As far as I can tell, the most common critique of Effective Altruism is something like this: It certainly sounds reasonable enough, but it feels rather suspicious and some of the outputs really seem rather wonky [1]. Of course, those in the movement know that this is due to bias which arises from misplaced confidence in commonsense morality, mere social convention. But I’d like to proceed from first-principals to show that perhaps commonsense morality should not be dismissed quite so easily.
There is a saying which I cannot find at the moment, which says something to the effect that basically all philosophy is just arguing about definitions. According to the introduction to EA, “Effective altruism is a project that aims to find the best ways to help others, and put them into practice”. As a fellow traveler to the movement, I have my own definition, which is two words: applied utilitarianism.
At this point, you might be thinking: “Aha! So, you don’t like utilitarianism, therefore you don’t like EA.” It is true that various flavors of utilitarianism (including the vanilla classical) do result in some strange results at times [2]. But I have my own definition of utilitarianism, the description of which is the secret agenda of this essay.
A Perfect Definition of Utilitarianism: To maximize total utility, where individual utility is the criteria which would be maximized if an omniscient version of an individual could choose a particular universe out of all possible ones, and total utility is the aggregation of individual utility by an omniscient daemon.
In short, utilitarianism seeks the best of all possible worlds, the golden path [3]. I call this a perfect definition because you can use proof of contradiction to show it is obviously correct [4]. However, it is not a very practical definition, as it relies on omniscience, a quality in short supply. What we want is a methodology from which any entity at a specific time and with a specific point of view can make actionable decisions.
The first use of omniscience in the definition is in knowing all potential universes to know the true objective wellbeing of an individual. One way we can approximate this objective wellbeing is by aggregating over time of the subjective wellbeing of an individual over their entire lifespan. This renders utility measurable by query: “How are you doing?”
That being said, omniscience is still required to know information which occurs in the future. In addition, although one can know their own subjective wellbeing with relative certainty, assessing this property in others is typically done with less certainty (especially given time constraints). We need to solve this problem by placing the calculation of utility inside of a probability distribution [5].
This allows us to specify a calculation for total utility which is actionable to any individual who resides in any particular context.
“Practical” Definition: As an individual, attempt to maximize the expectation of the aggregation over time and individuals of the subjective well-being of all individuals.
U(→x)≈E(∫time∫individualsP(→x))
That being said, it should be obvious why “practical” is written here in quotes: this definition is still computationally intractable. In addition, it should be noted that methods to aggregate individual into total utility are left undefined [6].
(Too) Practical Definition:As an individual, attempt to maximize the aggregation over time and individuals of the expected subjective well-being of all individuals, which is equivalent to always choosing the action which leads to the most increase in subjective well-being (under most commonly chosen aggregation functions).
U(→x)≈∫time∫individualsE(P(→x))
However, bringing the expectation function inside of the aggregation functions leads to various well-known paradoxes [2]. Famously, the aggregation of individual utility through mere addition leads to the repugnant conclusion [7]. A more reasonable aggregation function needs to consider that identical conditions may be highly pleasurable to one entity while resulting in negative utility in another, while still also accounting for nth-order effects such as the effect of one entity’s well-being on that of others [8]. As a result, I am reluctant to simplify my definition any further.
As a result, we are left with an objective function which is impossible to actually compute. Fortunately, we have a standard method to deal with intractable computations: heuristics. And I posit that this is exactly what commonsense morality is [9].
Selfishness – since you know your own subjective preferences, caring about your own wellbeing versus that of others is entirely reasonable. You can know exactly how much you will enjoy different experiences, while it is unclear if the enjoyment of others is truly as sublime as your own. So, we eat filet mignon.
Kinship bias – Rather than donating to complete strangers, it may be rational to assist those whom you know well, since you know their preferences and can gauge their reactions with a higher degree of certainty, no matter how much they protest “You shouldn’t have!”
In-group bias—In performing actions which you expect to please others, you ought to do so to people who have traits closer to yourself, since you are able to simulate their reactions with greater fidelity. Hence Scotscare [10].
Local bias – In donating, it may be strategic to donate to those you can interact with directly, as you can receive immediate feedback that the money reached the target and was received positively, whereas donations to far away locations may be lost or misused.
Loss aversion – Since one can easily simulate the negative utility a loss would deal oneself; it may be better to prioritize loss prevention rather than realizing a gain which actually isn’t all that great. As they say, the worst thing about realizing your heart’s desire is exactly that [11].
