[Ultimately_Untrue_Thought.git] / content / drafts / consilient-cultural-worldbuilding-and-the-incoherence-of-nondiscrimination.md

Title: Consilient Cultural Worldbuilding and the Incoherence of Nondiscrimination: Comment on <em>Planecrash</em>: "Crisis of Faith" Pages 33–34
Date: 2021-01-01
Category: commentary
Tags: Eliezer Yudkowsky, worldbuilding
Status: draft

Realistic worldbuilding is a difficult art: unable to model what someone else would do except by the ["empathic inference"](https://www.lesswrong.com/posts/9fpWoXpNv83BAHJdc/the-comedy-of-behaviorism) of imagining oneself in that position, authors tend to embarrass themselves writing [alleged aliens or AIs that _just happen_ act like humans](https://www.lesswrong.com/posts/Zkzzjg3h7hW5Z36hK/humans-in-funny-suits), or allegedly foreign cultures that _just happen_ to share all of the idiosyncratic taboos of the author's own culture. The manifestations of this can be very subtle, even to authors who know about the trap.

[TODO: minimal Planecrash summary]

[TODO: summarize Keltham's interview]

... and that's the part that broke my suspension of disbelief in Keltham being a realistic portrayal of someone who grew up in dath ilan as it has been described to us, rather than being written by people who live in Berkeley in the current year who don't know how to think outside of their own culture's assumptions.

It makes sense that Keltham feels bad for the women of Orision, who seem so much less self-actualized than the women of his world. It makes sense that he wants to smash the patriarchy, and reform their sexist customs about education and property.

But the _specific_ way in which he's formulating the problem—that the law should be ["_the same for men and women_, and halflings and tieflings and elves too"](https://www.glowfic.com/replies/1817402#reply-1817402)—seems distinctively American. The idea the government can't discriminate by race or sex as a _principle_ (as contrasted to most laws happening to not refer to race or sex because those categories happen to not be relevant to that specific law) is a specific form of Earth-craziness that only makes sense as a reaction to other Earth-craziness; it's not something you would ever spontaneously invent or think was a good idea if you _actually_ came a 140 IQ Society that thoroughly educated everyone in probability theory as normative reasoning. Let me explain the real principles first, and then dissect the example of military conscription (which is plausibly _the worst_ possible example the authors could have gone with).

The issue is that probability theory doesn't have any built-in concept of "protected classes." On pain of paradox, Bayesians _must_ condition on all available information. If groups differ in decision-relevant traits, _of course_ you should treat members of those groups differently! What we call "discrimination" in America on Earth is actually just Bayesian reasoning; P(H|E) = P(E|H)P(H)/P(E) doesn't _stop being true_ when H happens to be "I should hire this candidate" and E happens to be "The candidate is a halfling". Furthermore, there's no reason for the law to behave differently in this respect than a private individual: is Governance supposed to be _less_ Bayesian _because it's Governance_?!

(As an aside, it's actually kind of _hilarious_ how far Yudkowsky's "rationalist" movement has succeeded at winning status and mindshare in a Society whose [_de facto_ state religion](https://slatestarcodex.com/2019/07/08/gay-rites-are-civil-rites/) is [founded on eliminating "discrimination."](https://richardhanania.substack.com/p/woke-institutions-is-just-civil-rights) Did—did anyone besides me "get the joke"? I would have expected _Yudkowsky_ to get the joke, but I guess not??)

Of course, as Keltham correctly points out, if you have more specific information about an individual that [screens off](https://www.lesswrong.com/posts/5yFRd3cjLpm3Nd6Di/argument-screens-off-authority) information from their demographic category, then you should use the more specific information: once you measure someone's height, the fact that men are taller than women on average with an effect size of about 1.5 standard deviations is no longer relevant to the question of that person's height. (As the saying goes out of dath ilan, [hug the query](https://www.lesswrong.com/posts/2jp98zdLo898qExrr/hug-the-query)!) In very many situations, if there's a cost associated with acquiring more specific individuating information that renders information from demographic group-membership irrelevant, you should pay that cost in order to get the more specific information and therefore make better decisions.

