Hilary Greaves | On the desire to make a difference
HILARY GREAVES: (00:04) The sort of one-sentence summary of my talk is, when we're thinking about benevolence, many people are naturally drawn to frame everything in terms of making a difference. And if you're extremely careful about how you do it, which some people are, that is fine. However, you do need to be careful because if you're not careful, there are various pitfalls, which will lead you away from true benevolence, I will argue, if you're framing everything in difference-making terms. So the message is not βdon't make a differenceβ, it's βbe careful about talking this way. It can go wrongβ.
(00:34) Β So by way of introduction, I will take benevolence to be a concern to promote the good. In its purest form, I'll be assuming benevolence is a matter of preferring better things to worse things. However, as I said, many people are drawn to thinking if there would be benevolent activity in terms of specifically preferring that they themselves make a bigger, a better difference than that they themselves make a lesser or a worse difference. And this can go wrong. So you might think, when people talk in difference-making terms, this is completely innocuous. They're just very naturally and very usefully focusing on the thing that they have control over. There's no point in focusing on preferences that something go better on the other side of the world if it's a thing that you have no control over, think about the difference you can make.β So that aspect of it is completely fine. But there's an implicit assumption there that framing things in difference-making terms won't lead you in different directions to framing things directly in betterness terms and as I said, that's the thing I'm going to be arguing, does not always hold. This talk is based on joint work with several people, Will MacAskill, Andreas Mogensen, and that guy there, Teru Thomas, but all the bits you don't like are probably my fault and not his.
(01:49) So here's a bit of evidence that people talk this way all the time. This is a picture of the cover of Will MacAskill's book. This is a book in which he lays out sort of foundational tenets of Effective Altruism, and the title, even the title is βEffective Altruism and a Radical New Way to Make a Differenceβ. For another more or less randomly selected example, in Loren's introductory talk yesterday, I noticed other people may not have been so tuned on to this particular way of talking, he said he had π¨πͺ,π’π’π’ hours left to make a difference and that was the thing he wanted everyone to think about. So people talk this way all the time, and as I said that can be fine, but let's make sure that it is fine.
(02:27) So by way of warm up, here are a couple of the real-life effective altruist-like decision situations that got us started in thinking about this project in the first place. Suppose you're engaging in philanthropy and your aim is to do as much good as possible with your fixed resources, here are two choice points that you face. Firstly, whether or not to diversify your philanthropic portfolio. So this is the question of whether you should give all of your philanthropic pond to just one charity, let's say, or instead split it between two or more charities. The second choice point is the choice of cause area. For example, you might be considering spending your philanthropic resources on addressing existential risks or alternatively spending them on alleviating poverty. In both of these cases... Anecdotally, in the first case, many people prefer to diversify and they prefer to split their philanthropic pot, particularly if they're feeling subject to certain forms of uncertainty. In the second case, many people feel drawn to alleviating poverty rather than addressing x-risk, even while buying standard arguments that mitigating existential risk has higher expected value per unit spent. So what's going on with those? I will argue that at least one common way of rationalizing those common preferences is a matter of thinking about difference-making in a way that's perhaps morally suspect or anyway, it's not a matter of pure benevolence. So if you're trying to do benevolence, then you shouldn't be doing these things.
(03:54) Outline of the rest of the talk. First, relatively briefly, I'll talk about preferences to make a bigger difference in what I'll call a direct causal sense. And I'll argue that's not the thing we should be thinking about if we're trying to do benevolence. The rest of the talk will be about preferences to make a big difference in the second sense, what I'll call the outcome-comparison sense. But I will argue that in particular, once you bring uncertainty into the mix, if you're tempted to combine your tendency to think in terms of making a difference with a tendency, if you have one, to be averse to risk in what's then a very natural way, then you are certainly led to behaviors, preferences over behaviors and choices that are anti-altruistic in ways that really matter practically. Section 4 of the talk will be about, well if all of those things I've just said a true, like what's going on then with these preferences discussed in section 3 and section 5 will summarize.
