The expected value of longtermism

by Jeroen Bouterse

I’m not sold on longtermism myself, but its proponents sure have my sympathy for the eagerness with which its opponents mine their arguments for repugnant conclusions. The basic idea, that we ought to do more for the benefit of future lives than we are doing now, is often seen as either ridiculous or dangerous.

Usually, this takes the form of thought-experiments in which longtermists are supposed to accept major harm to humanity now in the hope of some huge potential benefit later. Physicist and YouTuber Sabine Hossenfelder, after boasting that she couldn’t be bothered to read William MacAskill’s book What We Owe the Future (“because if the future thinks I owe it, I’ll wait until it sends an invoice”), goes on to interpret a paper by MacAskill and his colleague Hilary Greaves in this way: “a few million deaths are acceptable, so long as we don’t go extinct.”[1] Phil Torres takes aim at an updated version of the same paper, in which Greaves and MacAskill argue that possible futures with huge numbers of people, even if they are unlikely, represent an expected value that outweighs that of the near future. He reads it as implying that MacAskill and Greaves might well press a button that lets 1 million people die now, to increase the chance that 100 trillion people will be born later.

That does sound inhumane, though luckily it seems unlikely that these two Oxford professors would get their hands on such a button, and that they would have a sufficiently strong credence in its reliability. Even aside from that, however, I have learned to stop worrying about what longtermists would do when presented with mega-trolley problems: every time some hostile rendition of a longtermist argument has shocked me into checking out the source, the original text turned out to be orders of magnitude less scary than I was led to believe. If you want to make the case that MacAskill and Greaves would support mass slaughter now for a tiny chance to benefit the future, you have to be willing to stretch specific parts of their argument and ignore the part where they, for instance, explicitly factor in a no-harm criterion (p. 5; p. 30). The idea that longtermist thinkers are making the world safe for cold-hearted utilitarianism, in which faraway ends justify any means, only gets off the ground if you are willing to believe that they are dishonest about where their arguments are taking them.

MacAskill’s What We Owe the Future is the opposite of dishonest: it levels with the reader every step of the way. OK, some may feel taken less than perfectly seriously by the four pages filled with little icons of people (each representing ten billion of them), which apparently serve simply to illustrate that the number of future people could be a lot. (MacAskill reveals afterwards that, under the assumptions he stated, it should actually have been more like twenty thousand pages.) Still, I didn’t feel manipulated, not really. I reckon MacAskill includes these drawings not because I, as a reader, am too stupid to comprehend large numbers; he does it because he knows that we all have weak intuitions for them. It is easy to accept that Some Very Large Number of future people may exist; it is not easy to appreciate how many individuals, families, communities are behind the Very Large Number. Anything that might serve as a crutch is welcome. MacAskill is not dumbing down his findings through such simple images or metaphors, but showing his work by sharing the intuition pumps that helped him himself appreciate what the math represents.

Would you still love life if you were a worm?

I take MacAskill’s argument for longtermism to apply to time the same argument that effective altruism already applied to space: our distance or proximity to other people has no normative implications whatsoever. What matters is, always, the difference we can make. The application of this argument in his previous book, Doing Good Better, is that almost everyone in the affluent world can make a meaningful difference, but that this potential is being severely underused. This is because people are not used to thinking about how to act effectively at long distances, and because when we try to act in the interest of strangers, we don’t naturally receive feedback about how successful we were. It stands to reason, then, that there is a lot of added value in thinking rigorously and collecting evidence.

In Doing Good Better, MacAskill showed that some charitable donations are really orders of magnitude better than others even though we, as the ones making the donations, will never notice the difference. What We Owe the Future seeks to overcome further restrictions to our individual and collective agency – restrictions that are again related to our distance to the lives we can influence, and the fact that we are not ourselves affected by the results of our actions.

The book starts with a thought-experiment that MacAskill will keep coming back to: he asks us to imagine living all human lives, in order of their starting point. “Your ‘life’ consists of all these lifetimes, lived consecutively” (p. 3). In all likelihood, most of this single, extremely long life will take place in the future. In that light, we should see ourselves as “an imprudent teenager” (p. 6, and also here), making decisions that might affect our entire future existence. Later in the book, MacAskill even extends this meditation to encompass all sentient life, because whether the future is a good place does not just depend on how good our human lives are, but also on the quality of lives of nonhuman animals. What would it be like to live through the experiences of all creatures capable of experience?

I like this metaphor as a way of underscoring how real all past, present and future lives are, but I also wonder if MacAskill is not trying to make it do too much. How far are we supposed to take the idea that all subjective life is part of the same story, and what would that actually imply for our evaluation of different possible futures?

