Consider three moral claims, each seemingly plausible and, in conjunction, accepted by various philosophers.1

The first is Impartiality: that the moral value of a life does not depend intrinsically on when or where it occurs; that, for instance, a human life lived millions of years in the future would be no more or less valuable than an otherwise identical life lived today.2

The second claim is that what matters for comparisons of moral betterness is the total sum of such value. Call this Additivity, the claim that: an outcome is at least as good as another if and only if the former contains at least as great a total sum of (some cardinal measure of) the value within each individual life.3

The third claim is that, when comparing risky options, expected value theory holds. This common view says that the morally best prospects over outcomes are those with the highest expected moral value—if we also accept Additivity, those with the highest probability-weighted sum of total moral value.4

But, the conjunction of these three claims has troubling implications. If we face options with certain probability distributions over total moral value, those options’ expected values will be undefined. We will be unable to say that their expected values are greater than, equal to, or less than any other. Nor, by expected value theory (and no stronger principle for comparing risky options) can we say which of our options are best. The probability distributions that generate this result include well-known troublemakers from decision theory, such as: the Pasadena game (originating in Nover & Hájek 2004), and the Agnesi game (see Alexander 2012; Poisson 1824). But the problem is not merely hypothetical, as those trouble-making cases from decision theory are. As I will argue, these three claims lead to a practical ethical problem. In practice, there is reason to think that the total moral value of the future follows a similarly problematic probability distribution. And, if so, every option we might ever choose in practice will have undefined expected value.5

If we have one of these problematic probability distributions over the total value of the future, and we accept Impartiality, Additivity, and expected value theory alone, then we face a dire reductio ad absurdum. For every option ever available to us in practice, we cannot evaluate it; we cannot compare it to any other such option, not even to options identical to itself. We can never say how our options compare in terms of moral betterness.

This implication seems absurd. But it is not immediately clear how we might avoid it, at least without abandoning Impartiality or Additivity—without admitting that the time at which a life is lived can matter morally, nor admitting that the ranking of outcomes deviates from that of their total values. If we find both claims compelling, can we hold onto them and extend our comparisons to prospects without slipping into absurdity?

One way we might do so—which I will discuss but not ultimately endorse—is by replacing expected value theory with an alternative theory that exhibits sensitivity to risk (e.g., expected utility theory with a non-linear utility function, or a version of risk-weighted expected utility theory). With the right profile of risk aversion and risk seeking, such theories may effectively replace prospects like the Pasadena game with better-behaved ones. Given this, we may have a novel argument for risk sensitivity in the moral context: it seems we may need to be risk-sensitive to compare our options at all in practice.

In this paper, I seek to determine, in effect, whether risk sensitivity is the only way out. If you find Impartiality, Additivity, and the risk neutrality of expected value theory convincing, is there some way to salvage them?6 Can the above reductio be avoided without allowing risk sensitivity, or denying Impartiality or Additivity?

To preserve risk neutrality, it is necessary to extend expected value theory to compare troublesome options. The literature already features various proposals for how to do so (e.g., Colyvan 2008; Easwaran 2008; 2014b; Meacham 2019). But, as it turns out, many of these existing proposals fail—I argue that they cannot compare the options we face in practice. Despite this, I describe a theory that may succeed in doing so. With such a theory, we can compare the problematic options I describe. And so we may be able to avoid the reductio that expected value theory, Impartiality, and Additivity brought upon us, and do so without endorsing risk sensitivity.

Decision theorists have long recognised prospects that lack well-defined, finite expected values. Some prospects lack such expected values because they feature outcomes with infinite value, such as in Pascal’s Wager. But I will set aside such prospects in this paper, and assume that outcomes must have only finite value.7

But even if we exclude infinitely valuable outcomes, some prospects still lack well-defined expected values. One frequently discussed such prospect is that of the Pasadena game.8

What is the game’s expected value? If we try to calculate it in the order the outcomes are listed, we obtain the series 1−12+13−14+…+(−1)n−1n+…. This series, also known as the alternating harmonic series, fails to be absolutely convergent. If we were to naively add it up in one order or another, so long as we picked the right order, we could obtain any total we wanted.9 So, we cannot say that the game has any particular expected value at all (see Nover & Hájek 2004)—in this sense, the Pasadena game defies expectations (or is expectation-defying). And so expected value theory cannot tell us how it compares to any outcome, nor to any other option, nor even to itself. If options were to be compared by expected value theory alone, then the Pasadena game would be no better than, no worse than, nor equally good as any other option.

