This competition entry has been selected for publication by the Forum team.
Effective altruism, as it is usually understood, is a commitment to taking the most effective available means to doing the most (impartial) good. This commitment is difficult to meet if we cannot tell which means would be the most effective one to doing the most (impartial) good.
Often, the “most effective means” is taken to be synonymous with “the strategy that maximises expected value”. Indeed, we find this assumption in DiGiovanni’s P1, the “normative premise”.
P1. Normative premise: To justify c-preferring A over B, it’s not enough to say (e.g.) that A seems heuristically good. Rather, we need to argue that A has higher “expected value” broadly speaking, meaning: In some sense we “expect” that, if we were idealized agents who could aggregate all of A’s and B’s possible consequences into literal EVs, then we’d say A has higher EV. Otherwise, it’s unacceptably arbitrary to c-prefer A.
Here, I object to the claim that taking the most effective means to doing good is not synonymous with maximising expected value. In fact, a better way to understand what taking the most effective means to doing (impartial) good means is “aiming to do what a rational agent who desires to do the most impartial good would do. Call this aim “Effective Altruism”:
Effective Altruism: the aim to do what a rational agent who desires to do the most impartial good would do.
So, what effective altruism demands depends to a large extent on what, exactly, it is rational for an agent who aims to do the most impartial good to do. My objection to the Normative Premise is that it is not always rational for an agent who desires the most impartial good to maximise expected value. This is because the complete story of what rationality demands includes additional requirements on rational deliberation and action. Agents who fail to meet these are not rational.
What are these requirements? In short, the requirements of substantive and structural rationality.
Substantive rationality requirements: requirements specifying how a rational agent’s attitudes should be based on reasons and evidence.
Structural rationality requirements: requirements specifying how a rational agent’s attitudes should fit together.
Decision-theory is part of the story. It aims to tell you what it would be rational for you to do, given a subset of your attitudes: your preferences, credences. Another important part of the story is how a rational agent should respond to her reasons and evidence and how a rational agent’s attitudes should fit together. Substantive requirements of rationality govern the former and structural requirements of rationality govern the latter.
Substantive and structural rationality requirements constrain which options it is rational for an agent to choose. So, rational agents should only maximise expected value when this would be compatible with also meeting the substantive and structural rationality requirements. Some options that maximise expected value are not compatible with substantive and structural rationality requirements. The clearest such example is options that have a very low probability of a very high value. Of course, the observation that these kinds of options cause problems for reconciling maximising expected value with our wider philosophical views about the rational agent is not new (see Allais 1953; Nover and Hájek 2004; Quiggin 1982; Schmeidler 1989, Buchak 2013, Monton 2019). What is new and worthy of discussion is the idea, often held or assumed by effective altruists, that maximising expected value could be the complete story of what is it rational for someone who cares about doing the most good to do.
Substantive rationality. Choosing these options violates substantive rationality requirements because it is not (substantively) rational to believe that these options will bring about good, because of what we know about their probability. Our knowledge about the probabilities constrains the credences it is rational to have about an option, and rational credences in turn constrain the coarse-grained attitudes it is rationally permissible to have - that is, what it is rationally permissible to believe about the option. For example, something you have a 0.9 credence in is something you should regard as more likely to be true than something you have a 0.1 credence in, and your coarse-grained epistemic attitudes should reflect this.
In this case, our knowledge about the probability of an option bringing about good (i.e. that the probability is very low) constrains our rational credence which in turn rationally prohibits us from believing that this option will bring about good. Furthermore, the most rational thing to believe about such an option is that it will fail to bring about good.
These constraints on rational choice cannot be fully captured by the claim that rational agents maximize expected value because decision theory has no mechanism for distinguishing between the differences in epistemic significance between different magnitudes of credence. An option with probability 0.01 can come out just as good as an option with probability 0.99, provided it is combined with a large enough value.
Structural rationality. Structural rationality is concerned with how well an agent’s attitudes fit together. It prohibits, for example: contradictory belief, akratic behaviour, and cyclical preferences. Choosing an option that has a very low probability of a very high value violates structural rationality requirements because it does not seem possible for a rational agent to coherently choose these options.
Another example of a violation of these more general coherence requirements is ‘Moorean’ absurdity, captured in assertions such as ‘P, but I don’t believe that P’ or ‘P, but I don’t know that P’. Yet another is having beliefs that one regards as not supported by one’s evidence. These combinations of attitudes are not logically inconsistent, but they nevertheless seem incompatible with having a fully rational first-person perspective. This more general sense of incoherence is intuitively recognizable, but difficult to define precisely. Worsnip, noting this, suggests defining incoherence in the following way:
a set of attitudes is incoherent just if it is constitutive of these attitudes that anyone holding them is disposed to revise at least one of them, under conditions of full transparency. (2021: 133)
In other words, combinations of attitudes are incoherent when we would not happily hold them together. Rational agents, intuitively, do not hold such attitudes. Using this standard analysis of incoherence, we find instances of it if we think that rational agents must always choose the option that maximizes expected value. Specifically, when we imagine agents choosing a low probability, high expected value option such as (g), these agents will sometimes have incoherent combinations of attitudes. Specifically, the desire to do the most good, and the intention to do something that they know is highly unlikely to bring about any good. Someone who makes such a choice would be correct to assert and believe statements such as the following, all of which have a Moorean flavour: ‘I am choosing E, but E is very unlikely to pay off’, ‘I am choosing E, but E is extremely unlikely to bring about any good’, or ‘I am choosing E, and I know that E will achieve nothing’. These assertions seem strange for someone primarily concerned with doing the most good. Nevertheless, they are correct descriptions of the agent’s actions, given the beliefs about their action that they are rational to hold. An ideal effective altruist who chooses such an option will, rationally, hold the following attitudes:
Desire to do the most (actual) good.
