Choose your (preference) utilitarianism carefully - part 1

Part 1 - Part 2


The overall goal of this three-part essay is to at least clarify and hopefully move people away from preference utilitarianism towards valence utilitarianism, (more discussion on why I use this terminology below, but consider it a synonym for hedonistic utilitarianism).

Utilitarianism is often ill-defined by supporters and critics alike, preference utilitarianism even more so. I briefly examine some of the axes of utilitarianism common to all popular forms, then look at some axes unique but essential to preference utilitarianism, which seem to have received little to no discussion - at least not this side of a paywall. This way I hope to clarify future discussions between valence and preference utilitarians and perhaps to clarify things for their critics too. That said I’m aiming the discussion primarily at utilitarians and utilitarian-sympathisers.

Part 1, below, only addresses the clarification. I hope that even if utilitarians (and their critics) disagree with all my subsequent arguments, this will still help to make future discussions more productive.

Part 2 is an argument against the coherence of the most visible arguments for preference utilitarianism. As such it doesn't constitute an argument for valence utilitarianism, and I intend it more as a reference for when people dismiss the latter with such common arguments.

Part 3 will argue that, given certain key assumptions about philosophical methodology that I think most preference utilitarians will at least be sympathetic to, valence utilitarianism is in fact a more justifiable position than preference utilitarianism.

Introduction - the axes of all utilitarianisms

Any well-defined form of utilitarianism (or rather of any theory of anything) must comprise a set of precise points on a - perhaps infinite - number of axes. Throughout, a ‘point’ in this sense is shorthand for ‘a specific and parsimonious proposition about utilitarianism’ (or about the subject in question, which itself could be a superordinate point about utilitarianism, or whatever subject). Meanwhile, something is an ‘axis’ in this sense only if a) it has multiple distinct points and b) each point is exclusive of each other point1 (though a and b don’t quite imply axiality - see appendix).

In discussing any form of utilitarianism (or other ethical view) clearly, it’s therefore essential to pare these axes down to those about which people could realistically disagree - I’ll refer to these as the ‘salient’ axes (and allow for the possibility of more and less salience, where relevant). A good guide to which these are is those axes about which people have disagreed.

For further clarity, we want to reduce the number of dimensions each of these salient axes has as far as possible - often to 1. Sometimes, though, this doesn’t seem to be possible, such as when we’re comparing points that represent different mathematical functions - eg averaging and totalising utilitarianism.

Many of these axes are well discussed, pertinent to almost any form of utilitarianism, and at least reasonably well understood, and I don’t propose to discuss them here beyond highlighting their salience. These include but probably aren’t restricted to the following:

  • What is utility? (for the sake of easy reference, I’ll give each axis a simple title - for this, the utility axis); eg happiness, fulfilled preferences, beauty, information
  • How drastically are we trying to adjust it?, aka what if any is the criterion for ‘right’ness? (sufficiency axis); eg satisficing, maximising2, scalar
  • How do we balance tradeoffs between positive and negative utility? (weighting axis); eg, negative, negative-leaning, positive (as in fully discounting negative utility - I don’t think anyone actually holds this), net (often called positive, but it would benefit from a distinct adjective)
  • What’s our default heuristic for applying it? (mentality axis); eg act, rule, two-level, global
  • How do we deal with changing populations? (population axis); eg average, total
  • To what extent do we discount future utility? (discounting axis); eg zero discount, >0 discount
  • How do we pinpoint the net zero utility point? (balancing axis); eg Tännsjö's test, experience tradeoffs
  • What is a utilon? (utilon axis)3 - I don’t know of any examples of serious discussion on this (other than generic dismissals of the question), but it’s ultimately a question utilitarians will need to answer if they wish to formalise their system.

Note that there’s nothing essential about the characterisation of the axes I’ve given above and those I’ll give below, and even if I’ve picked up all the major options, they could perhaps be recast differently to similar effect. I’ve selected them because I think they represent the way people actually conceive (and most commonly differ about) utilitarianism.

In some cases, certain points on one axis will have subcategories of their own - dependent or subordinate axes. What I’d like to explore in more depth now are the dependent axes of preference utilitarianism (ie ‘fulfilled preferences’ on the utility axis), because writings on the subject seem to contain a lot of unspecified assumptions, little analysis of these dependent axes, and consequently, a lot of confusion.

In the first part of this essay (this post) I’ll look at what I think are the six key dependent axes of PU (although to reiterate there’s nothing essential about the way I’ve cast them), and briefly note what requirements they have - ie how to discern where on them a particular form of PU sits.

In the second post, I’ll look at some popular and influential online arguments for PU and try to determine where they sit on relevant axes and whether they consistently adhere to the conclusions that their premises - once clarified - suggest.

In the last post, I’ll criticise many of the points on the PU axes, and try to show where they don’t actually contradict the conclusions their advocates hold or prove ultimately incoherent, they either require spooky metaphysical properties or imply conclusions that I think will seem almost universally undesirable.