A similar prospect is the Agnesi game. Unlike the Pasadena game, it gives a continuous (rather than discrete) probability distribution over possible values. It can result in an outcome of any real value v; its probability density over value is given by the following function, also known as the Witch of Agnesi or (an example of) the Cauchy distribution.10

On a graph, its distribution looks like this, symmetric about 0.

Try to take the expected value of this prospect and you will find that it has none (Poisson 1824). For continuous distributions like this, the expected value is given by the integral of v×p(v) from negative infinity to positive infinity (analogous to an expected sum: v×P(v) over all possible values v). But, for the Agnesi game, that integral between 0 and positive infinity is positively infinite! And, from 0 to negative infinity, it is negatively infinite! Sum these integrals together—equivalently, take the integral over all possible values of v—and the total is undefined. Much like the Pasadena game’s expected sum, the Agnesi game’s expected integral fails to converge absolutely. You might expect its expected value to converge to 0, since the distribution is symmetric about 0. But, no, it too defies expecations. So, expected value theory will fail to compare it to any outcome, to any other option, and even to itself.

You might think that neither of these prospects are realistic—that they are merely contrived, hypothetical options that we are sure never to encounter beyond the pages of a philosophy journal. As Hájek (2014: 565) says of one description of the Pasadena game, you might think that considering such an option is “…a highly idealised thought experiment about a physically impossible game.” If so, you might not be troubled that expected value theory cannot compare either of these prospects to any other. You might think that we should simply ignore such prospects, and that expected value theory still suffices for real-world decision-making.

Unfortunately, we cannot—there is reason to think we face such prospects in practice. When evaluating our options morally, if we consider the prospects for the moral value of the distant future and we maintain Impartiality and Additivity, then we have reason to think that every option ever available to us defies expectations. In the remainder of this section, I give two discrete arguments to this effect, the second more compelling than the first.11

But, first, a brief note on probabilities. I assume that the notion of an outcome’s probability that ultimately determines moral betterness must be one of two notions. The first is its evidential probability: how probable that outcome is to result from the given option, on the present evidence of the agent deciding between that option and others (see Williamson 2000: 209). The second possible notion is the outcome’s subjective probability: how confident the decision-making agent is that that outcome will result from a given option.12 If evidential probabilities are the morally relevant ones, and if our evidence prescribes expectation-defying prospects, then we will face difficulties. Or, if subjective probabilities are the relevant ones, agents who form their beliefs rationally given their evidence will face difficulties.

If you model the value of the distant future as involving any non-zero probability of the Aquila game, the Skewed Aquila game, or any similar such game, then you face a serious challenge. In practice, you cannot assign an expected moral value to any of the options ever available to you. So, if expected value theory (and nothing stronger) were the correct theory of moral betterness under risk, then no option ever available to you would be morally better or worse than any other. But to accept this implication would be absurd.

To plausibly compare any of our available future prospects, we must replace expected value theory with some stronger alternative. In later sections, I will discuss such alternative theories. But, first, what do we want them to achieve?

I propose five problem cases that those theories must be able to deal with—and deal with in the intuitively correct way—to be extensionally adequate.25 Extensional adequacy may require dealing with cases even more complicated than those I consider here. These cases will be crude simplifications of the options we face in practice: they exclude all sources of value in the world other than the possibility of an Aquila game generated after t. In practice, we face options in which many valuable events will occur before t, in which there is perhaps only a small probability of life surviving until t to generate something resembling an Aquila game, and in which the prospect for events after t is far more complicated than the Aquila game. Nonetheless, it will largely suffice to consider simplified cases like these—as noted above, our available options will inherit the problems of the Aquila game. If expected value theory fails in the cases below, it will fail in practice. And it turns out that it does fail, as do many stronger theories designed to deal with the original Pasadena and Agnesi games.

The first problem case, No Change, is (a simplification of) the decision scenario an agent faces when their available actions all produce exactly the same future prospect. For instance, an agent may choose between eating Sugar Puffs for breakfast and eating Frosties, but have no evidence for either option being more or less likely to influence the future in any particular way. (Agents with great foresight may have access to evidence supporting some story of why one cereal is more likely to produce better long-run outcomes, but suppose that the agent here lacks any such evidence.)

