2. Effective altruism in theory and in practice

EA is a broad tent, and different authors have very different views on what commitments it entails. To some extent this is due to the fact that EA is (at least) as much a social movement as a philosophical movement. This diversity of views and practices can make EA an elusive target to criticize. In this section we therefore proceed by defining our target more precisely. Our definition captures an important current in EA’s approach to global poverty – the initial impetus for EA. ^{Footnote 2} In particular, our target definition captures a revisionary approach to charitable giving which departs from prevailing common sense. Therefore, for simplicity, we will refer to our target simply as EA. In the conclusion we return to the question what our arguments entail for the broader EA movement.

A good starting point for defining our target is the definition used by the Centre for Effective Altruism (2021): ‘effective altruism is about using evidence and reason to figure out how to benefit others as much as possible, and taking action on that basis’. MacAskill unpacks this definition which is aimed for a general audience and highlights three features. The definition is non-normative, it is science-aligned, and it is maximizing. MacAskill (Reference MacAskill, Greaves and Pummer2019: 12–14) adds the requirements that EA is impartial – which we accept – and welfarist – which we do not include in our definition.

Let us take each of the components in turn. By non-normative MacAskill (Reference MacAskill, Greaves and Pummer2019: 15–17) means that EA asserts neither a moral obligation to give anything to charity nor a moral obligation to give effectively. While this makes EA a broad tent, we fear that it compromises too much of EA’s critical character. EA started out as a challenge to common-sense approaches and common sense plainly accepts that one can set oneself the personal project to give effectively. We propose instead a weak characterization of EA which retains some normative core. Following work by Pummer (Reference Pummer2016) and Horton (Reference Horton2017), we remain agnostic on the question of how much EA requires us to donate – if at all! All we require of EA is that it implies that we ought to donate effectively if we donate. This minimal requirement does not rule out stronger versions of EA that impose demanding unconditional obligations to give. ^{Footnote 3} Our concern is not with how much people should donate, but with whether the assessment methods employed by effective altruists provide good guidance as to where they should donate if they choose to donate.

Likewise, to avoid ruffling any feathers, we wish to remain agnostic about what is to count as good. The utilitarian roots of EA have led many effective altruists to embrace both a welfarist understanding of goodness and to embrace a view that ignores distributive questions. We do not take welfarism to be central to EA, indeed we believe that welfarism would be harmful in building the broad tent EA wants to build. So justice, freedom, virtue, biotic health, etc. are all on the table as possible final ends to be promoted. Our criticism in this paper is therefore largely independent from the question of what counts as ‘good’.

This leaves us with the two components of the definition which MacAskill (Reference MacAskill, Greaves and Pummer2019: 14–15) describes as uncontroversial: science-alignment and maximization. EA is science aligned in the sense that it (proudly) locates itself within the ‘evidence-based’ movement in social science, and insists (plausibly, given the high stakes of aid and development) that it is morally irresponsible to donate to some charity if you lack solid evidence that it is effective. ‘Solid evidence’ here indicates a high evidentiary threshold. In particular, claims about impact size not to be robustly established by studies capable of controlling for selection bias. In practice this usually means EA won’t recommend a charity unless randomized controlled trials (RCTs) have demonstrated its effectiveness (see GiveWell’s 2016 a stated procedure on evidence assessment). There are important exceptions – GiveWell (2013) briefly listed VillageReach among its top charities on the basis of convincing but non-randomized studies – but not many: all of GiveWell’s top nine charity recommendations are supported by multiple RCTs. This is very much in keeping with a strain of thinking within the evidence-based movement which views RCTs as the gold standard for establishing the causal efficacy of some intervention and estimating its average impact size (Banerjee and Duflo Reference Banerjee and Duflo2011; Karlan and Appel Reference Karlan and Appel2011; Leigh Reference Leigh2018), ^{Footnote 4} on the grounds that RCTs control for all confounding factors and therefore register only the causal influence of the intervention (see, however, Deaton and Cartwright Reference Deaton and Cartwright2018 for criticism of this assumption). On top of high-quality micro studies, charity evaluators also insist on high-quality regression analyses measuring the macro-level effect of programmes on variables of interest (e.g. poverty levels, health outcomes) as interventions found to be promising by controlled studies are scaled up and performed in different locations. This is needed to establish that the observed causal relation generalizes beyond the initial context, as sometimes a locally effective intervention becomes ineffective when scaled up (e.g. providing better education to individual students is good for them, but expanding education for everyone will have lesser effects due to the fact that education is partly a positional good; if education mainly helps in securing a government job, increased education won’t increase the amount of government jobs). These are heavy evidentiary burdens to discharge, and unsurprisingly they rule out most charities.

