As pointed out recently, suffering focused views imply that a population where everyone experiences extreme suffering is better than a population where everyone experiences extreme happiness plus a brief, mild instance of suffering, provided the latter population is sufficiently more numerous.
This is an overgeneralization of suffering-focused views. You can believe in Lexical Threshold Negative Utilitarianism (ie there is some point at which suffering is bad enough where it becomes infinitely worse than less bad experiences) where the threshold itself is applied at the person-level rather than the aggregate suffering over all beings level. In this case, many people experiencing mild suffering is trivially better than a smaller number of people experiencing extreme suffering. Not sure if I completely buy into this kind of philosophy but I think it’s plausible.
Yes, I agree that lexical NU doesn’t have that implication. My comment was addressed to the particular suffering-focused view I took Stijn to be defending, which he contrasted to CU. If his defence is of “suffering-focused views” as a whole, however, then it seems unfair to compare them to CU specifically, rather than to “classical views” generally. Classical views also avoid the repugnant and very repugnant conclusions, since some specific views in this family, such as critical level utilitarianism, don’t have this implication. [EDIT: Greg makes the same point in his comment; remarkably, we posted at exactly the same time.]
Concerning the merits of lexical NU, I just don’t see how it’s plausible to postulate a sharp value discontinuity along the suffering continuum. As discussed many times in the past, one can construct a series of pairwise comparisons involving painful experiences that differ only negligibly in their intensity. It is deeply counterintuitive that one of these experiences should be infinitely (!) worse than the other, but this is what the view implies. (I’ve only skimmed the essay, so please correct me if I’m misinterpreting it.)
Concerning the merits of lexical NU, I just don’t see how it’s plausible to postulate a sharp value discontinuity along the suffering continuum. As discussed many times in the past, one can construct a series of pairwise comparisons involving painful experiences that differ only negligibly in their intensity.
So, I agree that sharp values in discontinuity are not a great aspect for a moral system to have but consider
We put suffering and happiness on the same scale to reflect how they look in our utility functions. But really, there are lots of kinds of suffering that are qualitatively different. While we can do it sometimes, I’m not sure if we are always capable of making direct, optimized comparisons of qualitatively different experiences
We don’t actually have a great definition of what suffering is and, if we model it in terms of preferences, it bottoms out. AKA, there’s a point in suffering when I could imagine myself saying something like “This is the worst thing ever; get me out of here no matter what.” Our subjective experience of suffering and our actual ability to report it breaks down
It’s also super hard to really understand what it’s like to be in edge-case extreme suffering situations without actually being in one, and most people haven’t. Without that (and even potentially with it), trying to model ourselves in extreme suffering would require us to separate logical fallacies we would make in such a situation with our de-facto utility function. From an AI alignment perspective, this is hard.
If you’re an agent and you can’t reason about how bad something is while you’re in a situation and you don’t have a mental model of what that situation is like, getting into that kind of situation is a really bad idea. This isn’t just instrumentally inconvenient; it’s inconvenient in a “you know you’re suffering really badly but you can only model your experience as arbitrarily bad”
Even if we agree that our utility functions shouldn’t have strange discontinuities in suffering, there may still be a strange and discontinuous landscape of levels of suffering we can experience in the landscape of world-states. This is not directly incompatible with any kind of utilitarianism but it makes arguments along the lines of “imagine that we make this suffering just slightly, and virtually unnoticeably, worse” kind of weird. Especially in the context of extreme experiences that exist in a landscape we don’t fully understand and especially in a landscape where the above points apply
I’m a moral anti-realist. There’s no strict reason why we can’t have weird dicontinuities in our utility functions if that’s what we actually have. The “you wouldn’t want to trade a dramatic amount of resources to move from one state of suffering to an only infinitesimally worse one” makes sense but, per the above, we need to be careful about what that actually implies about how suffering works
This is all to say that suffering is really complicated and disentangling concerns about how utility functions and suffering work in reality from what logically makes sense is not an easy task. And I think part of the reason people are suffering-focused is because of these general problems. I’m still agnostic on whether something like negative lexical threshold utilitarianism is actually true but the point is that, in light of the above things, I don’t think that weird discontinuities is enough to dismiss it from the zone of plausibility.
We don’t actually have a great definition of what suffering is and, if we model it in terms of preferences, it bottoms out. AKA, there’s a point in suffering when I could imagine myself saying something like “This is the worst thing ever; get me out of here no matter what.”
