A (small) critique of total utilitarianism

In total utilitarianism, it is a morally neutral act to kill someone (in a painless and unexpected manner) and creating/​giving birth to another being of comparable happiness (or preference satisfaction or welfare). In fact if one can kill a billion people to create a billion and one, one is morally compelled to do so. And this is true for real people, not just thought experiment people—living people with dreams, aspirations, grudges and annoying or endearing quirks. To avoid causing extra pain to those left behind, it is better that you kill off whole families and communities, so that no one is left to mourn the dead. In fact the most morally compelling act would be to kill off the whole of the human species, and replace it with a slightly larger population.

We have many real world analogues to this thought experiment. For instance, it seems that there is only a small difference between the happiness of richer nations and poorer nations, while the first consume many more resources than the second. Hence to increase utility we should simply kill off all the rich, and let the poor multiply to take their place (continually bumping off any of the poor that gets too rich). Of course, the rich world also produces most of the farming surplus and the technology innovation, which allow us to support a larger population. So we should aim to kill everyone in the rich world apart from farmers and scientists—and enough support staff to keep these professions running (Carl Shulman correctly points out that we may require most of the rest of the economy as “support staff”. Still, it’s very likely that we could kill off a significant segment of the population—those with the highest consumption relative to their impact of farming and science—and still “improve” the situation).

Even if turns out to be problematic to implement in practice, a true total utilitarian should be thinking: “I really, really wish there was a way to do targeted killing of many people in the USA, Europe and Japan, large parts of Asia and Latin America and some parts of Africa—it makes me sick to the stomach to think that I can’t do that!” Or maybe: “I really really wish I could make everyone much poorer without affecting the size of the economy—I wake up at night with nightmare because these people remain above the poverty line!”

I won’t belabour the point. I find those actions personally repellent, and I believe that nearly everyone finds them somewhat repellent or at least did so at some point in their past. This doesn’t mean that it’s the wrong thing to do—after all, the accepted answer to the torture vs dust speck dilemma feels intuitively wrong, at least the first time. It does mean, however, that there must be very strong countervailing arguments to balance out this initial repulsion (maybe even a mathematical theorem). For without that… how to justify all this killing?

Hence for the rest of this post, I’ll be arguing that total utilitarianism is built on a foundation of dust, and thus provides no reason to go against your initial intuitive judgement in these problems. The points will be:

  1. Bayesianism and the fact that you should follow a utility function in no way compel you towards total utilitarianism. The similarity in names does not mean the concepts are on similarly rigorous foundations.

  2. Total utilitarianism is neither a simple, nor an elegant theory. In fact, it is under-defined and arbitrary.

  3. The most compelling argument for total utilitarianism (basically the one that establishes the repugnant conclusion), is a very long chain of imperfect reasoning, so there is no reason for the conclusion to be solid.

  4. Considering the preferences of non-existent beings does not establish total utilitarianism.

  5. When considering competing moral theories, total utilitarianism does not “win by default” thanks to its large values as the population increases.

  6. Population ethics is hard, just as normal ethics is.

A utility function does not compel total (or average) utilitarianism

There are strong reasons to suspect that the best decision process is one that maximises expected utility for a particular utility function. Any process that does not do so, leaves itself open to be money pumped or taken advantage of. This point has been reiterated again and again on Less Wrong, and rightly so.

Your utility function must be over states of the universe—but that’s the only restriction. The theorem says nothing further about the content of your utility function. If you prefer a world with a trillion ecstatic super-humans to one with a septillion subsistence farmers—or vice versa—then as long you maximise your expected utility, the money pumps can’t touch you, and the standard Bayesian arguments don’t influence you to change your mind. Your values are fully rigorous.

For instance, in the torture vs dust speck scenario, average utilitarianism also compels you to take torture, as do a host of other possible utility functions. A lot of arguments around this subject, that may implicitly feel to be in favour of total utilitarianism, turn out to be nothing of the sort. For instance, avoiding scope insensitivity does not compel you to total utilitarianism, and you can perfectly allow birth-death asymmetries or similar intuitions, while remaining an expected utility maximiser.