The maximization criterion is straightforward: in deciding where to donate, it is not enough to do some good. Rather, ‘we should do the most good we can’ (Singer Reference Singer2015: vii). More specifically, we should donate only to the most cost-effective charities, i.e. ‘the one[s] that help[] the most people the greatest amount per dollar’ (Alexander Reference Alexander2013). Once we’ve eliminated from consideration all charities for which we lack solid evidence of effectiveness, we rank those that remain by how much value (as measured by improvements on health, poverty, etc.) we can expect them to produce from additional funding.

Strictly speaking, we’re looking to compare the expected marginal rates of return on additional donations of each charity, and keep only those charities that have the highest expected return. And in practice, effective altruists have followed a very simple heuristic for measuring expected rates of return, which we will refer to as ‘myopic marginalism’. Here is one way. Start by calculating the past rate of return on donations. This is easy enough: simply divide the total size of the benefit generated by some intervention by the total cost of the programme. This first measure is a bit crude, since it only tells you about average return on donation, not the return on the last dollar, but it is used by EAs and it does tell you something (see MacAskill Reference MacAskill2015; Open Philanthropy Project 2017; GiveWell 2020 b; Giving What We Can 2021; and especially GiveWell’s 2021 explicit cost-effectiveness calculations in spreadsheets). A more sophisticated measure becomes possible if you have a time-series plotting the evolution of the programme’s costs and benefits: instead of looking at total costs and benefits, look at the ratio of the most recent change in size of the benefits to the change in the costs of the programme. This measure does give you the marginal return on the last dollar spent (see Budolfson and Spears Reference Budolfson, Spears, Greaves and Pummer2019 for further discussion). It is then predicted that the rate of return on the next dollar you donate to some organization will be very similar to the rate of return on the last dollar, up to however much more room the charity has for additional funding. With this information in hand, it is child’s play to identify the elite group of charities that will maximize the impact of the next dollar you donate (up to however much additional room for funding each charity has). And so, in just two easy steps, we’ve winnowed the space of charities worth considering donating to to just a handful, greatly simplifying the decision problems of donors.

There are many steps in the decision procedure we’ve described that one might take issue with. Critics of EA have, for example, criticized the optimizing logic of EA, its over-reliance and over-insistence on RCTs, and its use of cost-effectiveness analysis, which makes no room for permissible partiality and is claimed to overweigh the value of a statistical life. We take issue with none of this in the present paper, and will focus only on the inadequacy of myopic marginalism. As we will now see, myopic marginalism only yields accurate estimates of cost-effectiveness when the benefits of an intervention are continuous in its scale, because only then can we use past returns as a reliable guide to future returns. This would not be a problem if most or all cases were of this kind, but as we’ll show there are many cases where this does not hold.