(04:47) So firstly, to mention and then set aside preferences to make a bigger difference in the direct causal sense. Consider the following scenario, "Who jumps in?" Abdullah and Jacinta are each passing by when they see a child in danger of drowning. There's only space for one rescuer. Jacinta is keen that she rather than Abdullah gets to be that rescuer. Accordingly, she jumps in quickly before Abdullah has a chance. So I take it in this case, Jacinta's preference can't be a genuinely altruistic one. She wants that she is the one that makes a difference rather than that Abdullah is the one that makes a difference. But the child has no reason to care. From the point of view of the child's welfare, it's a matter of indifference which of these people jumps in. We can suppose, if they're equally effective rescuers. So what seems to be going on with Jacinta is that her concern is not with the child's welfare as such as with some notion of her herself getting to have the personal moral glory or something like that. Nor is this just a philosophers ivory tower example. The same theme crops up in, for example, real life discussions of career choice. So many of the people in this room will be familiar with these arguments from the effective altruist space in particular, that if you want to base your career choice on altruistic grounds, you don't necessarily want to be thinking about where you yourself get to most directly causally interact with the beneficiaries you're trying to help. You want to think about what will cause it to be the case that those beneficiaries are helped more. For example, if you're an animal lover, you might initially be thinking of expressing your altruism by going to work in an animal shelter. But if you realize that by instead taking a job in a bank and donating half your salary, you could hire three charity workers for that animal shelter so that three times as much benefiting animals happens. It seems like if your concern is the welfare of the animals, the genuinely benevolent thing, in at least some versions of this example, you should take the bank job rather than go for the thing that puts you in the closest causal contact with the beneficiaries.
(06:41) So for the rest of the talk, I'm going to set aside this direct causal stuff and talk about a different notion of difference-making, the outcome-comparison sense. In the outcome-comparison sense the difference I make to how good the outcome is, by doing A rather than by "doing nothing" is a matter of the difference in how good the outcome is in the case that I do A versus how good the outcome is in the case in which I do nothing. So this, of course, is only as well-defined as the notion of doing nothing, but in many of the examples we're interested in, I'm going to be making it tolerably clear what doing nothing amounts to.
(07:14) So if we're interested in difference-making in this sense, then it seems like we get to say the things about Jacinta that we want to from the point of view of benevolence. The reason why, from the point of view of benevolence, it's a matter of indifference whether Jacinta jumps in or lets Abdullah do it, is that by jumping in and elbowing Abdullah aside, Jacinta makes no difference in the outcome-comparison term. The outcome for the child is the same either way. So the moral so far is, if you're going to identify benevolent preferences with preferences to make a difference, you'd better be doing it in this sense and not the direct causal sense. So this point is well taken amongst the Effective Altruism community. I'm not saying there are people arguing the opposite, it's just a kind of foundational thing that it's worth being aware of to focus thoughts going forward.
(07:57) Now, let's dig more into this notion of making a bigger difference in this outcome-comparison sense. Even in this sense, and here's the thing that may be a little bit more surprising, it seems like there can potentially be some pathological anti-altruistic features of having these preferences. By way of, again, a sort of warm up, consider this example, "Mild or severe disaster". Jon has heard of a landslide in a neighboring town and Jon plans to join the rescue effort. With apologies, ignore the reference to uncertainty. That's distracting. What I want you to be thinking of instead, is that there are two different versions of this scenario. In the first version of the scenario, the disaster is more severe. So if Jon does nothing in that more severe version, (what did I say for the numbers) π€π’ people will die. Whereas if Jon goes to help with the rescue effort, only π£π¦ people will die. Jon's helping with the rescue effort will lead to six people being saved if the disaster is severe. There's a different version of the scenario in which the disaster is only mild in the first place so only three people are at risk of death. If Jon goes to help in that case, then two out of those three will be saved, one person will still die.
(09:06) So the thing I want to zoom in on this example is that, if Jon always prefers, end of story, a world in which he himself makes a bigger difference in the outcome-comparison sense, then it looks like he must prefer a world in which the disaster is severe to a world in which the disaster is mild, because if the disaster is severe, he gets to make a difference of six lives saved, whereas, if the disaster is only mild in the first place, he only gets to make a difference of two lives saved. So that kind of seems anti-altruistic. He prefers that things be worse for people he prefers the case in which more people die. Surely you think that can't be a matter of benevolence. But you might think that's too quick. And related to the reason why it might be too quick is the fact that this preference at least can't lead to anti altruistic action. It's baked into the story essentially, that Jon has no control over the matter of whether the disaster is mild or severe. And indeed, if he did, then his preferences based on his desire to make a difference would pan out differently because if he could make it the case that disaster was only mild, then that would count as part of the difference he could make. So far so inconclusive. It seems like there's something anti altruistic about these preferences. Maybe there is. But from a practical point of view, at least, it seems like we shouldn't be too worried about this thing. It can't lead him towards worse actions. So that was a case in which we haven't yet introduced uncertainty into the decision scenario that Jon faces. We just compared two different decision scenarios. Things get more anti altruistic when we do introduce uncertainty or at least things can get more anti altruistic if we're not careful when we do introduce uncertainty. That's what the rest of the talk is about. Specifically, I want to focus on what happens if you combine thinking in terms of making a difference with risk aversion in a way that seems natural.