Proponents or sympathizers of lexical NU (e.g. Tomasik) often make this claim, but I’m not at all persuaded. The hypothetical person you describe would beg for the suffering to stop even if continuing to experience it was necessary and sufficient to avoid an even more intense or longer episode of extreme suffering. So if this alleged datum of experience had the evidential force you attribute to it, it would actually undermine lexical NU.
It’s also super hard to really understand what it’s like to be in edge-case extreme suffering situations without actually being in one, and most people haven’t.
It’s even harderto understand what it’s like to experience comparably extreme happiness, since evolutionary pressures selected for brains capable of experiencing wider intensity ranges of suffering than of happiness. The kind of consideration you invoke here actually provides the basis for a debunking argument of the core intuition behind NU, as has been noted by Shulman and others. (Though admittedly many NUs appear not to be persuaded by this argument.)
I’m a moral anti-realist. There’s no strict reason why we can’t have weird dicontinuities in our utility functions if that’s what we actually have.
Humans have all sorts of weird and inconsistent attitudes. Regardless of whether you are a realist or an anti-realist, you need to reconcile this particular belief of yours with all the other beliefs you have, including the belief that an experience that is almost imperceptibly more intense than another experience can’t be infinitely (infinitely!) worse than it. Or, if you want a more vivid example, the belief that it would not be worth subjecting a quadrillion animals having perfectly happy lives to a lifetime of agony in factory farms solely to spare a single animal a mere second of slightly more intense agony just above the relevant critical threshold.
The hypothetical person you describe would beg for the suffering to stop even if continuing to experience it was necessary and sufficient to avoid an even more intense or longer episode of extreme suffering.
Yeah, I agree with this. More explicitly, I agree that it’s bad that the person won’t continue to experience suffering if it will cause them to experience worse suffering and that this implies that lexical trade-offs in suffering are weird. However
I said that “in terms of preferences, [suffering] bottoms out.” In this situation, you’re changing my example by proposing that there is a hypothetical yet worse form of suffering when I’m not convinced there is one after that point
The above point only addresses more intense suffering, not longer suffering. However I think you’re wrong about bringing up different lengths of suffering. When I talk about lexicality, I’m talking about valuing different experiences in different ways. A longer episode of extreme suffering and a shorter form of the same level of extreme suffering are in the same lexicality and can be traded off
It’s even harder to understand what it’s like to experience comparably extreme happiness, since evolutionary pressures selected for brains capable of experiencing wider intensity ranges of suffering than of happiness.
I agree with this and touched briefly on this in my writing. Even without the evolutionary argument, I’ll grant that imagining lexically worse forms of suffering also implies lexically better forms of happiness just as much. After all, in the same way that suffering could bottom out at “this is the worst thing ever and I’d do anything to make it stop”, happiness could ceiling at “this is the most amazing thing ever and I’d do anything to make it continue longer.”
Then you have to deal with the confusing problem of reconciling trade-offs between those kinds of experiences. Frankly, I have no idea how to do that.
Humans have all sorts of weird and inconsistent attitudes. Regardless of whether you are a realist or an anti-realist, you need to reconcile this particular belief of yours with all the other beliefs you have
I actually don’t need to do this for a couple reasons:
I said that I thought negative lexical utilitarianism was plausible. I think there’s something to it but I don’t have particularly strong opinions on it. This is true for total utilitarianism as well (though, frankly, I actually learn slightly more in favor of total utilitarianism at the moment)
The sorts of situations where lexical threshold utilitarianism differs from ordinary utilitarianism are extreme and I think my time is more pragmatically spent trying to help the world than it is on making my brain ethically self-consistent
As a side-note, negative lexical utilitarianism has infinitely bad forms of suffering so even giving it a small credence in your personal morality should imply that it dominates your personal morality. But, per the above bullet, this isn’t something I’m that interested in figuring out
Or, if you want a more vivid example, the belief that it would not be worth subjecting a quadrillion animals having perfectly happy lives to a lifetime of agony in factory farms solely to spare a single animal a mere second of slightly more intense agony just above the relevant critical threshold.
I would not trade a quadrillion animals having perfectly happy lives instead of agony in factory farms just to avoid a second of slightly more intense agony here. However, this isn’t the model of negative lexical utilitarianism I find plausible. The one I find plausible implies that there is no continuous space of subjective experiences spanning from bad to good; at some point things just hop from finitely bad suffering that can be reasoned about and traded to infinitely bad suffering that can’t be reasoned about and traded.