But crucially, getting individuating information is an [instrumental rather than a terminal value](https://www.lesswrong.com/posts/n5ucT5ZbPdhfGNLtP/terminal-values-and-instrumental-values); you should do it _when and because_ it improves your decisions, not because of some alleged principle that you're _not allowed to notice_ someone's race or sex. If there's a _cost_ associated with taking individual measurements, and the cost exceeds the amount you would save by making better decisions, then you shouldn't take the measurements. If your measurements have _error_, then your estimate of the true value of the trait being measured [regresses to the group mean](https://en.wikipedia.org/wiki/Regression_toward_the_mean) to some quantitative exent. Again, all this just falls out of _ordinary_ Bayesian decision theory, which continues to work even when some of the hypotheses are about groups of people.

If this still seems counterintuitive, it may help to consider that from the standpoint of Just Doing Bayesian Decision Theory, the distinction between "information from demographic group membership" and "information from individual measurements" isn't fundamental. The reason it seems unjust to notice race when you can just look at an individual's Strength, Intelligence, Wisdom, and Charisma scores, is because the relationship between race and any actual decision you might care about is merely statistical: it's not fair to always look to the orc if you need someone in your party to lift a fallen tree, just because orcs are stronger than other races _on average_, because it could easily be the case that this _particular_ orc is less suited to the task than other party members.

But the relationship between "measured traits" and any actual decision you might care about _is also merely statistical_. The reason we have a concept of "Intelligence" is because it turns out that people's performances on various mental tasks happen to positively correlate with each other, but that's just _on average_: it could easily be the case that this particular Intelligence 18 person is less suited to a particular task than some Intelligence 14 person. _Mathematically_, it's the same issue.

We don't typically _think_ of it as the same issue here in America on Earth. People do sometimes complain about inappropriate reliance on faulty "individual trait" proxies: that [holding a college degree isn't the same thing as being educated](/2022/Apr/student-dysphoria-and-a-previous-lifes-war/), that job interviews aren't the same thing as job performance, that IQ is not intelligence. But the objection doesn't pack the same moral force in our culture, as can be seen by how often complaints about "individual" proxies are _justified in terms of_ their effects on demographic groups, as when it is argued that ["whiteboard" coding tests are bad for diversity](https://shecancode.io/blog/its-time-to-end-whiteboard-interviews-for-software-engineers), or that [IQ is racist](https://www.nea.org/advocating-for-change/new-from-nea/racist-beginnings-standardized-testing).

The explanation for the difference in intuitions is as much political as it is moral. On account of being visible clusters in a ["thick" subspace of configuration space]((https://www.lesswrong.com/posts/esRZaPXSHgWzyB2NL/where-to-draw-the-boundaries)) (having many different correlates, [even if the effect size along any one dimension may not be very large](https://www.lesswrong.com/posts/cu7YY7WdgJBs3DpmJ/the-univariate-fallacy)), race and sex are _salient_ as [markers for coordination](/2020/Jan/book-review-the-origins-of-unfairness/). Groupings made on the basis of less visible and lower-dimensional traits, like "People with Intelligence 14", don't form a natural "interest group" in the same way, even if the lower-dimensional trait is more decision-relevant in many contexts. Conflict between interest groups in a democratic Society like America creates memetic selection pressure for "equality" memes that deny the existence of non-superficial group differences, as [the natural Schelling point for preventing group conflicts](/2020/Apr/book-review-human-diversity/#schelling-point-for-preventing-group-conflicts). It's an idea born of distrust in reasoning in an adversarial environment: if you let people _make probabilistic inferences_ using race or sex as inputs, they might motivatedly try to add _bad_ inferences to Society's shared maps that would give their own demographic an advantage in conflicts. It's safer to nip such Shenanigans in the bud by disallowing the whole class of thought to begin with: can't oppress people on the basis of race if race _doesn't exist!_