Belief that Option X is highly unlikely to bring about any (actual) good and will most likely achieve nothing.
Intention to choose Option X.
But, these attitudes exhibit incoherence. Rational agents concerned to do the most good should not do what they (rationally) believe is highly unlikely to bring about good. At least, assuming they have some other available option. If that is right, then the Normative Premise is false. Expected Value is not the metric by which we should compare options.
However, this does not mean that we should instead favour options that “seem heuristically good”. In the Normative Premise, DiGiovanni contrasts favouring an option because it maximizes expected value with favouring an option because it “seems heuristically good”. Complying with substantive and structural rationality is neither of these things. Heuristics approximate rationality while falling short of what is in fact required by ideal rationality. The substantive and structural requirements of rationality, if they are requirements of rationality, are requirements of ideal rationality, not heuristics. They describe how a rational agent should choose.
At this point, it is worth addressing an objection. Why should effective altruists care about substantive and structural requirements of rationality? Isn’t rationality supposed to “win” (to borrow the enigmatic slogan used by Jesse Clifton and repeated by DiGIovanni)? That is, isn’t it supposed to help us achieve our practical goals.
One might worry that we can easily imagine cases in which it would bring about more good for someone to have incoherent attitudes, or to believe against their evidence, and that this implies that substantive and structural requirements of rationality are not genuine requirements of rationality - they do not always “win”. However, this is a misunderstanding. If an eccentric philanthropist offers to donate a huge sum to a charity of my choice if I believe a contradiction, then I should endeavor to believe a contradiction. This would be practically rational in so far as I desire to do good. But, this is not incoherent. It is a coherent choice given by desires and beliefs - I desire to do good, I believe that I will do good if I believe a contradiction. So, I believe the contradiction. This is importantly different from the cases of structural incoherence exhibited by combinations of a desire to do good and an intention to do something known or rationally believed to be highly unlikely to being about good. The agent who believes the contradiction so that the eccentric philanthropist will donate to charity knowingly takes the best means available to her to bring about good. Thus, her desire to do good is coherent with her intention to take the best available means to doing good. Indeed, she maximizes expected value (which, in this case, is the rational thing to do). However, an agent who desires to do good and is faced with a fanatical option cannot coherently choose the option that maximizes expected value. This is because they rationally ought to believe that choosing the fanatical option would be a failure, or at least be highly likely to be a failure. Unlike the agent who believes the contradiction, it is not at all clear that they would not be knowingly taking the best available means to bringing about good (assuming they have some other option with better prospects available). While effective altruism might sometimes ask us to believe contradictions, it should not ask us to do things that, as rational agents, we should believe will fail to bring about our ends.
What is the upshot of this for cluelessness? The good news is that effective altruists do not need to worry about being unable to calculate the precise expected value of every possible result of their actions. Instead, they should focus on what their evidence in fact supports, what it would be rational to believe on the basis of that evidence, and which choices these beliefs rationalize. The, perhaps, bad news is that this might make the effective altruist project seem less exciting than many of its proponents like to think. In other words, what most of moral philosophy is doing and has been doing for about as long as there has been moral philosophy.
References
Allais, Maurice (1953) ‘Le Comportement de l’Homme Rationnel devant le Risque: Critique des Postulats et Axiomes de l’Ecole Americaine’, Econometrica 21: 503– 546. doi:10.2307/1907921.
Buchak, Lara (2013) Risk and Rationality. Oxford University Press.
Monton, Bradley (2019) ‘How to Avoid Maximizing Expected Utility’, Philosophers’ Imprint 19: 1-25. doi:http://hdl.handle.net/2027/spo.3521354.0019.018
Nover, Harris and Alan Hájek (2004) ‘Vexing Expectations’, Mind 113:237–249.doi:10.1093/mind/113.450.237.
Quiggin, John (1982) ‘A Theory of Anticipated Utility’, Journal of Economic Behavior and Organization 3: 323–343. doi:10.1016/0167-2681(82)90008-7.
Schmeidler, David (1989) ‘Subjective Probability and Expected Utility Without Additivity’, Econometrica 57: 571–589. doi:10.2307/1911053.
Worsnip, Alex (2021) Fitting Things Together: Coherence and the Demands of Structural Rationality. Oxford University Press.
DiGiovanni's very point is that the conflicting pieces of evidence we have do not support any action A over any B, from an impartial perspective.
So how do you come to a non-arbitrary belief about what this rational agent would do?
(I think the crux here for you may in fact not be DiGiovanni's normative premise but his empirical one.)