I guess you could argue that moralities are about how we should prefer subjective experiences as opposed to the subjective experiences themselves (...and thus that the above is completely compatible with total utilitarianism). However, as I mentioned
We don’t actually have a great definition of what suffering is and, if we model it in terms of preferences, it bottoms out. AKA, there’s a point in suffering when I could imagine myself saying something like “This is the worst thing ever; get me out of here no matter what.”
so I’m uncertain about the truth behind distinguishing subjective experience from preferences about them.
It is in the context of that uncertainty that I think negative lexical utilitarianism is plausible.
Yet in these cases, it is the lexicality, not the suffering focus, which is doing the work to avoid the counter-example. A total utilitarian could adopt lexicality in a similar way to avoid the (very/) repugnant conclusion (e.g., lives in a ‘repugnant region’ between zero and some ‘barely worth living’ should be ‘counted as zero’ when weighing things up, save as a tie-breaker between equally-good worlds). [I’m not recommending this approach—lexicality also has formidable costs across the scales from its potential to escape population ethics counter-examples].
It seems to miss the mark to say it is an advantage for suffering-focused views to avoid the (v/) repugnant conclusion, if the ‘suffering focus’ factor, taken alone, merely exchanges the (v/) repugnant conclusion for something which looks even worse by the lights of common intuition; and where the resources that can be called upon to avoid either counter-example are shared between SF and ¬SF views.
I think the tools to avoid all three of the the Repugnant Conclusion, the Very Repugnant Conclusion and the Very Sadistic Conclusion (or the similar conclusion you described here) left available to someone who accepts Mere Addition (or Dominance Addition) are worse than those available to someone who rejects it.
Using lexicality as you describe seems much worse than the way a suffering-focused view would use it, since it means rejecting Non-Elitism, so that you would prioritize the interests of a better off individual over a worse off one in a one-on-one comparison. Some degree of prioritarianism is widely viewed as plausible, and I’d imagine almost no one would find rejecting Non-Elitism acceptable. Rejecting Non-Elitism without using lexicality (like Geometrism) isn’t much better, either. You can avoid this by giving up General Non-Extreme Priority (with or without lexicality) instead, and I wouldn’t count this against such a view compared to a suffering-focused one.
However, under a total order over populations, to avoid the RC, someone who accepts Mere Addition must reject Non-Antiegalitarianism and Minimal Inequality Aversion (or Egalitarian Dominance, which is even harder to reject). Rejecting them isn’t as bad as rejecting Non-Elitism, although I’m not yet aware of any theory which rejects them but accepts Non-Elitism. From this paper:
As mentioned above, Sider’s theory violates this principle. Sider rejects his own theory, however, just because it favours unequal distributions of welfare. See Sider (1991, p. 270, fn 10). Ng states that ‘Non-Antiegalitarianism is extremely compelling’. See Ng (1989, p. 239, fn 4). Blackorby, Bossert and Donaldson (1997, p. 210), hold that ‘weak inequality aversion is satisfied by all ethically attractive . . . principles’. Fehige (1998, p. 12), asks rhetorically ‘. . . if one world has more utility than the other and distributes it equally, whereas the other doesn’t, then how can it fail to be better?’. In personal communication, Parfit suggests that the Non-Anti-Egalitarianism Principle might not be convincing in cases where the quality of the good things in life are much worse in the perfectly equal population. We might assume, however, that the good things in life are of the same quality in the compared populations, but that in the perfectly equal population these things are equally distributed. Cf. the discussion of appeals to non-welfarist values in the last section.
And the general Non-Sadism condition is so close to Mere Addition itself that rejecting it (and accepting the Sadistic Conclusion) is not that great a cost to someone who already rejects Mere Addition, since they’ve already accepted that adding lives with what might be understood as positive welfare can be bad, and if it is bad, it’s small step to accept that it can sometimes be worse than adding a smaller number of lives of negative welfare.
This is an overgeneralization of suffering-focused views. You can believe in Lexical Threshold Negative Utilitarianism (ie there is some point at which suffering is bad enough where it becomes infinitely worse than less bad experiences) where the threshold itself is applied at the person-level rather than the aggregate suffering over all beings level. In this case, many people experiencing mild suffering is trivially better than a smaller number of people experiencing extreme suffering. Not sure if I completely buy into this kind of philosophy but I think it’s plausible.
Yes, I agree that lexical NU doesn’t have that implication. My comment was addressed to the particular suffering-focused view I took Stijn to be defending, which he contrasted to CU. If his defence is of “suffering-focused views” as a whole, however, then it seems unfair to compare them to CU specifically, rather than to “classical views” generally. Classical views also avoid the repugnant and very repugnant conclusions, since some specific views in this family, such as critical level utilitarianism, don’t have this implication. [EDIT: Greg makes the same point in his comment; remarkably, we posted at exactly the same time.]