Short-term bias – It is better to prioritize increasing utility in the short term, since shortening the feedback loop can allow more nimble adjustments, while long-term projects will have much higher variance, which should be discouraged due to loss aversion.
The key difference though, is that commonsense morality actually thinks of itself as the answer, whereas the utilitarian knows it is merely a set of various heuristics seeking to approximate total utility. In which case, the strength of an altruism based on utilitarianism is a recognition that these heuristics are subject to biases and must be constantly examined and improved [12].
It should be uncontroversial that the best and most reliable method we have to rectify biases is through the empirical method with fast iterations [13]. It should also be pretty clear by now that as far as optimization go, the most successful general algorithm we have is gradient descent [14]. Therefore, my opinion is that the best way to go about improving utility is to take advantage of these tools: determine key performance indicators which are readily measurable (differentiable) and highly likely to improve total utility (likely to lead to global optima). We can then focus on most efficiently optimizing these parameters. I believe that this is the aspect of effective altruism which is most well received, and with good reason.
On the other hand, there are other “core tenets” of effective altruism which are receiving increased focus in recent years which are essentially untethered from the empirical approach.
Animal welfare – Regardless of the debate over whether the subjective well-being of non-humans is measurable by humans, my issue with this KPI is that assistance to humans can be paid forward with compounding altruism, while others of the kingdom are typically disinterested in the altruism project at best [15].
Excessive focus on infinities in the future [16]– aside from having unacceptably long feedback loops, our collective track record with predicting even the events of the medium-term future have generally been abysmal, owing to the confounding factors of higher-order effects [17].
AI-risk – A variant of x-risk, but one which is mediated through an unknown number of steps of uncertain probability (as compared to other risks like viral pandemic, nuclear war, or heat death, which directly result in dramatic population loss), and therefore particularly hard to estimate or measure.
These projects which presume to know better than the empirical approach, I perceive as being not even wrong [18]. Worse yet, within these topics they often they get bogged down in debates about details, which in a KPI which cannot even be measured properly can only be regarded as premature optimization. That is not to say these topics are not important: it is simply that whether they are effective or not is indeterminable. I cannot refer to them as effective altruism—only as altruism (and that is totally fine).
Footnotes
[1] For examples of this, see other criticisms here and here.
[2] A review of some common issues with classical utilitarianism can be found here.
[3] Of interest may be this review of the God Emperor of Dune. That being said, Leto is but one individual with their own point-of-view, while what we want is the objectively most perfect universe.
[4] Probably because it is a truism. If at any time we are presented with a universe with higher total utility, then we can say that superior option is actually our target universe.
[5] As MacAskill notes in Moral Uncertainty, probabilistic reasoning in utilitarianism and philosophy in general is grossly underused. Although even there, I find the treatment is rather superficial.
[6] An ideal aggregation daemon might be some sort of market where omniscient individuals trade off between each other on possible universes. Personally, I am not sure if such an aggregation is possible even between omniscient entities, as dimensionality reduction inevitably leads to loss of information. In this case, there are potentially infinite parameters which must all be reduced to a single factor of total utility, a curse of dimensionality in overdrive.
[7] See here for more details, or just read Reasons and Persons.
[8] To elaborate, even if almost everyone is living under terrific conditions, they may experience some negative utility if they realize that this comes at the expense of a scapegoat who lives under terrible conditions. Furthermore, the realization that others realize this could also result in negative utility, creating a sort of n-body problem which further complicates aggregation.
[9] See here for some more discussion on moral heuristics, commonsense morality, and utilitarianism.
[11] I can’t find a source for this either. Maybe I’ve just been reading too much Neil Gaiman.
[12] I don’t have to recommend the Scout Mindset, right?
[13] Well, there are some arguments against tight feedback loops, such as here. These tend to be in situations where the outcome difficult to measure, which is exactly what I am arguing against here.
[14] This assertion is based off its recent success in the machine learning field. That being said, there are other methods where analogies can be made. For example, you can think of simulating possible outcomes and choosing the best one as a sort of bisection method.
[15] I don’t deny that altruism in animals does exist, but I am mostly certain that it primary occurs within-species and therefore isn’t very effective.
[17] I am willing to change my mind here, if for example it turns out that long-term prediction markets can actually work. My own personal experience is that even short-term prediction markets aren’t actually all that good yet.