Scenario 1: No Change

Note that both options have identical probability distributions over value. But, still, bare expected value theory cannot say how they compare—neither option has well-defined expected value, so that value cannot be equal to itself. (The same goes if we swap the Aquila game for the Skewed Aquila game.) And this is all the more troubling when, intuitively, the correct ranking of options seems clear: Sugar Puffs and Frosties are equally good. It would be desirable for our theory to say this, that any instance of the Aquila game with such and such parameters is equally as good as any other with the same parameters.26

The second problem case, Improving the Present, is that which an agent faces when they can improve some aspect of the world with certainty,27 without otherwise changing the prospect. For instance, an agent may choose whether to save the life of a child in the present day. And, regardless of whether they do so or do not, their evidence may entail an identical probability distribution over what happens in the very distant future. If so, then, for my purposes, their options are equivalent to the following:

Scenario 2: Improving the Present

Here, both options are identical except that the latter has its probability distribution shifted by some bonus value s.28 But, again, expected value theory cannot compare any options of this sort. (And again, the same goes if we swap the Aquila game for the Skewed Aquila game here.) And, again, this is all the more troubling given that the intuitively correct ranking is clear: that, as long as s is positive, Save a Life should be better than Do Nothing. Improving every outcome should improve the option overall (so long as the outcomes’ probabilities are held fixed).

The third problem case, Improving the Future, is one that an agent may face if they attempt to improve events occurring after t, conditional on us surviving until then. In this case, the agent’s choices don’t make it more or less likely that we survive but, if we do survive until then, those choices make it more or less likely that the average life afterwards will have positive value (on whichever interval scale we represent value). It is a decision of whether to alter the skew of the Skewed Aquila game one way or the other. And, to an approximation, this is the sort of case an agent might face when they can affect humanity’s long-run prospects in some manner that is extremely persistent. Perhaps it applies when a political activist decides whether to campaign for a change to political institutions that would foreseeably improve decision-making. Doing so may make it ever so slightly more likely that humanity at large has better political institutions indefinitely far into the future, perhaps increasing the probability that future lives (and clusters of civilisation) have positive value on average.

Scenario 3: Improving the Future

Again, expected value theory alone cannot compare the two. Nor can it say that Campaign is better than Don’t Campaign—it cannot say that it is better to make it more likely that future lives are very good and less likely that they are very bad. Intuition demands that Campaign be ranked as better than Don’t Campaign.29

The fourth problem case, Reducing Extinction Risk, is that which an agent faces when they can affect humanity’s probability of long-term survival (and, a fortiori, the probability of morally valuable life surviving). If the agent does nothing, humanity will have some probability of surviving to t and beyond. If they intervene, humanity will have a greater probability of doing so. For my purposes, both options can be represented by some mixture of a low-value outcome (which, for simplicity, we can set to value 0) and the prospect obtained conditional on surviving the near term. For our purposes, those options are equivalent to the following.

Scenario 4: Reducing Extinction Risk

Here, both options are equivalent to having some probability of playing the Aquila game or Skewed Aquila game (with such and such parameters), with Intervene giving the higher probability. But, again, expected value theory cannot compare any two options fitting these descriptions. Expected value theory cannot say that Intervene is better; it cannot say reducing the risk of extinction is an improvement. Again, this is troubling.

Take the Aquila-game versions of Intervene and Do Nothing—each gives a probability of Aquila and a probability of value 0. It seems all the more troubling for expected value theory to say nothing in this case, given that the correct verdict may seem obvious. Both options seem equally good, in so far as expected value theory is plausible in the first place. After all, the Aquila game’s distribution is symmetric about 0. For any value v, it has the same probability (density) of v and –v. These should cancel out. According to expected value theory, they would do if we were dealing with a prospect whose expectation were defined. To uphold the spirit of expected value theory, they must cancel out for the Aquila game too, such that the game is valued at 0. What, then, of Intervene and Do Nothing? Each is a mixture of that option valued at 0 and a further outcome with the same value. They should each be valued at 0 overall, and so be equally good.