Concerning the merits of lexical NU, I just don’t see how it’s plausible to postulate a sharp value discontinuity along the suffering continuum. As discussed many times in the past, one can construct a series of pairwise comparisons involving painful experiences that differ only negligibly in their intensity. It is deeply counterintuitive that one of these experiences should be infinitely (!) worse than the other, but this is what the view implies. (I’ve only skimmed the essay, so please correct me if I’m misinterpreting it.)
So, I agree that sharp values in discontinuity are not a great aspect for a moral system to have but consider
We put suffering and happiness on the same scale to reflect how they look in our utility functions. But really, there are lots of kinds of suffering that are qualitatively different. While we can do it sometimes, I’m not sure if we are always capable of making direct, optimized comparisons of qualitatively different experiences
We don’t actually have a great definition of what suffering is and, if we model it in terms of preferences, it bottoms out. AKA, there’s a point in suffering when I could imagine myself saying something like “This is the worst thing ever; get me out of here no matter what.” Our subjective experience of suffering and our actual ability to report it breaks down
It’s also super hard to really understand what it’s like to be in edge-case extreme suffering situations without actually being in one, and most people haven’t. Without that (and even potentially with it), trying to model ourselves in extreme suffering would require us to separate logical fallacies we would make in such a situation with our de-facto utility function. From an AI alignment perspective, this is hard.
If you’re an agent and you can’t reason about how bad something is while you’re in a situation and you don’t have a mental model of what that situation is like, getting into that kind of situation is a really bad idea. This isn’t just instrumentally inconvenient; it’s inconvenient in a “you know you’re suffering really badly but you can only model your experience as arbitrarily bad”
Even if we agree that our utility functions shouldn’t have strange discontinuities in suffering, there may still be a strange and discontinuous landscape of levels of suffering we can experience in the landscape of world-states. This is not directly incompatible with any kind of utilitarianism but it makes arguments along the lines of “imagine that we make this suffering just slightly, and virtually unnoticeably, worse” kind of weird. Especially in the context of extreme experiences that exist in a landscape we don’t fully understand and especially in a landscape where the above points apply
I’m a moral anti-realist. There’s no strict reason why we can’t have weird dicontinuities in our utility functions if that’s what we actually have. The “you wouldn’t want to trade a dramatic amount of resources to move from one state of suffering to an only infinitesimally worse one” makes sense but, per the above, we need to be careful about what that actually implies about how suffering works
This is all to say that suffering is really complicated and disentangling concerns about how utility functions and suffering work in reality from what logically makes sense is not an easy task. And I think part of the reason people are suffering-focused is because of these general problems. I’m still agnostic on whether something like negative lexical threshold utilitarianism is actually true but the point is that, in light of the above things, I don’t think that weird discontinuities is enough to dismiss it from the zone of plausibility.
Proponents or sympathizers of lexical NU (e.g. Tomasik) often make this claim, but I’m not at all persuaded. The hypothetical person you describe would beg for the suffering to stop even if continuing to experience it was necessary and sufficient to avoid an even more intense or longer episode of extreme suffering. So if this alleged datum of experience had the evidential force you attribute to it, it would actually undermine lexical NU.
It’s even harder to understand what it’s like to experience comparably extreme happiness, since evolutionary pressures selected for brains capable of experiencing wider intensity ranges of suffering than of happiness. The kind of consideration you invoke here actually provides the basis for a debunking argument of the core intuition behind NU, as has been noted by Shulman and others. (Though admittedly many NUs appear not to be persuaded by this argument.)
Humans have all sorts of weird and inconsistent attitudes. Regardless of whether you are a realist or an anti-realist, you need to reconcile this particular belief of yours with all the other beliefs you have, including the belief that an experience that is almost imperceptibly more intense than another experience can’t be infinitely (infinitely!) worse than it. Or, if you want a more vivid example, the belief that it would not be worth subjecting a quadrillion animals having perfectly happy lives to a lifetime of agony in factory farms solely to spare a single animal a mere second of slightly more intense agony just above the relevant critical threshold.