First, let’s consider the perspective. When MacAskill says “you would spend nearly 80 percent of your time as a fish”, what is he keeping fixed? In what sense am I identical to, or continuous with the fish? Should I assume that no memories are passed on, no information about what it was like to be a fish – just like in reality? In that case, it is impossible to incorporate the full, actual experience of the fish into my overall evaluation of sentient life (just like in reality). Or should I imagine in some sense being able to ‘look back on’ all these lives from a transcendent point of view; a mind remembering being a fish and a dinosaur as if reminiscing about being a toddler and a child? “Ah, yes, being a pig before factory farming; those were the days.”

In that case, I accept the thought experiment, but I find it hard to believe that such a proto-divine mind would care as much about the absolute and relative quantities of similar lives as MacAskill seems to expect. People do not tend to evaluate their own individual lives via the area under their happiness curve, as MacAskill is well aware: he devotes an entire section to the complicated relation between preference satisfaction and hedonism in measuring happiness, noting that people who report significant hardship in life can end up rating it 10 out of 10 (p. 196). Subjectively, we don’t evaluate our lives through the metric of happiness times duration. Why, then, if we imagine a ‘universal subject’ (of all sentient life), should we speculate it to judge the aggregate of its own experiences through this lens?

On the other hand, why not? Fair enough. Let me give two reasons. The first is that if this fictional mind is anything like ours, its judgment about whether the project of sentience has been “worth it” will have a narrative component. It will see its time as a slaveholder partly in the light of its time enslaved (p. 3). This will influence the weight and meaning of any particular instance of pleasure or suffering. The second is that if this mind doesn’t do stories or connections, and identifies with all sentient life simultaneously, I see no reason for it to count (its own) indistinguishable experiences as separate. It may look back with negative judgment on its lives as a factory farm animal, wishing that it hadn’t been like that; but if some of the billions of lives it led as a caged chicken are similar to each other to the point of being near-indistinguishable, I expect they would appear to this universal subject’s memory ‘as one’. Which would mean that there is less to ‘add up’.

MacAskill’s quantitative approach fares better, I believe, under a social-scientific metaphor: if we accept that every experience (past, present or future) is real, that every life ‘counts’, then we can imagine taking surveys and proceed to do the actual counting from outside of these lives. Importantly, this involves abstracting from each individual life, or from each experience over a given timespan, the thing we are interested in. We could do this by asking people whether they are enjoying their life, or speculating what they would have said if we had been able to ask them, to take just one possible measure. Rather than considering the totality of all lives, then, we consider some measure of goodness – agreeing, for example, that a universe that has more lives classified ‘enjoyable’ or ‘worthwhile’ is better, all else being equal, than a universe that has fewer. Even in my quick sketch, this can take other forms than a hedonic calculus, and MacAskill is open about what exactly the measure is that we are optimizing for. However, there is little doubt that for his math to work, it must be a quantity that we can add up over multiple lives, and multiply by duration or by numbers of lives.

The fat tail

The question is then how confident we are that the things we measure (really or speculatively) track the things we value. Here I think longtermism is on shakier ground than short-term effective altruism, even though it basically works with the same math. I do not think this is a result just of the uncertainty of the future, though I agree with my fellow columnist Charlie Huenemann that it is in many respects impossible to factor in just how clueless we are about the far future. I think the problem lies in the different nature of the counterfactuals involved, even if we imagine having full information about the relevant empirical probabilities.

Quality-Adjusted Life Years (QALYs), the main protagonists in MacAskill’s previous book Doing Good Better, seem to me to map quite well to what we seek to achieve when we want to do good now. They make a fuzzy and implicit and likely incoherent desire precise, and allow us to estimate in a more rigorous and reliable way which interventions achieve better what we already vaguely know we want to achieve. Almost no matter your take on ethics, improving the health and living conditions of the global poor is a good thing; the question is how you and I can do that best, how we can contribute most effectively and what we should leave to better-placed actors. QALYs give us something to count, and the resulting numbers give us something to compare.

Effective altruism places our agency on the margin of an actually existing situation – the world as it is now. To some, this makes it too grounded in the status quo, as if it normatively accepts it. Jacobin Magazine recently published an attack on Peter Singer to the effect that his efforts over many decades to alleviate poverty, battle disease and reduce animal suffering went hand in hand with a resignation to existing structures, which betrays that he “fundamentally, isn’t thinking”. (I presume the authors are philosophizing tirelessly to bring about the revolution, and would have succeeded by now had Singer not been such a formidable reactionary.) If anything, longtermism cannot be accused of status quo bias: it is holistic and utopian, and has to imagine several ways in which the world could be radically different in the future, in order to decide which possible course we want to nudge it towards.

Like I said, however, the math remains the same: the thing we want to maximize is the difference in expected value of the trajectories of history with and without the effects of our agency (p. 32, 256, and here for details). It is in assigning these values that I believe the analogy between (short-term-ish) effective altruism and longtermism breaks down. Not because of the time difference: I agree with MacAskill that the child drowning in the pond is equally real whether they are a neighbor’s child, live ten thousand miles away, or ten thousand years away. The difference is that in the first two cases, the existence of the pond is a given. In the last case, on the other hand, no concrete scenario is defined in which we get to make a local intervention (e.g. removing the child from the pond); rather, any intervention in the present derives its value from the way in which it improves, by replacing it, the entire later world. It may be in our power to remove the pond in advance, or the child for that matter; or, if children and ponds are both valuable, we could aim to arbitrarily improve their absolute numbers, at an acceptable child/pond-ratio.