I have lofty ambitions here - I’m hoping to demonstrate that no distinct form of PU is ultimately tenable. But failing that, I hope that breaking it down this way will still prompt those who fault some point of the argument to refine their views so that HUs and PUs can better understand where we actually differ - and better arrange trade-offs when we do.

The myriad PUs, classified

Here then, are the six most salient dependent axes of preference utilitarianism, ie those that describe what could count as utility for PUs. I’ll refer to the poles on each axis as (axis)0 and (axis)1, and any intermediate view as (axis)X. We can then formally refer to subtypes of preference utilitarianism (points on these axes), and also exclude them with '¬', eg ¬(F0)R1PU, or ¬(F0 v R1)PU etc, or represent a range, eg C0..XPU.

How do we parse misinformed preferences? (information axis F) (F0 no adjustment / F1 adjust to what it would have been had the person been fully informed / FX somewhere in between)

How do we parse irrational preferences? (rationality axis R) (R0 no adjustment / R1 adjust to what it would have been had the person been fully rational / RX somewhere in between)

How long is a preference resolvable/thwartable? (duration axis D) (D0 During its expression only / DF1 During and future / DPF1 During, future and past (shorthand for DP1F1) / DPXFX Somewhere in between)

What creates a preference? (creation axis C) (C0 Phenomenal experience only / C1 Behaviour only / CX A combination of the two)

What resolves a preference? (resolution axis S) (S0 Phenomenal experience only / S1 External circumstances only / SX A combination of the two)

How do we parse irresolvable preferences? (irresolvability axis A) (A0Treat them as no preference / A1 replace them with something resolvable / Ax1 adjust to somewhere in between

(I'll provide frequent tooltips for the PU types so the reader needn't feel it's necessary to memorise this notation)

What distinguishes these categorisations is that each category, as far as I can perceive, has no analogous axis within valence utilitarianism. In other words to a valence utilitarian, such axes would either be meaningless, or have only one logical answer. But any well-defined and consistent form of preference utilitarianism must sit at some point on every one of these axes.

In a sense, all these PU axes are spectra - that is, I’ve tried to restrict them to a single dimension but on each there are an infinite number of possible answers a preference utilitarian could give within them. I suspect that in practice most people will find any point besides the edges and sometimes the centre too arbitrary, but it’s worth being aware of all the possibilities.

Strictly speaking though, they’re really n-dimensional spaces rather than spectra, since when weighing two factors, one could do so nonlinearly, probably in infinite alternative ways, and with infinitely many possible compromises between weighing methods. (Similarly, points that are a combination of two non-negating claims, eg on the creation and resolution axes could in principle overlap - but in practice people seem to view the relative weight of C0 as the inverse of C1, and similarly on the resolution axis). What matters in this essay though, is the exclusivity within axes, not the number of dimensions that exclusivity might reach in.

For the sake of simplicity I’ve assumed linear weightings between each pole on the axes. I can imagine people sincerely wanting to use alternatives eg exponential curves or log scales or some sort of minimum requirement of some factor. I think my criticisms never hinge on the exact approach to weighting, but if a PU does think my eventual criticism fails in such a case, I hope Part 1 at least will stand up enough to help her to clarify her overall position.

Exploring PU’s axes further

Now let’s go through each of them in more detail. I’ve tried to illustrate each point with an example where I can think of one, but doing so has hidden risks. Since most real-life situations involve numerous factors, often any PU prescription will necessarily entail decisions about other axes than the focal one. Where I’ve found this inescapable, I’ve highlighted the salient axis.

A quick note: the ordering of points in the descriptions below is counterintuitive, with the two poles first then the mid-position. This on pragmatic grounds: the mid-position is much easier to describe with reference to poles than before they've been defined.

For the sake of clarity, let’s assume in each case (or whenever relevant) that a disinterested PU is arbitrating on the example (and let’s assume she’s the only person in a position to settle matters in each case, such that we have no concerns about unwanted paternalism). We can then see which particular PU subtypes lead her to which decisions.

How do we parse misinformed preferences? (information axis F) (F0 no adjustment / F1 adjust to what it would have been had the person been fully informed / FX somewhere in between)

The points here are difficult to distinguish in practice, since often following an ill-informed preference will cause greater future disutility, which the PU might well not discount. So a non-misleading illustration of the points in isolation require minimal future consequences.

Imagine a wealthy octagenarian who, being a devoted cat fan had willed part of her estate to a cat shelter charity. Around the time of her death, the charity coincidentally had decided to reinvent itself as a dog shelter charity, but another charity nearby had reinvented itself as a cat shelter charity to pick up the slack. All else being equal in that sterile way only philosophical thought experiments allow (eg legal issues somehow being irrelevant, there being no social ramifications in the decision to change or not, and no substantial instrumental reason to do otherwise), a F0PU executor of the will would consider the octagenarian's initial choice to be her preference, whereas a F1PU would consider the second charity to be.