3. Marginalism and the problem of lumpiness

To get an intuitive handle on the problem, consider the following example. Suppose a farmer, Mnemosyne, needs to move a lump of hardened magma, so that she can till her land. The lump is too heavy for one person to lift on their own, but two people are just strong enough to move it the necessary distance. So Mnemosyne calls her wife, Scathach, and together they move the lump. In this example, what is the marginal productivity of each farmer? As the first worker on the scene, Mnemosyne’s marginal product is zero, and as the second worker on the scene, Scathach’s marginal product is one lifted lump. But it is not as though Scathach is a better worker than her wife: both were necessary to the lifting of the lump. Now suppose Mnemosyne has a lot of lumps lying around, and wants to hire workers to lift them. If Mnemosyne were a myopic marginalist, she would erroneously conclude from the fact that the first worker has a null marginal productivity that the next worker will also have a null marginal productivity. She would then never hire any workers at all, since she expects no output from any number of workers, and leave the lumps to litter her field. The myopia in myopic marginalism is apparent: when worker productivity is ‘lumpy’ and instead of continuous, it is not the marginal productivity of each worker that you need to look at; instead, you need to reason in terms of how productive variously sized worker coalitions would be, then hire the coalition which gives you best value for money. This will be the coalition whose ‘Shapley value’ (see Hart Reference Hart, Eatwell, Milgate and Newman1989) maximizes Mnemosyne’s payoff. Roughly speaking, for any given coalition of individuals engaged in a cooperative enterprise, the Shapley value assigns to each individual the share of the gains of cooperation that corresponds to their expected contribution in a random ordering of all the individuals involved. So in the case of Scathach and Mnemosyne, each receives an equal Shapley value, because if you average their marginal contributions to lump-lifting over all the number of ways in which the coalition could form (namely, Mnemosyne arrives first, or Scathach arrives first) you would arrive at the same number.

In the context of charitable giving, we may put this point sharply in perspective by considering the incentive structures that support the reproduction of oppressive cultural norms. Numerous oppressive social practices – e.g. FGM, child labour and (formerly) footbinding – persist due to coordination problems: roughly, so long as other members of society participate in the practice, it is in your interest to do so as well, but everyone would be better off if everyone stopped participating in the practice (see Mackie (Reference Mackie1996) on the issue of footbinding and FGM, and see Todaro and Smith (Reference Todaro and Smith2015: Ch. 8) and Basu (Reference Basu1999) on the issue of child labour). This is what is sometimes called the social conformity trap (Andreoni et al. Reference Andreoni, Nikiforakis and Siegenthaler2017), and it is a characteristic feature of a certain class of games where there are multiple Nash equilibria in pure strategy, such as the so-called stag hunt game (Cooper Reference Cooper1999: Ch. 1). Note, however, that the more people cease conforming, the smaller the benefits of conformity, and the smaller the costs of non-conformity. Eventually, there comes a point where defection becomes sufficiently widespread that the expected benefits of conformity over non-conformity taper down to zero, and as soon as that point is reached, everyone who has not yet defected will instantly defect, pushing us towards the optimum equilibrium. This is exactly what happened in the case of footbinding: a practice that had lasted for a thousand years ended within a generation as soon as a large enough minority stopped doing it (Mackie Reference Mackie1996: 1006). Crucially for us, the existence of multiple equilibria entails a discontinuous relation between the scale of interventions aimed at breaking social conformity traps and the benefits they provide.

To see how this all works, it may be helpful to examine one example in detail. Take the practice of female genital mutilation (FGM). The worst kind of FGM is infibulation, in which the clitoris, labia minora, and outer walls of the labia majora are removed (Mackie Reference Mackie1996: 1002). It is usually performed by adult women on girls around 8 years old, and it causes serious lifelong health problems. Urination and menstruation become painful and difficult, infections are common, as are stones in the urethra and bladder (Nussbaum Reference Nussbaum1999: 120); subsequent to the operation sexual intercourse becomes traumatically painful, and childbirth requires introcision and subsequent resewing of the genital area (Mackie Reference Mackie1996: 1003). And yet, FGM persists due to cultural norms: men will refuse to marry non-infibulated women, and women have no prospects outside of marriage (Mackie Reference Mackie1996: 1004). Moreover it is believed that being infibulated is healthier, cleaner and more beautiful than not being infibulated, and infibulation is thought medically necessary.