(10:52) Β So back to that diversification example. Here's the example in a little bit more detail. Suppose you have $π£π’π’π’ to spend philanthropically. You could donate to AMF or to SCI. Or alternatively, you could split your pot to donate to both like $ π§π’π’ to one and $ π§π’π’ to the other. Suppose you think there's only a π©π’% chance that AMF does any good. You have various uncertainties about the effectiveness of these charities and you also think, only for different reasons, there's only a π©π’% chance that SCI does any good. So this story is fairly plausible. AMF is a charity that distributes long-lasting insecticide treated bed nets in malarial regions. The case for thinking that that intervention is highly cost effective at doing good rest primarily on the fact that distributing such bed net saves, relatively speaking, many lives of children under five per $π£π’π’π’ donated. SCI has quite a different good-doing story behind it. The case for donating to SCI is that this charity will deworm children in regions where without such treatment, many children are infected with parasitic intestinal worms to the extent that it prevents them from going to school. So the case for funding SCI is generally more along the lines of let's make it the case that these children can go to school so they have better economic prospects in adulthood. The key for my purposes is, while you might have various uncertainties about whether these things really do good in the long run, they're likely to be quite different types of uncertainties because the path to doing long-run good are quite different in the two cases. For example, one goes by saving lives and one goes by improving lives primarily. And as I said, anecdotally in this sort of decision situation, many people who feel those uncertainties and those kind of uncorrelated uncertainties will prefer to split their pot between the two charities rather than to put all their eggs in one philanthropic basket.
(12:37) More on that second example. I put this up at the beginning. Suppose again, you have $π£π’π’π’ to spend philanthropically. This time you're thinking about the question of cause prioritization. You could direct your thousand philanthropic dollars towards poor country health. For example, one of the charities on the previous slide or alternatively, you could direct it towards mitigating existential risks. For example, you could use your resources to back some AI safety effort. Suppose you buy standard arguments to the effect that the expected number of lives saved or equivalent is vastly greater for the x-risk intervention. That is, you think the expected amount of good done is greater by x-risk mitigation, however, you think that there's a market difference in the probabilistic structure, if you like, of these two interventions. So you think that AMF or SCI save some relatively modest number of lives or does an equivalent amount of good for sure, whereas AI safety work is more a matter of almost certainly doing no good at all and then with tiny probability doing an enormous amount of good. Anecdotally in this situation, as I said, many people prefer to fund poverty alleviation or the health measure, even if they accept the expected value argument that favors x-risk mitigation. And on one backstory about why you might have that preference, the rationalization turns out to be very similar to the one that seems to be behind the diversification preference.
(13:57) So what's that rationalization? It's about risk aversion. So in general, as most people here are probably aware, risk aversion is a matter of dispreferring mean preserving spreads. That is, generally preferring actions that have something closer to a sure outcome and dispreferring actions for which the variation in possible outcomes across states of nature is more extreme other things like expected value being equal. So for example, if we're in a monetary context, many people would prefer that I give them $π£π’π’ for sure, rather than that I flip a fair coin and give them $π€π’π’ if the coin lands heads and nothing if the coin lands tails. If you have those preference, then you're risk averse with respect to the amount of money that I give you. So intuitively, you might think, both not diversifying and funding the x-risk intervention, so both of those things are reported anecdotally tend to be dispreferred by many people, are more risky than their alternatives. If you don't diversify, you're putting all your bets on the π©π’% chance that AMF is cost effective, whereas if you split your pot because the probabilities were reasonably uncorrelated, there's at least less chances that your intervention ends up doing no good. And somewhat suddenly, with the x-risk intervention, it seems like in that case, if you fund the x-risk mitigation, almost certainly with probability very close to one, you do basically no good. The argument for funding that intervention is based on this tiny little tail probability that you do an enormous amount of good. But you might think that as very risky from the philanthropic point of view.