As formulated, this also relies on the rejection of D0S0PU), which would imply the octagenarian's preferences didn't outlast her death, but we could isolate the question from that factor by imagining she had given power of attorney to someone else, who for some reason had to make the decision for her without consultation before her death.

‘Being fully informed’ is obviously an impossible ideal, but in practice presumably either means ‘as informed as the arbitrating PU’ or ‘as informed as possible’ - where some clear definition on what qualifies as ‘possible’ would be necessary to make this a well-formed position.

FxPU here requires the extra variable ‘well-informedness’, which could itself could comprise various factors.

(Theoretically both this axis and the one below could run negative - to prioritise preferences that are worse informed, or worse considered. This acknowledgement of that symmetry out of the way, I’m going to assume that no-one would take such a proposal seriously.)

How do we parse irrational preferences? (rationality axis R) (R0 no adjustment / R1 adjust to what it would have been had the person been fully rational / RX somewhere in between)

This axis is usually discussed in tandem with the one above, but they’re distinct in principle (although some pragmatist philosophers - eg Quine, and come to that me - would deny that there’s any fundamental difference between logical and factual questions. Readers who share this view might skip this section and assume the rationality and information axes are one and the same wherever it matters, if sometimes useful to separate in practice - unless they think of 'rationality' as encompassing something further, eg self-interested prudence).

If you do think it distinct, the nature of this axis and its implications depend on your definition of and beliefs about rational preferences. This too big a caveat to explore here - I’ll just give some examples of what it might mean, and hope that any ¬(R0)PUs unpersuaded by the later parts of this essay will clarify their position.

R0PU is consistent with respecting4 any preferences which seem likely to lead to bad outcomes. Superficially it seems like a ‘naive’ form of preference utilitarianism, which supposedly sanctions giving a cigarette to a child because he wants one, even if you think that he’ll consequently get addicted and rue the day you granted his wish. Actually doing so in ordinary circumstances would suggest certain positions on other axes though, for example, ¬(DPF1PU), since we expect him to develop a past-facing preference to have not taken up smoking, and heavy discounting of future utility (one of the axes common to any form of utilitarianism described in the 'Introduction - the axes of all utilitarianisms' section above) - any other point on that axis would allow you to acknowledge the child’s desire but recognise that his future desires will outweigh it.

R1PU might imply turning the child down (or rather, considering his 'true' preference to be not to have a cigarette), though this isn't clearcut. Decoupled from the concern about future regret, it's not obvious what a perfectly rational preference might look like, if not 'current preference with any lapses in logic resolved'. And resolving such lapses raises questions of epistemic uncertainty, since an honest arbitrator could have no way of knowing whether she or the preferrer (in this case the child) is more logically competent. That said, she might be able to allow for her own error rate. And in practice, there might be outside clues as to the ‘rational’ view if she looks for them, such as expert consensus on the issue in question.

Given an imperfectly rational arbitrator, R1 poses another problem: spotting whether a pertinent issue is of fact, logic, value or even something else. For example, on philosophical questions, such as PU itself - will there be universal agreement among PUs as to whether HU, deontology, intuitionism etc are mistaken or just different preferences?

As presented I don't think these are destructive criticisms of R1, but they raise some questions that an R1PU will need to answer before she can fully apply her system.

Similar to the mid-point on the previous axis, RxPU is just a question of how rational a person must be - again, an extra variable. In this case a PU might realistically favour a minimal requirement rather than a weighting - for example, an arbitrator might think an animal or human child not sufficiently rational to make potentially harmful decisions, but all adults of sound mind competent to do so.

Such a position is tricky to maintain consistently at a fundamental level though, since advocates of it usually grant that children are capable of determining some of their preferences: like whether they prefer chocolate or vanilla ice cream. It doesn’t seem very consistent to selectively prohibit only choices that we think have a harmful option on the grounds that children are irrational in those cases only.

Obviously we might have reasons unrelated to this axis for forbidding them (eg DPF1PU) - but invoking those to justify withholding drugs from a child would suggest we’re not really adhering to RxPU.

How long is a preference resolvable/thwartable? (duration axis D) (D0 During its existence only / DF1 During and in its future / DPF1 During, and in both its future and past / DPxFx Somewhere in between)

Not to be confused with the question of how much to value future people (or future utilons), this axis applies to past (and maybe future) preferences, whether or not their source still lives (or has yet to live). Like the other five PU axes, it’s part of the question of what counts as utility in the first place. Technically I’ve snuck two dimensions onto this axis (past and future relevance), but I think they’re similar enough to be discussed together.

D0PU means a preference can only be thwarted or satisfied while someone is holding it (or revealing it, depending on which you choose from C0..1PU above). This doesn’t restrict the D0PU as much as it seems it initially might, since she might allow for implicit and entailed preferences (wanting most daily pursuits might entail ‘a preference for’ not having one’s legs broken, for example). It does mean, though, if one also accepts C0PU, that when an agent5 is unconscious or in deep sleep, they can have no preferences. Given any other point on the creation axis, one might still think such preferences are impossible, depending on exactly how you conceive the relationship between behaviour and preference creation.