Incredibly, these norms make it worse to be non-infibulated so long as everyone else is infibulated, because in these communities marriage is so crucial to women’s prospects that it is worse to be healthy but celibate than to be married and in pain. And this is the key point: the only reason it is better to be infibulated in these communities is that being the only non-mutilated woman is competitively disadvantageous: if mothers stopped infibulating their daughters, the latter would suffer no competitive disadvantage with regards to their peers when looking for a husband, and then all could marry without suffering excruciating pain. (The same holds, of course, too if prospective husbands stopped preferring infibulated partners over non-infibulated partners.) And presumably no woman would then want to be the only one to undergo FGM, because that would be to choose marriage and pain over painless marriage (notice that this implies that it is better for those who do conform that sufficiently many others conform too, because only then is undergoing FGM advantageous in any way).

We can model formally the payoffs associated with FGM. In effect, women in communities which practice FGM have exactly two choices; call these ‘Don’t Conform’ and ‘Conform’. If all women choose Don’t Conform then they obtain the best possible payoff. In contrast, if all women choose Conform then they get a worse payoff, but it is still the second-best they can get away with. However, when only a few women choose ‘Don’t Conform’, the women who choose ‘Conform’ get comparatively better payoffs, whereas the women who choose ‘Don’t Conform’ are punished for departing from the cultural norm, and obtain a payoff that is significantly worse than what they would have obtained had they chosen ‘Conform’. This is the characteristic payoff structure of a close variant of what is known as the stag hunt game. Formally, in a simple setting where there are only two players in the game, the payoffs associated with each player’s choice in each possible setting can be represented by Table 1, where the numbers represent the utilities associated with each payoff.

Table 1.

The Social Conformity Trap

What’s important to realize here is that if each player believes the other will conform to the social practice, the best strategy is to conform as well. And if conformity is a stringently enforced social expectation, if everyone in your community has adhered to these standards in the past, then the reasonable belief to have is that everyone will conform. This is how people can get stuck in inferior equilibria. And indeed, Table 1 characterizes the payoff structure not just of FGM, but also of domestic servitude of women under patriarchy, beauty norms, religious conformity, and many other oppressive cultural practices. Note that if we generalize the above matrix to a situation with N players, then the benefits of either strategy will be increasing in the number of players who adopt that strategy: the more people conform, the costlier it is to not conform, but the more people defect, the less costly it is. And for any plausible model of an N-player coordination game of this sort, there will be some number n < N such that not conforming is the optimal strategy whenever you expect at least n other players not to conform.

How does the existence of multiple equilibria ‘lumpify’ the benefits of breaking social conformity traps? Consider that the first few defectors will not, on balance, benefit from it; in fact, short of potentially expensive compensating offsets, they will be strictly worse off for taking their moral stand. But the social benefit of convincing the nth woman to boycott the practice, whose defection pushes other people’s expectations to the tipping point, are immense, because this will rapidly cascade into the defeat of the social practice, hugely improving everyone’s welfare, and greatly improving the state of justice, freedom and equality. In other words, the marginal benefit of convincing the first n−1 players to defect is in every case negative (at least in terms of welfare), and the marginal benefit of convincing the nth player to defect is the sum total of all the benefits yielded by the ending of the practice. This problem generalizes. Tipping points aren’t limited to oppressive social practices, but are a characteristic feature of all games with multiple equilibria.

Any charity whose work involves pushing people from a lower to a higher equilibrium will generate lumpy benefits. And many charities’ work does involve this, as many are engaged in the project of challenging oppressive cultural norms. This is at the heart of Girls not Brides’ and Amnesty International’s work, for example. This is a feature that those charities share with many activist movements: progressive activism does tremendous good when it effects substantive political change (e.g. the civil rights movement), but very little if it fails (e.g. the Occupy movement), and meanwhile, whether or not it succeeds, mass activism consumes considerable time and resources.