But Keltham isn't _from_ America; you'd expect his thoughts to optimized for _solving problems_, not disallowing Shenanigans. Everything we've been told about dath ilan emphasizes that they should be collectively smart enough not to fall into this _crazy_ trap of political incentives making a certain class of correct Bayesian updates socially taboo in order to avert other social ills; the Keepers should have pre-emptively done the analysis in the preceding paragraph _without_ having to empirically see it eat their Society's sanity, and incorporated the appropriate counter-memes in their rationality training for children. To the dath ilani intuition, then, the quantitative extent to which the statement "It's wrong to make _X_ decision about someone just because they're _Y_" makes sense, depends quantitatively on how strongly _Y_ predicts the outcomes of _X_. Whether _Y_ is an "individual trait" like having Intelligence 18 or a demographic category like being female _does not matter_.

This is also how American people's intuitions work, too, in contexts where their [paranoid egalitarian meliorist](TODO: linky) memetic antibodies haven't been activated. Consider how the text of _Planecrash_ itself repeatedly contrasts Keltham to everyone else in the world of Golarion. No one (neither Watsonianly in the text, nor Doylistically in various discussions of the text on Discord) is shy about saying that Keltham is special in this setting _because he's dath ilani_. We don't insist on talking about how Keltham is smart _and_ knows about probability theory _and_ knows about chemistry _and_ doesn't know about Golarionian theology _and_ is accustomed to a high material standard of living _and_ is squeamish about seeing slave markets, as if these were separate, isolated facts about Keltham as an idiosyncratic individual. We connect these facts to Keltham's nationality even though, if you look, there are surely _also_ natives of Golarion who are smart (to some quantitative extent) and know about chemistry (to some quantitative extent) and disapprove of slavery (to some quantitative extent), because our whole high-dimensional picture of what Keltham _is_—comprising many, many traits to their respective quantitative extents—is, in fact, [_causally downstream_ of the "essential" fact](https://www.lesswrong.com/posts/vhp2sW6iBhNJwqcwP/blood-is-thicker-than-water) of his having grown up in another world. It's either not bigoted to _notice_, or a cognitive system requires some amount of "bigotry" in order to function.

However, just because noticing group differences is theoretically sound, doesn't mean it's always the right thing to focus on. Pragmatically, might it not be the case in practice, that statistical group differences are small enough, and that individual trait measurements are cheap and reliable enough, such that "don't discriminate by race or sex" is a useful _heuristic_?

It's an empirical issue—but sure, very often, yes. For most jobs—especially most jobs in an industrialized Society like dath ilan—"always test the individual's aptitude, never use sex as a proxy" is a fine rule, because most jobs primarily rely on human general intelligence: there was no _dentistry_ in the environment of evolutionary adaptedness, and thus there's no reason why women or men should make better dentists. In domains where sex differences are small, using sex as a proxy would just be _dumb_, not _unjust_.

But then it's _bizarre_ that Keltham persists in his no-legal-sex-discrimination stance when his interlocutor brings up _military conscription_ as a potential counterexample. Because, well, as unpleasant as it is for modern folk to think about ... there _was_ war in the environment of evolutionary adaptedness. Men's bodies are built for war. [Men's _emotions_ are built for war.](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3260849/) [(Males have more reproductive fitness to gain and less to lose by the prospect of risking death in a war where the victors gain mating opportunities.)](https://www.cep.ucsb.edu/papers/EvolutionofWar.pdf) The sex difference in muscle mass is [_2.6 standard deviations_](/papers/janssen_et_al-skeletal_muscle_mass_and_distribution.pdf). That means a woman as strong as the average man is at _the 99.5th percentile_ for women. That means if you just select everyone whose strength is greater than one standard deviation _below_ the male mean, you end up excluding _94.5%_ of women.