Or consider the Skewed-Aquila-game versions of the two options. If the Skewed Aquila game in question is skewed in the positive direction, then it can be obtained from the Aquila game by shifting probabilities such that it is more likely that future lives are very good and less likely that they are very bad. This is clearly an improvement, and so must be better than obtaining value 0. Then, Intervene would surely be better than Do Nothing—between the Skewed Aquila game and the outcome of value 0, it provides the higher probability of the better one. Or, if the Skewed Aquila game in question is skewed in the negative direction, it can be obtained from the Aquila game by making it more likely that future lives are very bad and less likely that they are very good. This is clearly worse than the original Aquila game, and so must be worse than an outcome of value 0. Then, Do Nothing must be better than Intervene—between the Skewed Aquila game and the outcome of value 0, it provides the higher probability of the better one.30

The fifth and most challenging problem case, Multifarious Changes, is a combination of the previous three. The agent is not merely improving/worsening the present with certainty, nor changing the probability of human extinction before t, nor changing the probability of a good future conditional on survival. They have all three effects, or any subset of them, at once. This, I think, is a more realistic representation of many of our options. For instance, attempts to improve the long-term future often have some moral cost in the present—e.g., the opportunity cost of spending one’s resources on lobbying for institutional change is that the same resources aren’t directly used to help the poor. Or, when attempting to reduce the risk of extinction, there is often a further effect on the well-being of future people in the event of survival—e.g., implementing some measure to reduce the incidence of deadly pandemics not only reduces the risk of extinction, but likely also causes future people to experience fewer pandemics in general, whether or not they rise to the level of threatening extinction. Likewise, attempting to make future lives better conditional on survival will often affect the probability of extinction—e.g., if one succeeds in changing political institutions to better respond to the public’s interests, those institutions would then likely also be better at responding to threats of extinction.

If an agent has any options that differ by at least two of those effects at once then, for my purposes, we can model their decision as follows. (Recall that a1a2 represents the skew of the Skewed Aquila game—the greater the fraction, the greater the game’s skew towards positive value.

Scenario 5: Multifarious Changes

From above, a fortiori, we know that expected value cannot compare (at least some) options fitting these descriptions. But nor, it turns out, can it compare any such options—any two such options will defy expectations. And again, this silence is troubling. It is not merely troubling because the correct ranking of the options is intuitively obvious; the correct ranking often won’t be. But it is troubling that our normative theories may fall silent in a decision that we plausibly face in practice. If an agent ever has the opportunity to influence humanity’s long-term future, it is plausible that they face this scenario, and they need guidance. For a decision theory to be plausible, it must offer such guidance in at least the cases we actually face in practice. But expected value theory cannot.

Given its failure in all five problem cases above, expected value theory alone cannot be the correct theory of instrumental moral betterness. If it were, no option ever available to us would be better than (or even comparable to) any other. And that would be absurd.

In later sections of this paper, I will argue that this absurdity can be avoided without rejecting the verdicts of expected value theory altogether—that the theory can be extended to deal with the problem cases raised above. But, before that, it’s worth briefly considering an alternative solution. That solution is to reject expected value theory altogether, not in favour of some extension of it, but rejecting even the verdicts it makes in less troublesome cases. In its place, we could adopt a risk-sensitive decision theory. This would allow us to avoid absurdity in the above problem cases, as I will show in this section. In effect, the discussion up to this point might be seen as forming a surprising argument in favour of risk sensitivity.

To illustrate how risk sensitivity avoids absurdity in the above cases, consider one theory that exhibits it: expected utility theory (specifically, a risk-sensitive version of it). This theory works much like expected value theory does. Where expected value theory says that the best options are those with the highest expected moral value, expected utility theory says that the best options are those with the highest expected utility.

What is utility? For my purposes, it is some representation of the betterness ranking over outcomes. But it need not be the same representation as the moral value function. Utility here is not the same thing as what moral theorists sometimes call utility—a cardinal measure of total welfare—but instead a purely decision-theoretic construct.31

In general, the utility of an outcome may be any increasing real-valued function of its moral value (at least when determining instrumental moral betterness), risk-sensitive or not, so long as that function is strictly increasing. In particular, the correct utility function for use in moral decisions might sometimes be concave: the higher the value of outcomes, the less their utility increases for each additional unit of value that is added to them. This tends to lead to risk-averse preferences. And/or the utility function may sometimes be convex: the higher the value of the outcomes, the more their utility increases for each additional unit of value. This tends to lead to risk-inclined preferences. One possible function, u(v), that is sometimes concave and sometimes convex is plotted below.