For an effective altruist charity evaluator, this presents a conundrum: what is the marginal value of donating an extra $10 to an organization such as Girls not Brides or Amnesty International who provide ‘lumpy’ benefits? Or the marginal value of the donation implicit in taking part in activism rather than working? As explained earlier, the effective altruistic method for computing the marginal value of a donation is to look at past returns on donations to various organizations; it is then predicted that the marginal rate of return on the next dollar you donate to some organization will be very similar to the rate of return on the last dollar. This decision procedure is fairly easy to apply, since knowledge of past returns is cheap to procure (assuming that you’ve run your studies and are keeping books, all the data should be in), and NGOs are usually happy to tell charity evaluators how much they can scale up their operations with additional funding. Indeed this is what makes marginalism attractive: it is informationally undemanding yet inferentially powerful. But past returns on investment is only a guide to returns on future investment if the relation between returns and investment is continuous. And in the case of games with multiple equilibria, this crucial presupposition fails: the return on the next dollar you spend may be hugely greater than the return on the last dollar you spent.

And unlike in the case of Mnemosyne and Scathach, here the Shapley value cannot come to our rescue, because we face problems of incomplete information that preclude its computation. Computing Shapley values requires much more knowledge than mere knowledge of past returns on donation: it requires knowledge of the total output of various hypothetical coalitions. Mnemonsyne was only able to identify the worker coalition that gave her best value for money because she knew from the get-go that it takes exactly two workers to lift one lump. But in games with multiple equilibria, we don’t know ahead of time what the magical value of n is which pushes expectations to the tipping point. All we know is that it lies between one and the total number of players, which is too bare a guide on which to base Shapley value estimates. Heck, in some cases we won’t know ahead of time what game form we’re in, and thus how many possible higher equilibria there are. Simply shifting one’s attention from one individual’s marginal impact to the impact of various donor coalitions is not going to be useful here. ^{Footnote 5}

One can try to guesstimate how many equilibria there are and what n’s value is by looking at past cases of successful collective action to defeat social norms that are similar to the kinds one is looking to overcome, but such guesstimates are unlikely to be very reliable or cheap to procure. For one thing, the historical record on successful attempts at overthrowing oppressive cultural norms may be sparse or non-existent (China is the only country where footbinding was ever practiced, so there would have been no inductive base from which to infer n’s value). Further, the value of n may be quite sensitive to local attitudes and conditions (e.g. payoffs may differ according to religiosity), and likewise the factors that drive expectations of conformity may vary from place to place, which makes it difficult to draw reliable inferences from other cases about ‘what works’ even in cases where something did work. And anyway, estimates are likely to suffer from reference class problems, as judgements about which cases are similar enough to new cases of interest to permit inferences are highly contestable (is footbinding structurally similar enough to FGM as a practice to permit inferences about how large a minority needs to defect for the social practice to fall? Is the practice of FGM in Afghanistan similar enough to the practice of FGM in Eritrea?). In sum, when lumpy benefits are at issue, cost-effectiveness analysis becomes too informationally greedy to yield reliable assessments. This is not to suggest that there are not some rare examples in which we have such information. In some infrastructure projects we have fairly good information about the necessary fixed costs to complete the project. Such projects are lumpy. Building a water treatment plant provides slightly lumpy benefits, because you get no benefits from it at all unless you invest enough to cover its fixed costs. But beyond that point, the relation between variable costs and total benefits is smooth: for every extra day you pay your employees and keep your plant running, you get an equal amount of sanitation. If you know the fixed costs, then it is trivial to extend myopic marginalism to this case. It is notable that the example of the water plant is another case of direct aid. The important goods we described, like breaking social conformity traps, are lumpy all the way down, and they pose more serious problems: there are only discrete tipping points at which large benefits accrue, and no ‘investment floors’ past which further investment or cooperation begins to yield smooth returns. In these cases, you do need to compute a Shapley value to figure out the value of an intervention, and the problems of incomplete information must be squarely faced. Unlike in the case of direct aid investments it is much more complicated to find out where the tipping points are.