DF1PU means that once formed, a preference endures for some nonzero time. Thus the DF1PU will object to stealing from a sleeping agent for eg (even given no other adverse consequences, like them noticing the the or having other preferences thwarted for want of the stolen object).

DPF1PU is sufficiently difficult to parse that it might be a mirage area. It is distinctly relevant only to past-facing preferences, such as ‘I wish I hadn’t drunk that curdled milk’ or ‘I wish I hadn’t killed so many people.’ In practice these are hard to distinguish from the future-or-present-referencing counterparts that inevitably accompany them - ‘I wish I could stop feeling sick’, and ‘I wish I could stop feeling guilty’ - but they’re perhaps not fundamentally equivalent. One might say ‘If I had the choice again I’d have taken it differently, but I don’t have any forward-facing discomfort about taking it the way I did’. Even this could be self-deception, but perhaps we can conceive an agent choosing on a coin flip to make a choice, and wanting out of curiosity to retake it if they had the chance. Even if we think their curiosity misguided since they presumably wouldn’t remember their ‘first’ choice if they could later change time, if we accept R0PU (or F0PU, if backward time travel turns out to be possible), we can still respect the preference they base on it.

Also even if a past-facing preference did always entail a future-facing preference, it doesn’t mean they couldn’t be considered two separate preferences. Conceivably this could make past-facing preferences twice as strong as their future-facing counterpart alone would be.

This all seems quite fiddly for what seems to be a relatively minor adjustment to the theory, but since, entropy aside, modern physics apparently treats both directions in time as essentially symmetrical6, it seems likely some modern ethicists will want to follow its example.

It's worth noting that these difficulties might be mirrored, perhaps less obviously, in future-facing preferences: 'I don't want to drink curdled milk' and 'I don't want to slaughter people' are difficult to tease apart from the concern that you might soon develop the ‘I wish I could stop feeling sick’, and ‘I wish I could stop feeling guilty’ preferences. As above, they needn't be logically the same, but if you were confident you would both exist and not have any future preferences against (the consequences of) the interim outcome, I suspect many people's intuition about the relevance of future-facing preferences would be weaker.

We could also conceive of preferences as only facing into the past, if we think that’s conceptually easier than preferences having to hang around in the aether. This is also hard to parse, and I’ve never heard anyone advocate it, so I won’t discuss it further - though I’m less confident that no-one would ever hold this position.

So what does DPF1 determine about our behaviour? I thought this confusing at first, until I realised so long as you’re clear on where your PU is on the (non-PU specific) population axis, it’s fairly straightforward. If you do accept past-facing preferences, you just remember to include them in your population-ethical account - that will determine what sort of impact you have on the number and nature of future preferences (such as maximising them), and then you can optimise for a trajectory with as many/as high an average of preferences fulfilled as possible, past-facing ones included.

DPFxPU would determine the discount or decay rate you attach to future- or past-facing preferences. If it’s a sufficiently low rate, then DPFxPU would still respect the preferences of our long-dead mammalian ancestors, and of any previous sentient (or non-sentient, for C1PUs) creatures, possibly outweighing the preferences of those who’re actually alive and so keeping them morally oppressed by our ancestral overlords. Equally, even if we discount future utility, it might still require us to create contemporary utility by fulfilling the past-facing preferences of our descendants.

Perhaps more counterintuitively, a sufficiently low discount rate might bind us to the preferences we’d formed earlier in life. If a child had always dreamed of becoming a paleontologist, then his adult self suddenly discovered a love of teaching, DPFxPU might dictate that he become a paleontologist anyway, since his past self had had the more enduring preference.

If the discount rate is sufficiently high though, someone on life support might ‘lose the preference’ they’d expressed before what befell them. If they’d wanted to stay on it, their preference would cease to have any force and a PU arbitrator could sanction removing them from it (or vice versa).

This 'range' might also include the concept of implicit preferences mentioned above - though arguably that issue merits its own axis.

What creates a preference? (creation axis C) (C0 Phenomenal experience only / C1 Behaviour only / CX A combination of the two)

These next two axes are perhaps the most counterintuitive in the sense of triggering importantly different intuitions in people (anintuitive?).7 I’m still not fully confident about how to most usefully present this and the resolution axis below, but a couple of remarks before I break them down:

Firstly, I originally had these as a single axis, seeing ‘behaviour’ and ‘external circumstances’ (from the resolution axis) as two different ways of seeing the same category - change in non-conscious states. I still lean towards that view, but they're not self-evidently equivalent, and arguing that they are here would be a needless digression.

Secondly, I should arguably put ‘behaviour’ and ‘phenomenal experience’ on two separate axes. But, while one might envisage subordinate axes for one or both concepts, they do seem exclusive. Whatever you think either entails, it can only be as relevant to preference-construction as the other one isn’t. See the appendix for more discussion on the structuring of axes here.