Yeah, I agree with this. More explicitly, I agree that it’s bad that the person won’t continue to experience suffering if it will cause them to experience worse suffering and that this implies that lexical trade-offs in suffering are weird. However
I said that “in terms of preferences, [suffering] bottoms out.” In this situation, you’re changing my example by proposing that there is a hypothetical yet worse form of suffering when I’m not convinced there is one after that point
The above point only addresses more intense suffering, not longer suffering. However I think you’re wrong about bringing up different lengths of suffering. When I talk about lexicality, I’m talking about valuing different experiences in different ways. A longer episode of extreme suffering and a shorter form of the same level of extreme suffering are in the same lexicality and can be traded off
I agree with this and touched briefly on this in my writing. Even without the evolutionary argument, I’ll grant that imagining lexically worse forms of suffering also implies lexically better forms of happiness just as much. After all, in the same way that suffering could bottom out at “this is the worst thing ever and I’d do anything to make it stop”, happiness could ceiling at “this is the most amazing thing ever and I’d do anything to make it continue longer.”
Then you have to deal with the confusing problem of reconciling trade-offs between those kinds of experiences. Frankly, I have no idea how to do that.
I actually don’t need to do this for a couple reasons:
I said that I thought negative lexical utilitarianism was plausible. I think there’s something to it but I don’t have particularly strong opinions on it. This is true for total utilitarianism as well (though, frankly, I actually learn slightly more in favor of total utilitarianism at the moment)
The sorts of situations where lexical threshold utilitarianism differs from ordinary utilitarianism are extreme and I think my time is more pragmatically spent trying to help the world than it is on making my brain ethically self-consistent
As a side-note, negative lexical utilitarianism has infinitely bad forms of suffering so even giving it a small credence in your personal morality should imply that it dominates your personal morality. But, per the above bullet, this isn’t something I’m that interested in figuring out
I would not trade a quadrillion animals having perfectly happy lives instead of agony in factory farms just to avoid a second of slightly more intense agony here. However, this isn’t the model of negative lexical utilitarianism I find plausible. The one I find plausible implies that there is no continuous space of subjective experiences spanning from bad to good; at some point things just hop from finitely bad suffering that can be reasoned about and traded to infinitely bad suffering that can’t be reasoned about and traded.
I guess you could argue that moralities are about how we should prefer subjective experiences as opposed to the subjective experiences themselves (...and thus that the above is completely compatible with total utilitarianism). However, as I mentioned
so I’m uncertain about the truth behind distinguishing subjective experience from preferences about them.
It is in the context of that uncertainty that I think negative lexical utilitarianism is plausible.
Yet in these cases, it is the lexicality, not the suffering focus, which is doing the work to avoid the counter-example. A total utilitarian could adopt lexicality in a similar way to avoid the (very/) repugnant conclusion (e.g., lives in a ‘repugnant region’ between zero and some ‘barely worth living’ should be ‘counted as zero’ when weighing things up, save as a tie-breaker between equally-good worlds). [I’m not recommending this approach—lexicality also has formidable costs across the scales from its potential to escape population ethics counter-examples].
It seems to miss the mark to say it is an advantage for suffering-focused views to avoid the (v/) repugnant conclusion, if the ‘suffering focus’ factor, taken alone, merely exchanges the (v/) repugnant conclusion for something which looks even worse by the lights of common intuition; and where the resources that can be called upon to avoid either counter-example are shared between SF and ¬SF views.
I think the tools to avoid all three of the the Repugnant Conclusion, the Very Repugnant Conclusion and the Very Sadistic Conclusion (or the similar conclusion you described here) left available to someone who accepts Mere Addition (or Dominance Addition) are worse than those available to someone who rejects it.
Using lexicality as you describe seems much worse than the way a suffering-focused view would use it, since it means rejecting Non-Elitism, so that you would prioritize the interests of a better off individual over a worse off one in a one-on-one comparison. Some degree of prioritarianism is widely viewed as plausible, and I’d imagine almost no one would find rejecting Non-Elitism acceptable. Rejecting Non-Elitism without using lexicality (like Geometrism) isn’t much better, either. You can avoid this by giving up General Non-Extreme Priority (with or without lexicality) instead, and I wouldn’t count this against such a view compared to a suffering-focused one.
However, under a total order over populations, to avoid the RC, someone who accepts Mere Addition must reject Non-Antiegalitarianism and Minimal Inequality Aversion (or Egalitarian Dominance, which is even harder to reject). Rejecting them isn’t as bad as rejecting Non-Elitism, although I’m not yet aware of any theory which rejects them but accepts Non-Elitism. From this paper:
And the general Non-Sadism condition is so close to Mere Addition itself that rejecting it (and accepting the Sadistic Conclusion) is not that great a cost to someone who already rejects Mere Addition, since they’ve already accepted that adding lives with what might be understood as positive welfare can be bad, and if it is bad, it’s small step to accept that it can sometimes be worse than adding a smaller number of lives of negative welfare.