4. Doubling down on myopic marginalism

What is the appropriate way to respond to all these observations? One possible response is that if the cost-effectiveness of lumpy-benefit providers such as Girls Not Brides cannot reliably be measured, however cost-effective they might in fact be at any level of funding, then tough luck for child brides and victims of FGM. Being well-intentioned in trying to overcome FGM and child marriage is not enough, we need to show that Girls Not Brides and similar organizations stand a good chance of actually changing things. There are reasons to be sceptical. Many well-intentioned charities do no or very little good in the world. For example, research by the US-based Coalition for Evidence-Based Policy indicates that up to 75% of all social programmes that were tested for effectiveness had no measurable positive effect – though some had measurably harmful ones (GiveWell 2019). All the while the stakes in development and aid are high. Money sent to ineffective charities cannot be used elsewhere. And EA is reminding us that there are charities that we know to be highly effective in saving many, possibly thousands of infants from easily preventable causes of death.

We aren’t terribly impressed by the flat response that ‘we’ve shown effectiveness, while you haven’t’. This response suggests that the failure of some aid programme to demonstrate cost-effectiveness according to EA’s method should make us highly confident that it is in fact less cost-effective than some programme which has demonstrated its effectiveness by those standards. But such high confidence is only warranted if the test employed permits accurate measurements of both programmes’ effects. In general, assume we have a measurement tool that is highly accurate in measuring A-type interventions but poor in measuring B-type interventions, i.e. with a probability close to 1 it would fail to register the effectiveness of B-type interventions even when these are effective. This measurement tool cannot give us a good guidance for comparing A-type interventions with B-type interventions. The judgement that A-type interventions fare better than B-type interventions is not due to anything about the interventions, but an artefact of our assessment methods. Myopic marginalism falls into this trap. Its impact assessment method is highly inaccurate when benefits are lumpy: if some programme is not at tipping point scale, then myopic marginalism will inevitably evaluate it as not cost-effective, regardless of how much more cost-effective it would be at the tipping point scale than any other programme.

Still, you might think that what counts in favour of EA’s recommendations is not that they are demonstrably more cost-effective but rather that we have better information about their cost-effectiveness. Our inability to accurately estimate the cost-effectiveness of lumpy-benefit-providing would then be in itself a good reason to err on the side of caution and only donate to the EA-approved programmes, since we at least know that those ones do a lot of good. This would be an example of what in decision theory is called ‘ambiguity aversion’ or ‘aversion to severe uncertainty’: you have no particular reason for thinking that f-ing is better than y-ing, but your greater uncertainty about the chances of y’s possible outcomes leads you to prefer f-ing. Aversion to severe uncertainty of this sort is explicitly endorsed by GiveWell (2016 b), who say ‘we generally prefer to give where we have strong evidence that donations can do a lot of good rather than where we have weak evidence that donations can do far more good’.

In response, we accept that this is one reason to prefer the EA approved programmes. An agent who is averse to severe uncertainty may, for this reason, systematically select the recommendations of EA rather than interventions that provide lumpy benefits. This suggestion is noteworthy since it goes against the grain of the utilitarian origins of EA, for utilitarianism is traditionally not sympathetic to ambiguity aversion. This is for two simple reasons: first, one of the strongest and most widely cited arguments for utilitarianism are the so-called ‘impartial observer’ theorems by John Harsanyi and later economists (Harsanyi Reference Harsanyi1955; Hammond Reference Hammond1981; Fishburn Reference Fishburn1984; Broome Reference Broome1991), which purport to show that any person who is impartial, sympathetic and rational must rank alternative courses of action by utilitarian standards. But crucially the notion of rationality required to obtain these results rule out ambiguity-aversion on the part of the social evaluator. Second, we can show that if a utilitarian is ambiguity-averse with respect to individual prospects (i.e. they prefer to give individuals unambiguous benefits to ambiguous ones), then utilitarians will sometimes have to choose between one course of action which is guaranteed to yield greater total well-being, and another which is guaranteed to give each individual a better (because less ambiguous) prospect – so either utilitarians must make everyone worse off in expectation, or they must fail to maximize total well-being (Rowe and Voorhoeve Reference Rowe and Voorhoeve2018: 262–265). Ambiguity-neutral utilitarianism, by contrast, will never face this choice: maximizing total well-being is always consistent with choosing what is expected to be best for everyone.