Empirical critique of EA from another direction
As far as I can tell, the most common critique of Effective Altruism is something like this: It certainly sounds reasonable enough, but it feels rather suspicious and some of the outputs really seem rather wonky [1]. Of course, those in the movement know that this is due to bias which arises from misplaced confidence in commonsense morality, mere social convention. But I’d like to proceed from first-principals to show that perhaps commonsense morality should not be dismissed quite so easily.
There is a saying which I cannot find at the moment, which says something to the effect that basically all philosophy is just arguing about definitions. According to the introduction to EA, “Effective altruism is a project that aims to find the best ways to help others, and put them into practice”. As a fellow traveler to the movement, I have my own definition, which is two words: applied utilitarianism.
At this point, you might be thinking: “Aha! So, you don’t like utilitarianism, therefore you don’t like EA.” It is true that various flavors of utilitarianism (including the vanilla classical) do result in some strange results at times [2]. But I have my own definition of utilitarianism, the description of which is the secret agenda of this essay.
A Perfect Definition of Utilitarianism: To maximize total utility, where individual utility is the criteria which would be maximized if an omniscient version of an individual could choose a particular universe out of all possible ones, and total utility is the aggregation of individual utility by an omniscient daemon.
In short, utilitarianism seeks the best of all possible worlds, the golden path [3]. I call this a perfect definition because you can use proof of contradiction to show it is obviously correct [4]. However, it is not a very practical definition, as it relies on omniscience, a quality in short supply. What we want is a methodology from which any entity at a specific time and with a specific point of view can make actionable decisions.
The first use of omniscience in the definition is in knowing all potential universes to know the true objective wellbeing of an individual. One way we can approximate this objective wellbeing is by aggregating over time of the subjective wellbeing of an individual over their entire lifespan. This renders utility measurable by query: “How are you doing?”
That being said, omniscience is still required to know information which occurs in the future. In addition, although one can know their own subjective wellbeing with relative certainty, assessing this property in others is typically done with less certainty (especially given time constraints). We need to solve this problem by placing the calculation of utility inside of a probability distribution [5].
This allows us to specify a calculation for total utility which is actionable to any individual who resides in any particular context.
“Practical” Definition: As an individual, attempt to maximize the expectation of the aggregation over time and individuals of the subjective well-being of all individuals.
U(→x)≈E(∫time∫individualsP(→x))That being said, it should be obvious why “practical” is written here in quotes: this definition is still computationally intractable. In addition, it should be noted that methods to aggregate individual into total utility are left undefined [6].
(Too) Practical Definition: As an individual, attempt to maximize the aggregation over time and individuals of the expected subjective well-being of all individuals, which is equivalent to always choosing the action which leads to the most increase in subjective well-being (under most commonly chosen aggregation functions).
U(→x)≈∫time∫individualsE(P(→x))However, bringing the expectation function inside of the aggregation functions leads to various well-known paradoxes [2]. Famously, the aggregation of individual utility through mere addition leads to the repugnant conclusion [7]. A more reasonable aggregation function needs to consider that identical conditions may be highly pleasurable to one entity while resulting in negative utility in another, while still also accounting for nth-order effects such as the effect of one entity’s well-being on that of others [8]. As a result, I am reluctant to simplify my definition any further.
As a result, we are left with an objective function which is impossible to actually compute. Fortunately, we have a standard method to deal with intractable computations: heuristics. And I posit that this is exactly what commonsense morality is [9].
Selfishness – since you know your own subjective preferences, caring about your own wellbeing versus that of others is entirely reasonable. You can know exactly how much you will enjoy different experiences, while it is unclear if the enjoyment of others is truly as sublime as your own. So, we eat filet mignon.
Kinship bias – Rather than donating to complete strangers, it may be rational to assist those whom you know well, since you know their preferences and can gauge their reactions with a higher degree of certainty, no matter how much they protest “You shouldn’t have!”
In-group bias—In performing actions which you expect to please others, you ought to do so to people who have traits closer to yourself, since you are able to simulate their reactions with greater fidelity. Hence Scotscare [10].
Local bias – In donating, it may be strategic to donate to those you can interact with directly, as you can receive immediate feedback that the money reached the target and was received positively, whereas donations to far away locations may be lost or misused.
Loss aversion – Since one can easily simulate the negative utility a loss would deal oneself; it may be better to prioritize loss prevention rather than realizing a gain which actually isn’t all that great. As they say, the worst thing about realizing your heart’s desire is exactly that [11].
Short-term bias – It is better to prioritize increasing utility in the short term, since shortening the feedback loop can allow more nimble adjustments, while long-term projects will have much higher variance, which should be discouraged due to loss aversion.