But how does switching from expected value theory to expected utility theory, with a non-linear utility function, affect our comparisons of expectation-defying options? To see how, note that such options posed a problem for expected value theory only because the probability densities of outcomes didn’t approach zero quickly enough as value approaches positive and negative infinity. If those extreme outcomes just had lower (perhaps much lower) absolute values, the options would no longer defy expectations, and expected value theory could evaluate them. But, in effect, they do have lower absolute “values” if we switch to expected utility theory with a utility function like that plotted above—we lower the contribution that those extreme outcomes make to the expected utility calculation. Then, for the purpose of calculating expected utility, an expectation-defying option no longer defies expectations!

Take, for example, the Aquila game. Its troublesome distribution was given by p(v)=ab+|v|v (for some a,b>0). With a utility function that is concave enough for large positive values and convex enough for large negative values, we can turn that expectation-defying distribution over value into a much tamer distribution over utility. For instance, set u(v)=|v| such as that given by p(u)=4a|u|2b+|u|3 (which gives an expected utility of 0).

This works in all five of the problem cases described above. In the first (No Change), where we must compare the Aquila game to an identical prospect, a utility function as above lets us say that the two options are equally good—each gets an expected utility and, since the prospects are identical, those expected utilities are equal. So, not only can expected utility theory say something here, but it says the intuitively correct thing. In the second case (Improving the Present), expected utility theory with a utility function as above lets us say that the Aquila game sweetened by value s is indeed better than the same Aquila game without it. In the third case (Improving the Future), it says that increasing the Skewed Aquila game’s skew in the positive direction is indeed an improvement. In the fourth case (Reducing Extinction Risk), where we compare two mixtures of the (Skewed) Aquila game, expected utility theory can again provide a comparison (although what it says will depend on the exact utility function). And, in the fifth (Multifarious Changes), again, it can compare any (perhaps sweetened mixture of) one Skewed Aquila game to another. In all five cases, it satisfies the desiderata I gave above.

And we can do the same with any pair of expectation-defying options; we need only adopt a utility function that is concave (convex) enough for large positive (negative) values. We need only accept a certain sensitivity to risk and the problem is solved. Thus, expected utility theory can deliver verdicts in those scenarios where expected value theory did not.32

And, thus, we seem to have a new and surprising argument in favour of risk sensitivity. Risk-sensitive decision theories can be compatible with Impartiality, Additivity, and the various empirical claims above, while expected value theory alone is not. Following this argument, perhaps we must reject expected value theory in favour of risk sensitivity.

But whether this argument succeeds depends on whether this is the only way to deal with those problem cases. Ultimately, it turns out that risk sensitivity is unnecessary to solve the problem. As I show below, we can extend expected value theory to deal with those problem cases. But it is worth keeping in mind what is at stake here: the result not only exonerates expected value theory from being incompatible with Impartiality and Additivity; it also undermines an otherwise compelling argument in favour of alternative theories that accommodate risk sensitivity.33

Can we deal with prospects like the Aquila game without embracing risk sensitivity? Rather than rejecting expected value theory altogether, can we extend the theory to deal with such problem cases? And, if so, how?

In this section, I consider several possible extensions. As we will see, many of the extensions so far proposed in the literature cannot deal with a prospect as troublesome as the Aquila game. But some can.34

As demonstrated above, expected value theory can be extended to deal with the problem cases from earlier—even if our prospects for the total value of the future are described by versions of the (Skewed) Aquila game, we can still compare the options available to us. But the comparisons we reach are perhaps surprising.

Recall Multifarious Changes. In that case, our options may be any two versions of the Skewed Aquila game, sweetened by some value s and/or mixed with an outcome of value 0. Specifically, consider a version of the case where we must choose between a) sweetening the game by some large amount s, and b) slightly increasing the positive skew of the Skewed Aquila game and/or slightly increasing its probability. According to Invariant Value Theory*, the latter will always be better (provided that it is skewed in the positive direction).44 And this holds no matter how large the sweetener in (a) is, and no matter how slight the changes in probability in (b). In effect, changes of the sort made in (b) are infinitely more valuable than any finite sweetening as in (a). This is perhaps surprising and counterintuitive. If the Skewed Aquila game does describe our prospects over the total moral value of the future, then some interventions focused on the long-term future will be, in effect, infinitely valuable. Increasing the probability that future lives are positive and/or get lived at all will always be more valuable than improving (finite numbers of) present lives. And this will still hold no matter how small those changes in probability, and no matter how many present lives you might otherwise improve.