In any case, the appeal to aversion to severe uncertainty does not help the case of EA. EA wants to say that only a small number of charities are donation-worthy. Its power of discriminating among the vast number of charities in the world is its greatest asset. Aversion to severe uncertainty is, however, at best rationally permissible. To yield the prescriptions EA makes for the charitable giving of others, the proponent of EA would need to argue the substantially stronger claim that one is rationally required to be averse to severe uncertainty. Otherwise, EA cannot tell an agent who is neutral to severe uncertainty that they are making a mistake in donating to lumpy-benefiting-providing organizations. Nor is aversion to severe uncertainty somehow constitutive of caring about giving effectively: provided your attitude to severe uncertainty is rationally permissible, what counts as ‘the’ effective intervention sensitively depends on just what your attitude is. If you’re permissibly averse, the effective intervention is intervention that strikes the best balance between maximizing expected benefits and reducing severe uncertainty, while if you’re permissibly neutral the effective intervention is the intervention that simply maximizes expected benefits. We take it that both aversion and neutrality with respect to severe uncertainty are rationally permissible attitudes, so neither can be uniquely constitutive of caring about effectiveness. The bottom line is that effectiveness is not linked to any particular attitude towards severe uncertainty.

Second, the argument from severe uncertainty would have more force if EA approved aid programmes and programmes which provide lumpy benefits were all in the business of providing the same sorts of benefits. Ambiguity aversion is most sensible in cases such as the Ellsberg paradox (Ellsberg Reference Ellsberg1961), where you have a choice between an ambiguous lottery (unknown probabilities) and an unambiguous lottery (known probabilities), but both lotteries return cash prizes. For one thing, if the kinds of goods you have a choice between are so different as to be difficult to compare in value precisely, then even if you are ambiguity averse, your ambiguity aversion may fail to support any determinate ranking among your options. ^{Footnote 6} This contrasts with Ellsberg paradox cases, where any positive level of ambiguity aversion will require you to choose the less ambiguous lottery. But let’s set this point aside.

The appeal of ambiguity aversion is the appeal of ‘playing it safe’. But while we may want to play it safe if we have to select based on poor information when the outcomes are always money, things can look very different if we are talking about taking a chance on poor information to promote goods of justice and liberty versus taking a bet on surer information to promote the rather different goods of health and income. Charitable giving is not just a one-off decision. It is also not just a decision done by a single agent. If followers of EA are convinced that they should be averse to severe uncertainty, then they would all systematically choose direct aid programmes designed to improve health and income. Other values like justice and liberty would, routinely, get the short end of the stick. This makes the case of severe uncertainty aversion in EA different from that of a simple Ellsberg paradox. In a simple Ellsberg paradox, there is just a single choice in which we might want to play it safe. But playing it safe over and over again, all the while there is a bias in what the safe option is, is less compelling. Here the appeal of playing it safe competes with the appeal of promoting under-served values.