To the options: C0PU means that a preference begins when (and, on D0PU, only persists as long as) someone phenomenally desires something. Though many PUs might allow for implicit preferences, such as their planet still having a breathable atmosphere, these would still have to arise from a conscious preference that entailed them. On this claim, conscious desire (whatever it is) is both necessary and sufficient to form a preference.

C1PU means that any apparently goal-seeking behaviour counts as a preference - to use a classic example, a thermostat seems to have a preference to regulate the temperature. An economist who takes the idea of revealed preference as wholly trumping what an agent (thinks she) wants would be a C1PU.

CXPU means both phenomenal experience and goal-seeking behaviour are necessary to consider something a preference (or perhaps that either alone would suffice). Here I can imagine CXPUs frequently opting for a threshold rather than a linear weighting - either that some amount of consciousness is sufficient but behaviour determines the strength of preference, or vice versa.

We have an intuitive sense of what it means to have a stronger phenomenal desire for p than q, but the same doesn’t seem to be true of behavioural desire. So a PU advocating C1PU or non-threshold versions of CxPU would need to give some equivalently practical sense of how we could judge the intensity of preference from behaviour alone. Or otherwise, they’d need to eschew the idea of more and less intensity, and treat all preferences as equally weighted.

In either case, per the running theme of this essay, I hope they’d make their position clear.

What resolves a preference? (resolution axis S) (S0 Phenomenal experience only / S1 External circumstances only / SX Some combination of the two)

How does a preference come to be thwarted or satisfied? This is the counterpart of the creation axis, separated to some (possibly 0) degree by the duration axis. Indeed, a D0PU might prefer to unify the resolution and creation axes, since highly closely analogous events happening in what seems like the same place at the same time are likely to be different descriptions of a single phenomenon.

Clearly this is closely analogous to the creation axis, whether or not it’s genuinely distinct. Assuming some persistence of preferences (ie ¬D0PU), though, and therefore some gap between creation and resolution we probably wouldn't consider them logically equivalent.

S0PU means that a preference is satisfied iff (or to the extent that) an agent believes that whatever she desired has come to be – if she happens to be wrong, her sense of satisfaction is nonetheless sufficient (though she might revisit the desire, or another very like it, if she learns of her mistake in future).

S1PU means that her phenomenal experience is irrelevant to the preference in question (though she might form a separate preference to have/not have a certain phenomenal experience). Even if an agent is convinced her desire has been thwarted, if she happens to be wrong, she’s gained utility. As with the creation axis, we have an intuitive sense of degree where phenomenal experience is concerned that doesn't obviously apply to 'external circumstances'. We can easily intuit believing more or less strongly that external circumstances have resolved a preference, but less so to what extent they have. So unless an S1PU wishes to say that preference satisfaction is a binary state, she’ll hopefully now see the need to specify something what types of circumstances constitute greater and lesser degrees of preference satisfaction.

SXPU, like CXPU, requires an account of both types of satisfaction, and a way of weighing them (or an SXPU could take a threshold view, but that seems like it would be less intuitive for resolution than creation).

How do we parse irresolvable preferences? (irresolvability axis A) (A0Treat them as no preference / A1 replace them with something well formed / Ax adjust to somewhere in between

Irresolvable preferences are a fairly wide net. They include preferences whose very statement entails an error (whether factual or logical) - for example, ‘I want Clark Kent to beat Superman in the high jump’ or ‘I want to spend more than £1000 but less than £100’. Unlike misinformed or irrational preferences, which seem to have plausible equivalents once we adjust for the apparent error, there isn't really a meaningful adjustment we can make here to leave us with anything we can act on.

We can also include preferences that simply have no parsable referent, such as 'I wish I had a nice tnetennba', and preferences whose referent can never be confirmed, denied, or even investigated, such as 'I hope the universe is infinite' (or if you think we might somehow find a (dis)proof of that, something like 'I hope there are habitable planets outside the observable universe').

This axis is the only one whose points I’ve never actually seen discussed, but it seems highly significant to PU decision-making and subject to disagreement, hence its appearance here.

This might not be a relevant axis to C1PUs, since they might think that one cannot reveal a preference with no tangible target.

As for the different points A0PU means we conceptually replace the irresolvable statement with ‘no preference’. A1PU seems like it would require minimally ‘fixing’ their preference - imagining what they’d want given all the facts/perfect rationality/physically accessible referents and treating whatever preference they'd have formed as their real preference. Or as an alternative or subclass of this approach, we might simply believe all irresolvable preferences to be auto-thwarted.8

Despite their similarity, an A1PU needn’t necessarily also be either a F1PU or a R1PU, since in this situation an arbitrator might feel forced to make some adjustment to allow the preference to make sense in preference to ignoring it, and might for eg view A1 as the least of all evils.