(15:23) Well is it? It depends on what kind of risk aversion we're talking about. It is crucial to distinguish between what I'll call standard risk aversion, which is the standard thing and this difference-making risk aversion which is the one that's going on in the arguments that I just gestured at on the previous slide. So standard risk aversion in the philanthropic context will be risk aversion with respect to total goodness, if you like, risk aversion about how good the complete outcome is. That's not what's going on in the arguments just gestured at. Instead it has to be what I'll call difference-making risk aversion, that is, being averse to risk in the difference that you yourself make. If you are risk averse, in the standard sense, I won't go through the details, but you would not in general get a preference for diversifying your own individual philanthropic portfolio and you probably wouldn't and you wouldn't in general, get a preference for poverty alleviation over x-risk mitigation.
(16:15) So what I want to argue in most of the rest of the talk then is, while difference-making risk aversion seems to be what's behind these intuitive preferences that I've anecdotally reported lots of people have, there's something anti altruistic about that thing, at least. So if you're going to be thinking in terms of making a difference, at least avoid being risk averse with respect to the difference you make. Then we try to convince you then that there's something anti altruistic about those preferences. So there will be two arguments for this conclusion. One focuses on the notion of Collective Defeat. That one's coming up first. The second one will be about Stochastic Dominance. That one comes second.
(16:51) So first of all, collective defeat. Before I say what collective defeat is, let me just sketch the scenario that the discussion will be raised around. So here's a decision scenario. It's going to involve two people, you and me if you like, say for rescue lifesaving. You're a philanthropist. You're facing a choice between two interventions you could fund. We're going to call one of them safe and one of them risky. Here's what they do. If you choose to fund safe, then compared to the scenario in which you do nothing, which in this case means something like keep your dollars in the bank or spend them on very expensive coffee or something like that. If you fund safe compared to that βdo nothingβ scenario, π€ additional lives will be saved. If you fund risky, then again, compared to the scenario in which you do nothing, you're either going to save π’ or π§ lives and it's π§π’/π§π’ probability. So you can think if you like of a flip of a fair coin, π’ lives saved by you, if that coin lands heads and π§ lives saved by you, if that coin lands tails. I face a very similar decision. I also face the choice between safe and risky. Although in the event that I choose risky, the correlation between how many lives I save and the outcome of the coin flip is the opposite to the one that was in your case. So it's the same coin both ways. But whereas if the coin lands heads, you will save π’ lives, if the coin lands heads, I will save π§. If the coin lands tails, you will save π§, but if the coin lands tails and I choose risky, then I will save π’.
(18:12) Why is this scenario interesting to consider from the point of view of the discussion we're currently engaged in? To see that, let's compare the scenario in which both you and I fund safe to the scenario in which both you and I fund risky. Two observations are worthy of note. The first one is that the outcome is certainly better if we both fund safe than if we both fund risky. If we both fund safe, then we know that π¦ lives are saved, π€ by you and π€ by me. If we both fund risky, we know that π§ lives are saved in total, although whether it's me or you that is the one who does the saving is subject to randomness. It depends on which way the coin landed. So the outcome is certainly better if we both fund risky, but that may not be the thing that's directly driving our decisions. Because remember, we're thinking in terms of difference-making and we're risk averse with respect to the difference that we individually make. And risky, as the name suggests, is more risky from that point of view. So if you have difference-making risk averse preferences, if you're risk averse with respect to the difference that you yourself make, you may well prefer to fund safe. In some version of this scenario with a number specified appropriately, you will prefer to fund safe rather than risky and similarly for me. So it looks like that there's going to be a danger that a pair of agents with these difference-making risk averse preferences will both choose safe, thereby with certainty bringing about a worse outcome. So when that happens, I'm going to say that these patterns of preferences lead to collective defeat of goodness.