It is relevant in this context that EA-approved interventions are (almost) all providers of direct aid, i.e. private goods such as income, medications and bednets, whose consumption is linked to improvements in consumption levels or health. ^{Footnote 7} Direct aid is favoured by EA in large part because direct act typically has continuous marginal benefits. Providing one bednet is not too different from providing two bednets, and so on. The improvements brought about by direct aid obviously matter, and are always a major aim of interventions aimed at ending oppressive cultural practices – the best reason for ending infibulation is that it is traumatically painful and cruelly impairs women’s lifetime health. But there are special reasons to care about ending oppression that go beyond their immediate benefits in terms of health and income: oppressive social practices such as FGM and child marriage are high crimes against their victims, not mere misfortunes, like malaria; they crush individual liberty, and create classes of dominators and dominated. We have compelling interests in halting high crimes, protecting liberty, and undermining structures of domination that are not reducible to our other compelling interests in relieving the burdens of poverty and ill health. (Note that we are not begging the question against welfarist construals of EA. It might be, for example, that liberty and non-domination are both components of or contributors to well-being and so their value is reducible to that of well-being; we just need to deny that it is reducible to the value of health and income.) If these ends can only be advanced by donating to EA non-approved causes, that is a reason to take risks on such causes by donating to them. And not to put too fine a point on it, but the end of infibulation would be a very good thing, just as the end of footbinding was a good thing, and as the end of just any oppressive social practice would be a good thing. So if things go well in your support of some NGO devoted to challenging these practices, things will go very well indeed, and this again gives you reason to take a chance on such causes.

This means that we can have reason to choose interventions with lumpy, discontinuous benefits, precisely because they serve different ends from the direct aid interventions favoured by EA. There is another reason why we may prefer to donate to organizations carrying out such interventions. We earlier mentioned that it is well-known that many charities fail to provide any substantial benefits at all. It is another well-known fact that aid interventions that do a great deal of good in the short term can do immense harm over the long term. Against such a background, one might reasonably be concerned with giving greater weight to ruling out interventions that have a risk of making things worse than making sure to identify interventions that are effective in making things better. Consider an analogy. Refusing to convict criminal defendants unless their guilt is established beyond a reasonable doubt implies that many criminals will walk, but that is appropriate given the stakes in criminal defence trials. Likewise, refusing to recommend donating to some organization on the ground that there is a risk of perpetuating harm implies refusing to recommend many charities that are in fact donation-worthy, but that is appropriate given the stakes in aid and development.

It is worth pointing out that because programmes aimed at undermining oppressive cultural practices are, unlike the GiveWell approved programmes, not in the business of providing direct aid, they don’t raise many of the concerns about harmful effects of direct aid. As de Waal (Reference de Waal1997) has documented, disaster relief aid to war zones, or even countries suffering peacetime famines (e.g. North Korea in the 90s) can only reach their intended recipients with the consent of local warlords and tyrants. Such aid saves many lives, de Waal shows, but Angus Deaton and others have argued that by providing support to oppressive regimes, such aid prolongs political violence and repression, undermining long-term development and costing more lives down the line.

Concerns about the risks of aid are not limited to war zones or authoritarian regimes. Aid can have negative consequences for state capacity. Effective state institutions are extremely important for development, and weak institutions can lead to poverty traps (Sen Reference Sen1999: 111–203; Acemoglu and Robinson Reference Acemoglu and Robinson2012; Deaton Reference Deaton2013: 267–324). In countries with weak institutions (which are often major recipients of aid), the provision of aid can lead to a substitution effect: when NGOs provide goods and services that the state could itself deliver through taxation, the state can scale back its own efforts to provide these goods (or doesn’t build up the capacity to deliver them), so the short-term effect of the intervention is a wash, and its long-term effect is seriously negative because it undermines state institutions (Wenar Reference Wenar, Illingworth, Pogge and Wenar2011: 115, 125–126; Deaton Reference Deaton2013: 292–294; Acemoglu Reference Acemoglu2015; Clough Reference Clough2015). Substitution effects are a real concern for direct aid programmes such as the ones recommended by GiveWell (acknowledged today by most NGOs and charity evaluators), since they do provide goods which the government also provides. But they’re not a concern for charities that challenge social conformity traps, because reforming oppressive norms and practices is different. Either governments of developing countries have no interest in providing them, or if they are interested, they often need the support of civil society organizations. Building an NGO-run school can crowd out public investment in education, but advocating alongside the state against child marriage is more likely to be mutually reinforcing. It is somewhat ironic, therefore, that the appeal of EA’s focus on direct aid lay in a form of playing it safe (aversion to severe uncertainty) given that a different way of caring about playing it safe (aversion to causing harm) counsels in the opposite direction.