AxPU presumably means somehow diminishing the weighting of the adjusted preference, perhaps to accommodate the PU’s reduced confidence that she’s accurately representing the agent.

Overlapping with valence

One last point of note on the PU axes: if you choose the leftmost point on each of them (ie P0 for every value of P), it is, as far as I can tell, functionally a subset of valence utilitarianism. At the very least, we could class them both under the fairly specific label 'mental state utilitarianism'. Your form of utility is exclusively felt experience (although you might think this itself might be broken down further), has no influence on the future or past except inasmuch as it predicts similar felt experience in each, and appeals to the emotions people actually feel rather than those they ideally would.

It avoids most of the obvious counterintuitive conclusions from P0 on any given axis by having a very low (typically zero) future discounting rate - meaning that if a preference seems 'irrational' or 'misinformed', resolving it is highly likely to lead to future thwarted preferences and thus a worse world overall. Where it still raises counterintuitive conclusions, we’ll see in parts 2 and 3 that even forms of PU that are often assumed to avoid such conclusions typically end up recommending the same actions.

Whether you then call this PU or HU is a matter of preference, but I suspect many preference utilitarians of holding a stereotypical view of HU. On this stereotype, HU supposedly excludes a whole range of positive/negative emotions from consideration, claiming they’re ‘too rich’ to be encapsulated by the concepts of happiness, pleasure and suffering. Such critics apparently don’t realise that for most HUs these words are placeholders, used slightly idiosyncratically to precisely cover the whole range of possible negative and positive emotions, since the English language has no better words for the job - or at least, none in common use.

To clear up this confusion once and for all, I would like to rebrand hedonistic utilitarianism (and hence this particular form of PU) as ‘valence utilitarianism’9 - which links it neatly with the better defined psychological concept.

Concluding thoughts

For now, let’s treat these axes as having just three points, ignoring the continuity between them. I suspect that each would find some support among PU advocates - I’ll offer a number of examples in part II, but none of the essays of the essays I know of are specific enough to describe a PU that's exactly definable on all the axes.

If different points on each axis do have some support, and if I’m right that HU doesn’t have its own unique axes, the numerous configurations of PU multiplied out mean it has hundreds of times as many advocated forms as HU.10

This is not yet a criticism of PU. Even if you consider parsimony an important property for any ethical theory (as I do), 'number of axes' os not something we should be concerned about the parsimony of, since we know there are infinite conceivable axes, and we've just been focusing on those people have happened to dispute.

That said, I am very suspicious of PU arguments which resort to playing with a new axis as an ad hoc way of fixing an apparent problem with their ethics. Such a wriggle is virtually always possible (see appendix), and I won’t argue against this approach in any depth here, except to say that it seems very much against the spirit of disinterested inquiry from which most utilitarians feel their ethos springs.

As we’ll see in part 2, many of the PU-supporting arguments I’ve seen waver, usually implicitly, between inconsistent PU-forms (that is to say, inconsistencies given the actual approach of the writer, not ones which cannot conceivably be loopholed out of), and it’s this which I want to discourage. Though I’m planning to criticise such writings, and to argue against many of the possible points on the PU axes in part 3, I hope that if PUs and its critics alike merely become more specific about what theories they have in mind and what they really imply, I can call this essay a success.

Appendix - on axes, their value and limitations

As Lila Reiber pointed out, the idea of axes is ultimately a metaphor, and in her view a distracting one. Most of the other commenters liked it (as do I), so I kept it in, but felt like it warranted further discussion outside the main essay.

While it is a metaphor, I think it’s a robust one, since it allows us to better clarify our views by highlighting the different relationships between such concepts as a) satisficing vs rule util, and ii) satisficing vs scalar util.

The former, it seems, is an orthogonal relationship on the spatial metaphor, the latter an colinear (and hence exclusive) one. One can be either a satisficing util and a rule util, or both or neither, but one cannot for practical purposes be both a satisficing and maximising util. It seems there is a categorical difference between their differences.

But that ‘for practical purposes’ caveat is an important one. The boundaries between many (perhaps all) axes are blurry and over-restrictive, and set primarily (perhaps exclusively) by the positions people have taken, whether they’ve done so with sound reason or not.

For example, ‘total vs average’ and ‘negative vs net vs positive’ seem to be distinct axes, but it’s messier than that. For one thing, the order in which you parse the adjectives seems to change the output. A negative-average utilitarian might seek to ensure that, whatever population he looks at, its overall average utility is non-negative and otherwise remain ambivalent about outcomes. An average-negative utilitarian might seek to minimise the average negative utilons of each member of the population.

More importantly, average middling utilitarianism over a fixed population seems like it could dispense with the balancing axis altogether. As long as we can distinguish between multiple utilities, saying which is greater than another and by how much, the notion that some experience is ‘zero utility’ is as irrelevant as pointing to the ‘start’ of a circle when measuring its circumference.