(19:44) So turning that into an argument... So that's just a fact. That is true of that scenario. I want to try and turn it into an argument against the proposition that difference-making risk averse preferences are matters of true benevolence. That's a slightly thornier thing to do. But here's the basic idea. We'll define collective defeat of goodness as follows. Say that preferences P can lead to collective defeat of goodness, just if there exists situations like the one on the previous slide in which agents acting in accordance with those preferences firstly, produce an option profile that's statewise dominated in terms of goodness by some other option profile. And secondly, that this is true regardless of their credences about each other's actions. So that happens in the scenario on the previous slide. The intended argument then is, firstly, that difference-making risk averse preferences can lead to this phenomenon. That's what I tried to show on the previous slide. Second premise is going to be the premise that purely benevolent preferences can't do that thing. And then conclusion, difference-making risk averse preferences in that case aren't purely benevolent. So the underlying premise here is the important one. That's the thorny one. That's the one that I haven't argued for. I think it's kind of plausible, although I also think it's not completely straightforward to argue for it.
(20:51) So now the print gets a bit smaller. So digging into whether that second premise is true. You might think, well isn't there a kind of partners-in-guilt argument here? Because we know that even under completely standard consequentialist type reasoning, something a bit like collective defeat can happen, can't it? So isn't this just the same thing again. So I think it's not that simple. It's true that something a bit like collective defeat can happen, even under preferences that are clearly the best contenders for benevolence that are in town. But there's an extra aspect of collective defeat going on in these cases. So the things we already know quite aside from thinking about difference-making is firstly, if agents have inaccurate credences about what each other will do, then you can have some kind of collective defeat type thing going on. So for example, suppose you and I both want to meet for lunch. The main thing we care about is just that we meet each other, but we can't communicate, we've forgotten our phones. There are two possible places we could go to lunch, the curry place and the pizza place and as bad luck would have it have it, you go to the pizza place and I go to the curry place. And so we fail to meet. So we collectively bring about a sub optimal outcome, but intuitively in this case, the culprit is our false beliefs about what the other person would do. The culprit isn't the preferences themselves.
(22:07) Second case in which something a bit like collective defeat can happen. If there are multiple Nash equilibria, and again the agents can't communicate, then it could turn out that they coordinate on the suboptimal Nash equilibrium even while having correct beliefs about what one another will do. So again, with curry and pizza. Suppose that you and I both strictly prefer curry to pizza, but we both believe that the other one will go to pizza, and the main thing we prefer is that we meet. We care about that more than about what we have for lunch. So because I believe that you will go to the pizza place, I go to the pizza place, because you believe that I will go to the pizza place, you go to the pizza place and we meet for pizza. We were both correct as it turns out about what the other one will do. Nonetheless, it's a suboptimal outcome. It would have been better if we'd both gone for curry.
(22:47) So that's another way in which you could have these sort of collective defeat type thing could happen. But again, that's not what's going on in the difference-making case. And then finally, I won't go through it in detail, but if the agents have different empirical credences about which options can lead to which outcomes, then they can fail to coordinate because what's driving their preferences is the expected utility of the options and not directly the utilities of the outcomes. However, under standard consequentialist type reasoning, if you have a shared utility function on option profiles and you have the same empirical credences, then you can't have collective defeat. So you can't have collective defeat in the sense that's been defined on the previous slide. In the safe versus risky life-saving example, the kind of collective defeat that's going on is over and above any of the kinds that we're used to, and at least intuitively, it feels like the culprit in this case is the fact that what's driving us is our desire about the differences we ourselves make. It's sort of an essentially agent-centered thing that's driving our actions rather than genuine agent-neutral benevolence.
(23:50) So you may or may not find that convincing. In case you don't, here's a different argument for the proposition that difference-making risk averse preferences aren't matters of pure benevolence. So to lead on to that, notice that what goes on in the collective defeat cases is that the outcome will be better not if just one agent acted differently. The key observation is that things would be better if both agents acted differently at the same time. And then you might think, well here's a reason for not really being bothered about that, for not thinking it shows that there's any failure of benevolence going on. No individual agent can bring it about that both agents act differently. Then that's the reason why you might be skeptical about the underlying premise in the collective defeat argument.
(24:34) So now I want to go on to the stochastic dominance based argument, where we're not going to be thinking about two agents acting and what would happen if both agents acted differently. There's only going to be one agent in the story. And so I'm going to argue there's going to be a tension between what it seems is mandated by pure benevolence and what a difference-making risk averse agent might choose.