Notably, this implication is not peculiar to Invariant Value Theory*. It is also implied by the other extant extensions of expected value theory that can deal with each of these cases (i.e., those of Easwaran 2014b; Meacham 2019). And this is to be expected—the differences in the probability distributions of (a) and (b) are roughly analogous to the St Petersburg game, which any genuinely risk-neutral theory must say is better than any finite value (see Hájek & Nover 2006: 706). I suspect that no faithful extension of expected value theory will be able to avoid saying that it is always more valuable to increase the skew or probability of a (positively) Skewed Aquila game than to gain any finite value with certainty.

Even if the Skewed Aquila game doesn’t accurately describe our options in practice, we may still encounter a similar implication. Invariant Value Theory* says much the same for analogous changes to any probability distribution with undefined expected value, so long as it can evaluate those changes at all. (So, I suspect, will other faithful extensions of expected value theory.) So we might accept a model of the future very different from the one from earlier that gave us the Skewed Aquila game, or we might assign only a small probability to the earlier model alongside many others. And still, if the resulting probability distributions have undefined expected values (and if Invariant Value Theory* can compare them at all), then much the same will hold—it will be more valuable to increase the probability of future lives overall being positive or, if they are positive, the probability of future lives being lived at all than to improve (finite numbers of) present lives. Admittedly, this result is only suggestive—there might be a single correct model of the future which gives a defined expected value or, more plausibly, the collection of plausible models might result in probability distributions so troublesome that even Invariant Value Theory* cannot compare them. In either such case, the lessons drawn here from the Skewed Aquila game would have no bearing on practical decision-making. But, if neither is the case, we have quite radical implications for how to evaluate actions that have a small probability of greatly altering the future.

This discussion started with three seemingly plausible normative claims: Impartiality, Additivity, and expected value theory. In practice, are these claims compatible? Or, in practice, do they lead to absurdity?

As I have argued, there is some reason to think that the total expected moral value of the future is undefined. There is at least one plausible model of morally valuable events in the distant future that, if we accept Impartiality and Additivity, gives a probability distribution (the Aquila game or Skewed Aquila game) over moral value that has undefined expectation. Assign any non-zero probability to this model and, no matter what other models we might consider nor what we might expect to happen in the near-term future, our overall prospects for total moral value will inherit that undefined expected value. And this is bad news for expected value theory. If it alone were the correct decision theory, then no option ever available to us in practice would ever be morally better than any other. And this would be absurd.

One possible response to this absurdity is to abandon the verdicts of expected value theory altogether, in favour of some alternative theory that exhibits risk sensitivity. As demonstrated above, by doing so, we can effectively turn any expectation-defying option into a better-behaved one. If this is the only way to avoid absurdity, while holding onto Impartiality and Additivity, we may have a surprising argument in favour of risk sensitivity. But is this the only possible solution? Or can we preserve the risk-neutral verdicts of expected value theory somehow?

It turns out that we can—risk sensitivity may be unnecessary. We can extend expected value theory to deal with the expectation-defying options described here. Admittedly, not just any old extension of expected value theory will do—some proposals are insufficient (e.g., Relative Expectation Theory and Weak Expectation Theory). But other proposals do better, including Invariant Value Theory(*). As demonstrated above, with such an extension, we can deliver comparisons even in those various problem cases involving the (Skewed) Aquila game. (Doing so also brings on some surprising implications, as described in the previous section.)

Does this mean that Impartiality, Additivity, and (some extension of) expected value theory are perfectly compatible in practice; that we need not accept risk sensitivity? Maybe; maybe not. If the five problem cases given above accurately describe the decisions we face in practice, then yes. Even if they loosely describe our real-world decisions—if our real-world options have probability distributions that behave sufficiently like the (Skewed) Aquila game—then the answer is yes. But our real-world options may also be far more complicated. For instance, the model of the distant future described above is just one possible model. There may be far more complicated models to which we should assign some probability. And, if those models give us even more challenging probability distributions, perhaps neither Invariant Value Theory(*) nor even further extensions of expected value theory will be able to compare the options we then face. If so, we may have a compelling argument for risk sensitivity once more.

This somewhat limits the conclusions that can be drawn here. We cannot conclude that Impartiality, Additivity, and expected value theory are guaranteed to be compatible in practice. But at least one seemingly troublesome argument against their compatibility has been undermined—on one fairly plausible model of our future, on which they seemed to conflict, they have been shown to cohere perfectly well. Perhaps there are other plausible models of the future on which they do still conflict, but this remains to be seen. For now, absent such models being proposed, it seems that we can safely endorse all three principles.