For similar reasons, the description of an axis as ‘two or more distinct points only if each point excludes all the others’ is under-restrictive. In many cases (again, perhaps all, with enough imagination), two points surely could have been included on the same axis and haven’t. Let’s look at some possibilities:

As Kaj Sotala observed, neither ‘[not-necessarily-conscious] behaviour’ nor ‘consciousness’ is normally defined by the absence of the other. But in answer to the question ‘what creates a preference?’, we can see that each bounds the prominence of the other. If behaviour alone determines preference then, while that behaviour might intersect with consciousness on occasion, consciousness cannot even partially determine preference. Similarly, if behaviour determines 90% of a preference (whatever we might understand by that), consciousness cannot determine more than the remaining 10% - and so on. Furthermore, if we’re considering consciousness and behaviour as the only possible sources of preference, we know that each must constitute precisely the remaining proportion – ie precisely 10% in the example above. As long as we don’t consider anything else, it seems that the two ideas belong as firmly together (or, if you prefer, as firmly apart) as satisficing and scalar util.

If, on the other hand, we allow the possibility of a third variable (or more), we find ourselves with a multi-dimensional axis, which changes the landscape. Behaviour and consciousness still cap each other, but don’t fully determine the import of the other, since the remaining proportion of preference-creation can be split between the remaining variables.

What could other variables be? Brian Tomasik suggests ‘unconscious reflexes, structural features (e.g, laws of a country, even if not enforced), verbalized attitudes, etc.’ These sound to me like subsets of behaviour, but perhaps that’s a sign that (preference-revealing-) ‘behaviour’ itself needs to be broken down more.

Even if we accept pure C0PU, we have to contend with the fact that ‘consciousness’ is not well understood, perhaps not well defined, and probably has numerous salient subordinate axes even within our current understanding. It seems reasonable to expect that behaviour could have comparable intricacy.

Alternatively, I can at least conceive of (non-salient) variables that emerge eg from pure maths, like preferences somehow emerging from digits of pi.

But by the time the points on ‘an axis’ sit on multiple dimensions and themselves comprise salient axes (which may have multiple dimensions), the intuitive value of the axis metaphor breaks down.

So is there a categorical difference between, say, a) satisficing vs scalar util and b) a point on one of the multi-dimensional salient subordinate axes of behaviour vs a point on one of the multi-dimensional salient subordinate axes of consciousness? For that matter, if we ultimately want to map ethics to the physical world, should we be thinking of the lowest level as in some sense waves instead of/as well as points? I honestly have no idea yet. I don’t even know if they’re useful questions.

The point of this appendix, besides navel-gazing, is this: too much of philosophy either appeals to pure intuition or is written as though the author has privileged access to a precise ingredients list of primitive concepts11. I wish to avoid the latter here, while still allowing i) that there might be such primitive concepts, as well as concepts that, for the sake of ethical discussion we can treat as primitive, ii) that we need some conceptual framework for dealing with them, iii) that the most helpful form of such a framework depends largely or entirely on the finite set of ideas humans actually debate, rather than on some ‘scale model’ of the infinite set of conceivable ideas, and iv) that the framework I’ve offered is as good as any I know, better than many - and far better than the absence of an explicit framework in which many ethical arguments operate.

Hopefully others will be able to both use this framework and refine it.

1 If a reader can show that in any case a utilitarianism-form could logically have multiple answers within the axis, I’ve probably got something about the axis wrong. That said, there’s no obvious way to prove that ethics should be consistent over time - one can conceive an ethical equivalent of grue, eg deontoutilitarianism, which states that until a certain point (in time or space or within other dimensions) certain acts are absolutely wrong, and beyond that point maximising utility is all that matters. I suspect the hideous moniker isn’t the only reason I’ve never heard anyone propose such a monstrosity though, so for the rest of this post I’ll assume that the reader requires ethical principles to be universal and unchanging.

2 On originally writing this essay, I knew of no living utilitarians who actually hold the maximising view.

Even the classic writings of Mill and Bentham aren't obviously in favour of it. Bentham's formulation of his principle of utility was 'that principle which approves or disapproves of every action whatsoever according to the tendency it appears to have to augment or diminish the happiness of the party whose interest is in question', which seems ambiguous between satisficing, maximising and even scalar.

Similarly, Mill’s actual statement, in his essay ‘Utilitarianism’, is ‘actions are right in proportion as they tend to promote happiness, wrong as they tend to produce the reverse of happiness’ (emphasis mine). This sounds to me more like scalar utilitarianism as spelled out in the main test immediately after this reference.

Dale Miller argued in an email exchange that the ‘in proportion’ actually modified ‘actions’ rather than 'actions are right' - ie that Mill meant that the proportion of actions that ‘tend to promote happiness’ is the same proportion as ‘actions that are right’, not that some actions can be 'rightish'. I'm not sure how convincing I find this, but you can draw your own conclusion - nothing rides on Mill's opinion other than historical curiosity.