Total utilitarianism is not simple nor elegant, but is arbitrary

Total utilitarianism is defined as maximising the sum of everyone’s individual utility function. That’s a simple definition. But what are these individual utility functions? Do people act like expected utility maximisers? In a word… no. In another word… NO. In yet another word… NO!

So what are these utilities? Are they the utility that the individuals “should have”? According to what and who’s criteria? Is it “welfare”? How is that defined? Is it happiness? Again, how is that defined? Is it preferences? On what scale? And what if the individual disagrees with the utility they are supposed to have? What if their revealed preferences are different again?

There are (various different) ways to start resolving these problems, and philosophers have spent a lot of ink and time doing so. The point remains that total utilitarianism cannot claim to be a simple theory, if the objects that it sums over are so poorly and controversially defined.

And the sum itself is a huge problem. There is no natural scale on which to compare utility functions. Divide one utility function by a billion, multiply the other by eπ, and they are still perfectly valid utility functions. In a study group at the FHI, we’ve been looking at various ways of combining utility functions—equivalently, of doing interpersonal utility comparisons (IUC). Turns out it’s very hard, there seems no natural way of doing this, and a lot has also been written about this, concluding little. Unless your theory comes with a particular IUC method, the only way of summing these utilities is to do an essentially arbitrary choice for each individual before summing. Thus standard total utilitarianism is an arbitrary sum of ill defined, non-natural objects.

Why then is it so popular? Well, one reason is that there are models that make use of something like total utilitarianism to great effect. Classical economic theory, for instance, models everyone as perfectly rational expected utility maximisers. It gives good predictions—but it remains a model, with a domain of validity. You wouldn’t conclude from that economic model that, say, mental illnesses don’t exist. Similarly, modelling each life as having the same value and maximising expected lives saved is sensible and intuitive in many scenarios—but not necessarily all.

Maybe if we had a bit more information about the affected populations, we could use a more sophisticated model, such as one incorporating quality adjusted life years (QALY). Or maybe we could let other factors affect our thinking—what if we had to choose between saving a population of 1000 versus a population of 1001, of same average QALYs, but where the first set contained the entire Awá tribe/​culture of 300 people, and the second is made up of representatives from much larger ethnic groups, much more culturally replaceable? Should we let that influence our decision? Well maybe we should, maybe we shouldn’t, but it would be wrong to say “well, I would really like to save the Awá, but the model I settled on earlier won’t allow me to, so I best follow the model”. The models are there precisely to model our moral intuitions (the clue is in the name), not freeze them.

The repugnant conclusion is at the end of a flimsy chain

There is a seemingly sound argument for the repugnant conclusion, which goes some way towards making total utilitarianism plausible. It goes like this:

  1. Start with a population of very happy/​utilitied/​welfared/​preference satisfied people.

  2. Add other people whose lives are worth living, but whose average “utility” is less than that of the initial population.

  3. Redistribute “utility” in an egalitarian way across the whole population, increasing the average a little as you do so (but making sure the top rank have their utility lowered).

  4. Repeat as often as required.

  5. End up with a huge population whose lives are barely worth living.

If all these steps increase the quality of the outcome (and it seems intuitively that they do), then the end state much be better than the starting state, agreeing with total utilitarianism. So, what could go wrong with this reasoning? Well, as seen before, the term “utility” is very much undefined, as is its scale—hence egalitarian is extremely undefined. So this argument is not mathematically precise, its rigour is illusionary. And when you recast the argument in qualitative terms, as you must, it become much weaker.