(24:56) So stochastic dominance. If we didn't have uncertainty in the picture, it will be obvious what it would mean for preferences to track betterness. If you don't have uncertainty in the picture then to each option there corresponds an outcome and for your preferences over options to track betterness is just for you to always prefer the option with certainty in these scenarios β it has a better outcome than the other option. What we want to do is keep hold of that idea that if you're purely benevolent, then your preferences must track the betterness relation, but extend the betterness relation that we're talking about to the certain contexts that include uncertainty. Now, there's no obviously correct way of doing this in full generality because, if you like, in full generality in the presence of uncertainty, there's no such thing as the outcome of A in the relevant sense. We're trying to do ex ante evaluation of prospects rather than evaluation of outcomes directly. However, sometimes it's the case that one option stochastically dominates another and the idea is going to be that when that happens, at least it makes sense to say that the first option is better than the second from the point of view of the goodness of outcomes. If A stochastically dominates B, then A is better with respect to goodness of outcomes than B.
(26:08) So here's the definition of stochastic dominance. We say that A stochastically dominates B, if firstly for every possible outcome O, the probability of getting an outcome that's at least as good as O is at least as high on A as it is on B. And secondly, for some possible outcome O, the probability of getting an outcome that's at least that good, is strictly higher on A than it is on B. Okay. I'm guessing stochastic dominance will be very familiar to some people in this room and unfamiliar to others. So if you're in the first category, take a two-minute nap. If you're in the second category, here's an example that might help.
(26:43) So just a very simple example to illustrate stochastic dominance and to help you get into the frame of mind of thinking, oh yeah, this is a case where A stochastically dominates B. That kind of means A is better than B. So we're going to consider three options. And if you take option A, let's suppose you have a β probability of getting each of the following amounts of value β probability of getting π’,Β β probability of getting π§, β probability of getting π£π’. You can think of these as being, if you like, numbers of lives saved. If you choose option B, things are similar, so β probability of getting π’ and still β probability of getting 10. But in that middle β probability case, instead of getting π§ as you did on A, if you choose B, you get π¨. And if you choose option, C, this is a variation on A in a different sense, there's still a β probability of getting π£π’ units of value and you still might get π’ and might get π§ if you don't get 10. But compared to A, we've shifted some of the probability mass from getting π’ to getting π§. So here, you've got only β probability of getting that worst outcome π’ units of value. You've got a large probability Β½ of getting the middle thing π§ and still the β of getting π£π’.
(27:54) So if you stare at these things and you stare at the definition of stochastic dominance, you see that B stochastically dominate A and C stochastically dominates A, and the reasons for those are related to the stories I told as I introduced the examples. B stochastic dominates A because you've kind of taken that middle case and made the outcome better. C stochastically dominates A because you've focused on those first two outcomes and shifted a bit of probability mass from the worse one to the better one. So that should make it intuitive that this stochastic dominance thing is appropriately regarded as a betterness relation on prospects. It's, of course, only a partial ordering because often, you'll have two options and neither of them stochastically dominates the other. But the idea is when one option does stochastically dominate the other, then any sane view of betterness of prospects will regard the first as a better prospect than the second one. So because of that then, the idea is β so now you can wake up again if you went to sleep β if one option stochastically dominates another with respect to outcome goodness, then benevolent preferences have to track that stochastic dominance relation. They have to prefer the first option to the second.
(28:58) And the point of hammering that is that difference-making risk-averse preferences don't meet this condition. They violate stochastic dominance. So here's the example that establishes that. I call it βswap and sweetenβ. So a fair coin is going to be flipped. The agents choice set is either βdo nothingβ or do A or do B. We don't really care about the βdo nothingβ option. That's only there to define how much difference you make if you do A or B. What we're interested in is whether you do A or whether you do B. The possible outcome goodnesses are as follows: If you do nothing, you're going to get π£π’ units of goodness. There will be π£π’ units of goodness if the coin lands heads and π’ if the coin lands tails. If you do A then you get an extra π£π’ either way. So in terms of outcome goodness then, π€π’ units of goodness if heads and π£π’ units of goodness, if tails. If you do B then you're only going to get a tiny little bit extra goodness, namely x, so think of x as being some very small number, if the coin lands heads and you're going to get a whole π€π’+x units of goodness if the coin lands tails. So the outcome goodness are the ones in that left hand table there.