In any case since I wrote this essay in 2015, Peter Singer and Katarzyna de Lazari-Radek explicitly rejected the scalar view in The Point of View of the Universe; so I cannot claim that the maximising view has no backers. Nonetheless, when I informally polled the utilitarianism Facebook group it received no votes from approx 15 respondents, and I've never met anyone who claimed to hold it.

But nonetheless the maximising view is often characterised as being the 'standard' form of utilitarianism, so I include it here since it's a logically coherent form of utilitarianism, strawman or no, and it’s helpful to be able to distinguish it from other points on the sufficiency axis.

3 This is arguably a different way of looking at the utility axis, but since I think for each point on that, the questions you’d ask about a utilon could be quite different, and since I think it’s an underexplored and important part of utilitarianism, I include it here for completeness.

4 Throughout these essays I mean ‘respect’ purely as shorthand for ‘seek to fulfil, all things being equal’.

5 Throughout these essays I mean ‘agent’ purely as shorthand for ‘entity which might have properties the theory in question requires’ (in this case one which might be able to prefer things).

6 I’m basing this claim mainly on Brian Greene’s pop-physics in The Fabric of the Cosmos, so am open to being corrected by actual physicists.

7 I had a relevant exchange with Kaj Sotala and Brian Tomasik suggesting the sort of intuitions one might have here:

Kaj: One option that might fit under C1PU, or possibly as its own option: every preference that can be seen to be logically implied by one of the agent's existing preferences counts.

So to pick a slightly silly example, if I had a preference for having a library near the local school because I valued children having easy access to books, then this would also logically imply that I would value every school on the planet having a library nearby, and that I would also value this to be true for every intelligent species in the universe that had close enough equivalents for "childhood", "school", and "library".

Interestingly, I think that this option becomes equivalent with negative PU, since then each agent has an infinite amount of preferences and the unsatisfied preferences always outnumber the satisfied ones.

Brian: There would be infinitely many satisfied preferences too, like the preference to not have a nail in your foot, not have your head set on fire, to have hands, etc. (except in rare cases).

Me: How could you show logical implication between seemingly analogous situations, especially without reference to mental states? You might prefer a library next to some subset of all schools because your kid goes to that one, because you work for a local library construction firm, or because it will give you a chance to place a copy of the necronomicon on the right leyline that will let you unleash Cthulu on the world.

8 I owe this thought to Simon Rosenqvist. If we think some preferences can’t be either fulfilled or thwarted, this seems equivalent in practice to ignoring them. If we do think being unfulfillable == being thwarted, this seems to have profound implications for PUs, who would then have a strong prescription to minimise false beliefs among the population, even where they do no other damage, to prevent such preferences arising.

9 If you dislike ‘valence utilitarianism’, I would suggest calling this hedonistic rather than preference utilitarianism for three aesthetic reasons:

Firstly, history: hedonistic utilitarianism was, roughly, the form originally conceived by Jeremy Bentham, against which PU was supposed to contrast and improve. If the form you settle on is functionally indistinguishable, it seems ingracious to use a name whose existence implies rejection of his ideas.

Secondly, its lack of necessary external references fits better with the way we normally think of happiness than of preference. Happiness generally stands on its own as a concept that we can conceive without reference to other objects; ie while we might be happy about something, we wouldn’t think it sounded peculiar to just say ‘I am happy’.

By contrast, preferences are almost always grammatically transitive. We would think it peculiar to say any variation on ‘I prefer’ without adding a grammatical object that’s being preferred.

Thirdly, it’s just clearer. If we refer to this as PU, it could confuse the hell out of people wondering what HU actually is, and perpetuate the idea that hedonistic utilitarians are only interested in the sorts of ‘happiness’ that one gets from short-term physical stimulation. Plus, given how imprecise the phrase ‘preference utilitarianism’ turns out to be, when we have a ready-made more precise one, why not use it?

None of this discussion is relevant to its function, of course, hence its relegation to a footnote.

10 Some reviewers complained that I’m unjustifiably ignoring axes unique to HU, but I think this is false. It’s certainly not obviously true: as long as one doesn’t accept C1PU, then any questions about happiness seem likely to to have equivalents regarding conscious preference-satisfaction. They may even wind up being the same questions if (per above) we accept that what we’re actually examining is positive and negative valence.

I’ll revisit this in part 3, but in any case it’s not a claim PUs can assert without example - they need to actually specify what these unique HU axes might be.

11 For example, the SEP entry on consequentialism describes classical utilitarianism as ‘actually a complex combination of many distinct claims’ (then listed), which the author admits ‘claims could be clarified, supplemented, and subdivided further’ - but at no point does he acknowledge that the entire list could change around and still describe the same product.

Thanks to Brian Tomasik, Kaj Sotala, Carl Shulman, Simon Rosenqvist, Lila Reiber, Mihnea Mafti, Peter Hurford, Ruairi Donnelly, Lance Bush and Scott Alexander for, between them, a huge number of helpful comments on this piece.

Continue to Part 2


Preference utilitarianism


theoretical ethics preference utilitarianism