Notwithstanding that Keltham grew up in a peaceful Society that [screened off its history](https://www.glowfic.com/replies/1612939#reply-1612939) (such that he wouldn't have read histories of some analogue of Genghis Khan), it seems like Keltham should know this stuff? We're told that dath ilan [has very advanced evolutionary psychology](https://www.glowfic.com/replies/1801140#reply-1801140), and there's no apparent reason for them to have spent any of their eugenics bandwidth selecting for reduced sexual dimorphism (which is [slower to evolve than monomorphic traits, anyway](/papers/rogers-mukherjee-quantitative_genetics_of_sexual_dimorphism.pdf)). We're told that [ordinary dath ilani are good at reasoning about effect sizes](https://www.glowfic.com/replies/1783037#reply-1783037).

But if Keltham _does_ know this stuff, why is he talking like a UC Berkeley graduate? ["Strength is an _externally visible and measurable_ quality that determines who you want in your army; you don't need to go by the presence of penises,"](https://www.glowfic.com/replies/1817422#reply-1817422) he says. When his interlocutor objects that strong women would get drafted, which would be terrible, Keltham asks how it would be _more_ terrible than men getting drafted. When the interlocutor replies that the woman's marriage prospects would be damaged by a history living among men in the army, Keltham muses that it sounds like she's implying that ["the army would need strong enough internal governance to prevent women in it from being raped, but you could do that with cheaper truthspells?"](https://www.glowfic.com/replies/1817432#reply-1817432)

There's just _so much_ wrong with this entire exchange from the perspective of anyone who knows anything about humans and isn't playing dumb for a religious American audience.

Firstly, if you decided that strength is the quality that determines who you want in your army, you're going to be drafting almost all men _anyway_. (Again, a sex difference of _2.6 standard deviations_ and a selection threshold 1 standard deviation below the male mean gives you a male:female ratio of (1 − Φ(−1))/(1 − Φ(1.6)) ≈ 15.4:1, where Φ is the [cumulative distribution function](https://en.wikipedia.org/wiki/Cumulative_distribution_function) of the normal distribution.)

To this, the Berkeley graduate might reply, "So then the optimal army has 15 men for every woman; what's the problem with that? Surely you don't want to make your army _less strong_ just to satisfy some weird æsthetic that all your soldiers should have the same kind of genitals?"

A minor counterreply would be that, if people's sex is public information but there are administrative costs associated with strength-testing everyone, you probably wouldn't _bother_ testing the women, for the same reason that, if you were mining for spellsilver ore, and one mine had fifteen times as much ore as the other, you wouldn't even set up your tools at the poorer mine until you had completely exhausted the first.

But more fundamentally, even if you assume strength-testing is free, we haven't yet taken into account all _other_ sex differences that are relevant to military performance. It's not just that any other individual traits (_e.g._, aggression) that you select for will stack multiplicatively, resulting in even more extreme ratios. There are group-level effects that aren't captured by measuring the traits of individual soldiers: the social dynamics of a squad of fifteen men and one woman are going to be different from those of a squad of sixteen men. Even if you've selected a woman for strength and every martial virtue to equal any man, do the _men_ know that in their hindbrains, or are they going to be biased to want to protect her or seek her favor in a way that they wouldn't in an all-male environment? You could command them not to—but does that actually _work_?

[TODO: and who is this helping, exactly? Is it worse for a woman to be drafted? Yes! (Quote what Thellim says her world believes about male nature.) Keltham acknowledges the possibility of rape, and then schemes about trying to solve it with truthspells—but why are you even trying to solve this problem at all?]

[TODO: okay, we want to accomodate exceptions; that's important. (We also want to accomodate exceptions like people without college degrees: college _or_ have an awesome portfolio is fine.) If there are women who really want to fight to defend their homeland, then either induct them or set up a special women's company depending on the empirical social design trade-offs (lower cohesion _vs._ lost skills due to no cross-sex mentorship). But "draft men, but accept women volunteers" is a _Pareto improvement_ over "Draft everyone based on strength"; it's not ilani to _ignore Pareto improvements_ because of American taboos.]

[TODO: misrepresentation of the Light: Dath ilan has a concept of "the Light"—the vector in policyspace perpendicular outwards from the Pareto curve, in which everyone's interests coincide.]