(30:02) So here we've got two tables. The table on the left shows you how good in total the outcome will be if you do nothing, if you choose A, if you choose B. The table on the right shows you how much difference you make compared to the βdo nothingβ option, if you do nothing, if you choose A, if you choose B. So of course, the first line of the table on the right has to be 0. Zero because the comparison is to the scenario in which you do nothing. And then you have the other numbers by doing the subtractions. So π£π’ here is because the difference between this cell and this cell is π£π’. x here is because the difference between this cell and this cell is x and so on.
(30:39) So the first observation is that B stochastically dominates A. Why is that? Well, we compare the Β½ probability in which you do A and the coin lands tails, where you get π£π’ units of goodness. There's a Β½ probability if you do B of getting π£π’+x instead of π£π’. So we're kind of doing the diagonal comparisons to see if the stochastic dominance relation holds. So instead of a probability of a half of π£π’, if you do B, you get π£π’+x. Instead of the probability of Β½ that A gives you π€π’, if you do B, that probability of half of π€π’+x, so B stochastically dominates A. A different question is, which of these options will a difference-making risk averse agent prefer, and you can see there's a danger they're going to prefer A to B. If you specify the numbers appropriately, they will prefer A to B. And intuitively that's because in terms of the difference the agent themselves makes, A is completely safe. They definitely get to make 10 units of difference, whereas B is risky, sort of either almost nothing or loads. So we have this scenario in which B stochastically dominates A and so according to what I was arguing on the previous slide, B is better than A on any reasonable view, and yet, the difference-making risk averse agent is going to prefer A to B. And so I'm going to conclude the difference-making risk averse preferences can't be matters of pure benevolence. So here's a sort of writing it out the way philosophers do so you can point to the premise you disagree with if you don't like the argument.
(32:00) Here is the argument. Firstly, difference-making, risk-averse preferences can lead an individual agent to prefer an option that's stochastically dominated with respect to goodness. So that's what the βswap and sweetenβ example was there to show. Secondly, I thump the table and insist, purely benevolent preferences, relatedly preferences of a benevolent third party, if you like, cannot prefer an option that's stochastically dominated with respect to goodness and therefore difference-making risk averse preferences are not purely benevolent. So again, the first premise is what I was trying to just prove by showing you an example. The second premise, maybe there's some room to argue, but I think this is actually harder to argue within this case than in the case of the first argument. And the conclusion just follows.
(32:45) This slide, I won't go into details. So this slide basically says things are somewhat similar, at least somewhat similar for ambiguity aversion, as they are for risk aversion. So again, you can find some situations where there's some intuitive preference, or at least seems to be intuitive to lots of people, and in particular, it's a preference that turns out to be important in various philanthropic decision-making scenarios where it seems like the preference goes against the arguments that straightforward expected value maximization suggests, but it seems like the preferences could maybe be rationalized in terms of ambiguity aversion. Again, though, it turns out that the preferences in question can only kind of obviously be rationalized by ambiguity aversion if it is aversion to ambiguity in the difference you yourself make, what we might call difference-making ambiguity aversion, rather than ambiguity aversion in a more standard sense where you're averse to ambiguity and how much total value there is in the world, for example.
(33:38) So to give you an example to illustrate where these preferences come up, think, again, of that existential risk or health example. One thing that bothers many people about funding x-risk rather than funding health interventions is that if you're doing health interventions, you can do randomized control trials to establish what works. And maybe you're not certain, but at least you know what the probabilities are. You feel like you're on solid, experimental ground when you write down the numbers for your expected value calculation. If you're doing existential risk mitigation, then you consider funding an AI safety or contributing your funds to an AI safety effort and you think, βWell, I really have not much clue by how much I changed the probability of doom if I do that thing. If you forced me to write down the number for the purpose of doing an expected value calculation, I kind of feel like I'm making a number upβ. I know the number is very small. It's sort of plausible that the number's big enough to get expected value to work out in favor of x-risk mitigation, but there seems to be in a relevant sense quite wide uncertainty about what the probabilities are. And so that's why it seems plausible that ambiguity aversion might sway you against funding x-risk mitigation. You'd like to do randomized control trials, but then you realize that if your concern is preventing extinction, that's not really going to pan out. And so you're inclined to stay away from that territory and do something where experiments can help guide your judgments. But then again, you can see when you think about it, you have to be averse to ambiguity in the difference you yourself make for that reasoning to make sense. If you're just averse to ambiguity and the total prospect faced by humanity, there seems no particular reason to think that you funding existential risk mitigation will reduce that total uncertainty rather than increasing it, after all, it was already extremely ambiguous what the prospect of doom was.