Going through the iteration, there will come a point when the human world is going to lose its last anime, its last opera, its last copy of the Lord of the Rings, its last mathematics, its last online discussion board, its last football game—anything that might cause more-than-appropriate enjoyment. At that stage, would you be entirely sure that the loss was worthwhile, in exchange of a weakly defined “more equal” society? More to the point, would you be sure that when iterating this process billions of times, every redistribution will be an improvement? This is a conjunctive statement, so you have to be nearly entirely certain of every link in the chain, if you want to believe the outcome. And, to reiterate, these links cannot be reduced to simple mathematical statements—you have to be certain that each step is qualitatively better than the previous one.

And you also have to be certain that your theory does not allow path dependency. One can take the perfectly valid position that “If there were an existing poorer population, then the right thing to do would be to redistribute wealth, and thus lose the last copy of Akira. However, currently there is no existing poor population, hence I would oppose it coming into being, precisely because it would result in the lose of Akira.” You can reject this type of reasoning, and a variety of others that block the repugnant conclusion at some stage of the chain (the Stanford Encyclopaedia of Philosophy has a good entry on the Repugnant Conclusion and the arguments surrounding it). But most reasons for doing so already pre-suppose total utilitarianism. In that case, you cannot use the above as an argument for your theory.

Hypothetical beings have hypothetical (and complicated) things say to you

There is another major strand of argument for total utilitarianism, which claims that we owe it to non-existent beings to satisfy their preferences, that they would prefer to exist rather than remain non-existent, and hence we should bring them into existence. How does this argument fare?

First of all, it should be emphasised that one is free to accept or reject that argument without any fear of inconsistency. If one maintains that never-existent beings have no relevant preferences, then one will never stumble over a problem. They don’t exist, they can’t make decisions, they can’t contradict anything. In order to raise them to the point where their decisions are relevant, one has to raise them to existence, in reality or in simulation. By the time they can answer “would you like to exist?”, they already do, so you are talking about whether or not to kill them, not whether or not to let them exist.

But secondly, it seems that the “non-existent beings” argument is often advanced for the sole purpose of arguing for total utilitarianism, rather than as a defensible position in its own right. Rarely are its implication analysed. What would a proper theory of non-existent beings look like?

Well, for a start the whole happiness/​utility/​preference problem comes back with extra sting. It’s hard enough to make a utility function out of real world people, but how to do so with hypothetical people? Is it an essentially arbitrary process (dependent entirely on “which types of people we think of first”), or is it done properly, teasing out the “choices” and “life experiences” of the hypotheticals? In that last case, if we do it in too much detail, we could argue that we’ve already created the being in simulation, so it comes back to the death issue.

But imagine that we’ve somehow extracted a utility function from the preferences of non-existent beings. Apparently, they would prefer to exist rather than not exist. But is this true? There are many people in the world who would prefer not to commit suicide, but would not mind much if external events ended their lives—they cling to life as a habit. Presumably non-existent versions of them “would not mind” remaining non-existent.

Even for those that would prefer to exist, we can ask questions about the intensity of that desire, and how it compares with their other desires. For instance, among these hypothetical beings, some would be mothers of hypothetical infants, leaders of hypothetical religions, inmates of hypothetical prisons, and would only prefer to exist if they could bring/​couldn’t bring the rest of their hypothetical world with them. But this is ridiculous—we can’t bring the hypothetical world with them, they would grow up in ours—so are we only really talking about the preferences of hypothetical babies, or hypothetical (and non-conscious) foetuses?

If we do look at adults, bracketing the issue above, then we get some that would prefer that they not exist, as long as certain others do—or conversely that they not exist, as long as others also not exist. How should we take that into account? Assuming the universe infinite, any hypothetical being would exist somewhere. Is mere existence enough, or do we have to have a large measure or density of existence? Do we need them to exist close to us? Are their own preferences relevant—ie we only have a duty to bring into the world, those beings that would desire to exist in multiple copies everywhere? Or do we feel these have already “enough existence” and select the under-counted beings? What if very few hypothetical beings are total utilitarians—is that relevant?