The key difference though, is that commonsense morality actually thinks of itself as the answer, whereas the utilitarian knows it is merely a set of various heuristics seeking to approximate total utility. In which case, the strength of an altruism based on utilitarianism is a recognition that these heuristics are subject to biases and must be constantly examined and improved [12].
It should be uncontroversial that the best and most reliable method we have to rectify biases is through the empirical method with fast iterations [13]. It should also be pretty clear by now that as far as optimization go, the most successful general algorithm we have is gradient descent [14]. Therefore, my opinion is that the best way to go about improving utility is to take advantage of these tools: determine key performance indicators which are readily measurable (differentiable) and highly likely to improve total utility (likely to lead to global optima). We can then focus on most efficiently optimizing these parameters. I believe that this is the aspect of effective altruism which is most well received, and with good reason.
On the other hand, there are other “core tenets” of effective altruism which are receiving increased focus in recent years which are essentially untethered from the empirical approach.
Animal welfare – Regardless of the debate over whether the subjective well-being of non-humans is measurable by humans, my issue with this KPI is that assistance to humans can be paid forward with compounding altruism, while others of the kingdom are typically disinterested in the altruism project at best [15].
Excessive focus on infinities in the future [16]– aside from having unacceptably long feedback loops, our collective track record with predicting even the events of the medium-term future have generally been abysmal, owing to the confounding factors of higher-order effects [17].
AI-risk – A variant of x-risk, but one which is mediated through an unknown number of steps of uncertain probability (as compared to other risks like viral pandemic, nuclear war, or heat death, which directly result in dramatic population loss), and therefore particularly hard to estimate or measure.
These projects which presume to know better than the empirical approach, I perceive as being not even wrong [18]. Worse yet, within these topics they often they get bogged down in debates about details, which in a KPI which cannot even be measured properly can only be regarded as premature optimization. That is not to say these topics are not important: it is simply that whether they are effective or not is indeterminable. I cannot refer to them as effective altruism—only as altruism (and that is totally fine).
Footnotes
[1] For examples of this, see other criticisms here and here.
[2] A review of some common issues with classical utilitarianism can be found here.
[3] Of interest may be this review of the God Emperor of Dune. That being said, Leto is but one individual with their own point-of-view, while what we want is the objectively most perfect universe.
[4] Probably because it is a truism. If at any time we are presented with a universe with higher total utility, then we can say that superior option is actually our target universe.
[5] As MacAskill notes in Moral Uncertainty, probabilistic reasoning in utilitarianism and philosophy in general is grossly underused. Although even there, I find the treatment is rather superficial.
[6] An ideal aggregation daemon might be some sort of market where omniscient individuals trade off between each other on possible universes. Personally, I am not sure if such an aggregation is possible even between omniscient entities, as dimensionality reduction inevitably leads to loss of information. In this case, there are potentially infinite parameters which must all be reduced to a single factor of total utility, a curse of dimensionality in overdrive.
[7] See here for more details, or just read Reasons and Persons.
[8] To elaborate, even if almost everyone is living under terrific conditions, they may experience some negative utility if they realize that this comes at the expense of a scapegoat who lives under terrible conditions. Furthermore, the realization that others realize this could also result in negative utility, creating a sort of n-body problem which further complicates aggregation.
[9] See here for some more discussion on moral heuristics, commonsense morality, and utilitarianism.
[10] MacAskill’s favorite example of well-intentioned but somewhat dubious charities, a charity focusing on helping downtrodden Scots living in London.
[11] I can’t find a source for this either. Maybe I’ve just been reading too much Neil Gaiman.
[12] I don’t have to recommend the Scout Mindset, right?
[13] Well, there are some arguments against tight feedback loops, such as here. These tend to be in situations where the outcome difficult to measure, which is exactly what I am arguing against here.
[14] This assertion is based off its recent success in the machine learning field. That being said, there are other methods where analogies can be made. For example, you can think of simulating possible outcomes and choosing the best one as a sort of bisection method.
[15] I don’t deny that altruism in animals does exist, but I am mostly certain that it primary occurs within-species and therefore isn’t very effective.
[16] Especially given MacAskill’s recent release of What We Owe the Future.
[17] I am willing to change my mind here, if for example it turns out that long-term prediction markets can actually work. My own personal experience is that even short-term prediction markets aren’t actually all that good yet.
[14] Not falsifiable, so not useful.