(35:23) So the last thing here, which I didn't go into detail about is that there's an analogue of the collective defeat argument for difference-making ambiguity aversion. Perhaps interestingly, if you are convinced by the second argument and not the first, concerning risk aversion, there isn't a direct analogue of the stochastic dominance argument for ambiguity aversion. So maybe difference-making ambiguity aversion is on better ground than difference-making risk aversion.
(35:51) So where are we now? So we've framed everything in terms of our desire to make a difference. We've tried to combine that with risk aversion and what seemed like a natural way. So we've been led to difference-making risk aversion. Then we've noticed that difference-making risk aversion seems to be in tension with what pure benevolence would recommend in at least some cases. So this leads to the question of what we should make of this. Is there any available defense of the idea that a desire to engage in the project of benevolence would lead one to have difference-making risk averse preferences? So I think that's an interesting question to ask partly because many people find that even after they're aware of all the things I just said, they still feel a pull towards difference-making risk aversion. So what's up with that?
(36:33) Here's a conjecture. Perhaps the desire to make a difference, an aversion to risk in the difference one makes, is a matter of something like desiring that your own life be meaningful. You desire that you make a contribution, not just that the world go well, but that you do something significant with your life. If so, while you have an agent-centered preference, it is about how your life goes. But it's a kind of combination of agent-centered preference with benevolence because you wanted that you make a difference to how good things are. But at least there's that agent-centered element in there. So then, it seems like the thing to say about those preferences is, yeah okay, maybe for the reasons stated in this talk, it is not a matter of pure benevolence but unless you're a thoroughgoing consequentialist who thinks that benevolence is all there is to morality, it doesn't immediately follow that there's anything morally odious about those preferences. Maybe it's just a kind of preference that belongs in the non-consequentialist part of morality, if there is such a part. So I think that's an interesting line of thought to explore. It certainly has some plausibility to it. It also strikes me though, that having the difference-making risk averse preferences for that kind of reason might be unstable under recognition of the points noticed in this talk. Because again, the line of thought driving difference-making risk aversion was in the minds of most people who engage in this kind of behavior, supposed to be driven by benevolence in the first place. We started off thinking we want to contribute to this project of benevolence with our lives, how do we go about doing that? Or at least that's how many of us got started on it. And if you are trying to do that thing and then you realize that difference-making risk aversion is not doing that thing, then you might think, well, the truly meaningful thing to do with my life then would be to do the thing that's in line with actual benevolence, not this kind of somewhat messed out halfway house that in certain situations at least pushes against what pure benevolence would recommend.
(38:29) In summary, altruists are often drawn to thinking in terms of making a difference. This is, in principle in tension with altruism because altruism is supposed to be an agent-neutral matter and being concerned with the difference you yourself make is at least in principle, a self-centered, an agent-relative concern. I've argued in this talk that it can also be in tension with altruism in practice firstly, if it amounts to a desire to make a difference in what we call the direct causal sense, as in the example of Jacinta. And secondly, if it leads to difference-making risk aversion, even if difference-making is interpreted in the outcome-comparison sense. It seems to have more prospects, initially, of being in line with pure benevolence. That last bullet point was because of two reasons: firstly, the argument about collective defeat and secondly, the argument that difference-making risk averse preferences violates stochastic dominance with respect to goodness.
(39:21) So to arrive at genuine altruism, I suggested, we need to be focusing on wanting the world to be good fundamentally speaking, rather than wanting that we ourselves make a difference to that. And if you're finding yourself inclined to frame things in terms of making a difference, that can be fine if you are interested in, say maximizing the expected value of the difference you make to goodness and you didn't have any risk aversion in the picture. Then for at least for practical purposes, your preferences would coincide with those that you would get by simply trying to maximize expected value. So you would not then be led astray by your tendency to think in terms of making a difference. But it does seem to be the case that once people start framing things in terms of difference-making, they are at least often driven to these versions of it that are in tension with benevolence. So those are the things that we need to guard against if we're trying to do pure benevolence and we're framing things in terms of difference-making. Then I said at the end that in those cases where difference-making preferences take a form that is in tension with genuine benevolence. What's going on might be a matter of wanting one's life to be meaningful, but I at least suggested that the explanation might not ultimately be coherent.
(40:28) Thank you.