I believe it is unlikely that QALYs (or something similar that effectively measures the sum of all instantaneous wellbeing multiplied by its duration) are as reliable a proxy for real value in all possible worlds as they are in the actual world. For instance, it is much less clear to me that a universe containing 2N (subjectively-)worthwhile lifespans is (all else being equal) approximately twice as good as a universe containing N, than it is that saving 2n lives now is approximately twice as good as saving n. (I can’t guarantee that reading this sentence twice will make it twice as good.) To MacAskill and Greaves, this question matters, because they are trying to quantify how bad a civilization-ending disaster would be in terms of expected value lost, so that they can make rational decisions about the cost-effectiveness of measures lowering the probability of such a disaster.

In saying that these longtermist arguments hinge on the specifics of your moral philosophy much more than effective altruism, I am not saying anything MacAskill hasn’t thought of: his book contains an entire chapter about population ethics. He himself tends towards the view that total wellbeing remains the thing to maximize, even though this view comes with Derek Parfit’s famous ‘repugnant conclusion’: no matter how many exceedingly happy people live in a given world A, there is always some world B full of people whose lives are just barely above the threshold of being ‘worth it’, but with a population so huge that we still ought to prefer B over A. MacAskill agrees that this is a counterintuitive conclusion, but intuition can be wrong, and the premises that lead to it are strong (180-181).

However, he believes it is possible to hedge even with respect to theories of population ethics. He tends to view uncertainty about the relevant values as amenable in principle to the same probabilistic treatment as empirical uncertainty. Hilary Greaves and Toby Ord have called this the ‘expected moral value’ approach to axiological uncertainty. And, just like empirical uncertainty tends to do in longtermist discussions, the fat tail wags the decision-theoretical dog: if extremely large populations are at stake, theories that maximize total wellbeing make such a large difference if they are right, that they dominate our rational decision-making even if we ascribe only low credence to them.

Long story short: they are taking a weighted average of ethical theories! I wasn’t even aware that this was an option. I suppose that should the debate about whether this is even a valid move get stuck – yes, there are skeptics – we could always assign some conservative probability to it being rational, and multiply by that. (Can we do that, though? I hope for the longtermists that from here on, it isn’t Pascal’s wager all the way up. At some point even they wouldn’t be able to afford multiplying by negative powers of ten.)

Before I succumb to the temptation to ridicule rather than argue after all, I should remind myself what longtermists are using these calculations for in practice. They are not running around pressing buttons that kill a million people now to maybe-possibly save a trillion lives later. What they are doing, is pressing upon us how much sense it makes to make relatively modest investments in projects that may warn us about incoming asteroids, decrease the likelihood of climate disaster, prevent a crippling pandemic, or improve the chances that artificial intelligence will be durably aligned with human interests. Their quantifying tendencies, though I have no doubt that they are sincere, also serve this rhetorical purpose: they help convince us that there are some courses of action that stand out, simply by putting a number to them.

Longtermists don’t trust our unaided common sense to weigh the value of distant lives even nearly correctly. In this, surely, they are correct. Perhaps they err slightly on the other side: perhaps they trust in numbers too much, and get a bit lost in assigning values to possible histories. Perhaps, in tending to scale value linearly with quantity of lives, they are somewhat biased in favor of possible futures in which there is a massive number of people. Perhaps they are sometimes misguided in practice as to which courses of action should even be on the table. Still, we could do much worse. Surely the world is not suffering from an excess of foresightedness.



William MacAskill, Doing Good Better: Effective Altruism and a Radical New Way to Make a Difference. Faber & Faber Ltd: 2016 (paperback).

William MacAskill, What We Owe The Future. Basic Books: 2022.

[1] In the older version of the paper, Greaves and MacAskill say that the effects of our actions on the next 100 or even 1000 years “can be ignored”. On its own, this may plausibly be read as saying that what happens in the next century or millennium does not matter at all. That this is not in fact their argument, is evident from reading the entire paper: the point is that the long-term impact of our actions is dominated by their far-future effects, so that in the context of determining the long-term (overall) expected value of our decisions, short-term effects are negligible. This is still controversial, of course, but it is also quite straightforwardly the well-known longtermist argument that the future is at least potentially very big and therefore very important. In the 2021 version of their paper, Greaves and MacAskill omit the sentence, removing the possible ambiguity. What does Hossenfelder make of this? “So in the next 100 years, anything goes so long as we don’t go extinct. Interestingly enough, the above passage was later removed from their paper and can no longer be found in the 2021 version.” It’s a cover-up!