On a more personal note, every time we make a decision, we eliminate a particular being. We can not longer be the person who took the other job offer, or read the other book at that time and place. As these differences accumulate, we diverge quite a bit from what we could have been. When we do so, do we feel that we’re killing off these extra hypothetical beings? Why not? Should we be compelled to lead double lives, assuming two (or more) completely separate identities, to increase the number of beings in the world? If not, why not?

These are some of the questions that a theory of non-existent beings would have to grapple with, before it can become an “obvious” argument for total utilitarianism.

Moral uncertainty: total utilitarianism doesn’t win by default

An argument that I have met occasionally is that while other ethical theories such as average utilitarianism, birth-death asymmetry, path dependence, preferences of non-loss of culture, etc… may have some validity, total utilitarianism wins as the population increases because the others don’t scale in the same way. By the time we reach the trillion trillion trillion mark, total utilitarianism will completely dominate, even if we gave it little weight at the beginning.

But this is the wrong way to compare competing moral theories. Just as different people’s utilities don’t have a common scale, different moral utilities don’t have a common scale. For instance, would you say that square-total utilitarianism is certainly wrong? This theory is simply total utilitarianism further multiplied by the population; it would correspond roughly to the number of connections between people. Or what about exponential-square-total utilitarianism? This would correspond roughly to the set of possible connections between people. As long as we think that exponential-square-total utilitarianism is not certainly completely wrong, then the same argument as above would show it quickly dominating as population increases.

Or what about 3^^^3 average utilitarianism—which is simply average utilitarianism, multiplied by 3^^^3? Obviously that example is silly—we know that rescaling shouldn’t change anything about the theory. But similarly, dividing total utilitarianism by 3^^^3 shouldn’t change anything, so total utilitarianism’s scaling advantage is illusory.

As mentioned before, comparing different utility functions is a hard and subtle process. One method that seems to have surprisingly nice properties (to such an extent that I recommend always using as a first try) is to normalise the lowest possible attainable utility to zero, the highest attainable utility to one, multiply by the weight you give to the theory, and then add the normalised utilities together.

For instance, assume you equally valued average utilitarianism and total utilitarianism, giving them both weights of one (and you had solved all the definitional problems above). Among the choices you were facing, the worst outcome for both theories is an empty world. The best outcome for average utilitarianism would be ten people with an average “utility” of 100. The best outcome for total utilitarianism would be a quadrillion people with an average “utility” of 1. Then how would either of those compare to ten trillion people with an average utility of 60? Well, the normalised utility of this for the average utilitarian is 0.6, while for the total utilitarian it’s also 60/​100=0.6, and 0.6+0.6=1.2. This is better that the utility for the small world (1+10-9) or the large world (0.01+1), so it beats either of the extremal choices.

Extending this method, we can bring in such theories as exponential-square-total utilitarianism (probably with small weights!), without needing to fear that it will swamp all other moral theories. And with this normalisation (or similar ones), even small weights to moral theories such as “culture has some intrinsic value” will often prevent total utilitarianism from walking away with all of the marbles.

(Population) ethics is still hard

What is the conclusion? At Less Wrong, we’re used to realising that ethics is hard, that value is fragile, that there is no single easy moral theory to safely program the AI with. But it seemed for a while that population ethics might be different—that there may be natural and easy ways to determine what to do with large groups, even though we couldn’t decide what to do with individuals. I’ve argued strongly here that it’s not the case—that population ethics remain hard, that we have to figure out what theory we want to have without access to easy shortcuts.

But in another way it’s liberating. To those who are mainly total utilitarians but internally doubt that a world with infinitely many barely happy people surrounded by nothing but “muzak and potatoes” is really among the best of the best—well, you don’t have to convince yourself of that. You may choose to believe it, or you may choose not to. No voice in the sky or in the math will force you either way. You can start putting together a moral theory that incorporates all your moral intuitions—those that drove you to total utilitarianism, and those that don’t quite fit in that framework.