$ licenses the opposite recommendation, namely pick the institution and let the preferences follow.
My reply is only partial. A designer does not get to condition on an institution that does not yet exist and may not survive; from the design position, institution-induced variation and intrinsic variation are the same uncertainty, because which institutional future obtains is also unknown. That defends width as the decision-relevant quantity without defending it as a fact about the species. Whether it is enough depends on how much control a designer really has over the institutional trajectory, which is exactly what §5 says is unresolved.
Why width is the operative quantity is easy to see informally, though it needs a condition stated. A designer choosing between a prescriptive institution — tuned to a specific conception of what people want — and a hedged one that performs adequately across many is making a bet. The prescriptive option is better when the bet lands and much worse when it misses; the hedged option costs something in every case. The crossover is where the price of hedging equals the expected cost of guessing wrong.
The condition: this makes width decisive only when the hedged option’s cost is roughly independent of where the target sits, and the prescriptive option’s loss grows with distance from its assumed target. Both can fail. Under sharply asymmetric loss the centre matters enormously — if being wrong in one direction is catastrophic and in the other merely inconvenient, aim away from the catastrophe whatever the spread. And if every distribution in a wide range happens to favour the same institution, width is irrelevant and you should simply adopt it. So the claim is not that the centre never matters; it is that for the loss structures institutional design typically faces, width dominates.
Three consequences.
The canon was arguing about the wrong parameter. Hobbes-versus-Rousseau is a dispute about the centre. If width dominates, the dispute is close to irrelevant to the design; what matters is that neither of them knew, and both wrote as though they did. This is why Hobbes and Marx — who agree about almost nothing — both end up recommending arrangements with very little hedging in them. Confidence, not pessimism, is what licenses the strong prescription.
Madison and Rawls were right for the wrong stated reason. Both hedge against the whole range rather than betting on a location. Madison justifies it with a claim about politicians, Rawls with a claim about fairness. The justification available here is different and, I think, stronger: it is simply what you do when the width is large, whatever you happen to think about politicians or fairness.
Width is measurable, so this is falsifiable. If further cross-cultural and historical work found human preference distributions to be far more uniform than the current record suggests, the argument here would say to abandon hedging and pick a substantive target. That is the shape of a position worth holding: one that specifies what would overturn it.
4. How to get $\mathcal{H}$ wrong
Four failure modes, each with a political tradition attached.
Assuming away tribalism. The characteristic error of liberal design: model humans as individual preference-holders and treat coalitional dynamics as a distortion to be corrected rather than a permanent feature to engineer around. Institutions built this way work well until the first serious in-group/out-group activation, then fail fast.
Assuming away cooperation. The mirror error, made by homo economicus modelling and by security-first political theory. It produces institutions that crowd out the intrinsic cooperation they could have recruited — a documented effect, not a theoretical worry. Ostrom’s commons are the standing counterexample.
WEIRD extrapolation. Estimating $\mathcal{H}$ from whatever population you have access to. This is the failure mode most directly relevant to AI: preference data in RLHF pipelines comes from a narrow, non-random slice of humanity, and the resulting model of what people want is a point estimate presented as a universal. It is precisely the error of taking the width to be small because your sample is homogeneous.
Treating $\mathcal{H}$ as exogenous. The subject of §5, and the worst of the four.
5. $\mathcal{H}$ is endogenous
Rousseau’s insight and Marx’s central claim: institutions form the preferences they aggregate. $\mathcal{H}$ is not a fact about humans that a designer reads off and builds against. It is partly a product of the design.
This breaks the clean separation the whole exercise assumes. A designer wants to choose an institution that does well against $\mathcal{H}$ — but $\mathcal{H}$ depends on the institution, so this is a fixed-point problem, not an optimisation. Three consequences follow, and none of them is comfortable:
- Multiple stable configurations. More than one institution/preference pair can be self-consistent. This is the formal statement of something historians already believe: societies lock into mutually reinforcing arrangements, and both good and bad ones are stable.
- Reform converges to the wrong thing. Iteratively redesigning against the preferences you currently observe converges, at best, to a self-consistent institution — not to a good one. Reaching the latter would require reasoning about the preferences an institution will create, which no real reform process does.
- Bad institutions damage the inputs to the next round. They do not merely fail to serve preferences; they degrade the preference distribution the next design has to work with. Weimar is the canonical case; repatrimonialisation is the slow version.
And there is a genuinely nasty interaction with §3 that I do not know how to resolve. The argument for hedged, procedural design is that the width is large. But if $\mathcal{H}$ is endogenous, a sufficiently powerful institution can narrow it — produce a population that reliably wants what the institution supplies. By the argument of §3, that institution would then be justified in becoming more prescriptive. An institution that homogenises its population earns, by this criterion, the right to commit harder to a substantive target. That is a coherent description of a totalitarian equilibrium, and the argument as it stands does not rule it out.
The only defence I can see is to fix $\mathcal{H}$ at a counterfactual estimate the institution cannot influence — some notion of what people would have wanted under conditions the institution did not set. Whether such a counterfactual is definable, let alone estimable, is open, and I think it is the most important unsolved problem in this area.
6. Open problems
- Estimate the width. The cross-cultural experimental record and the historical record of preference change are jointly enough to attempt a real number, or at least a defensible bound. This is the most valuable empirical task here, because the case for procedural design turns on it.
- Estimate the price of hedging for a specific design problem. Together with (1) this converts the central claim into a calculation.
- Model positional goals. Recognition is a robust feature of $\mathcal{H}$ and is zero-sum. What happens to an objective built on individual goal-achievement when a large fraction of goals are positional? I suspect the answer is unflattering.
- Pin $\mathcal{H}$ to something institutions cannot move (§5).
- Take Hayek seriously. If the knowledge required to estimate $\mathcal{H}$ is tacit and dispersed, what is the best available estimate, and is it good enough to design against? “Institutions should be evolved rather than designed” is a hypothesis with real evidence behind it and it has not been met here.
7. Closing
The usual question is whether people are fundamentally good or fundamentally bad, and it has absorbed four centuries of argument. The more useful question — and, unlike the first, an answerable one — is how much they vary, and how much of that variation the institutions themselves produced.
That reframing changes what a designer should do. Narrow $\mathcal{H}$ and you should identify the right conception of the good and build for it; that is what Hobbes and Marx each did, from opposite premises. Wide $\mathcal{H}$ and you should build institutions that work whoever turns up, and accept the cost of doing so. The evidence we have says wide.
Further reading. Fukuyama, The Origins of Political Order (2011), for the sociobiological derivation and for repatrimonialisation. Henrich, The Secret of Our Success (2015) and The WEIRDest People in the World (2020) for norm psychology and the sampling problem; Henrich et al. (2001) for the fifteen-society experiments. Bowles and Gintis, A Cooperative Species (2011) for strong reciprocity. Ostrom, Governing the Commons (1990) for what self-organisation actually achieves. Hayek, Law, Legislation and Liberty, for the case against the whole enterprise. Madison, Federalist 51, for the best two sentences anyone has written on designing under uncertainty about human nature.
This post is part of a series on governance mechanisms. What follows from a wide $\mathcal{H}$ — what a designer should aim at — is taken up in agency. The historical case that institutions matter this much is in why care.