This competition entry has been selected for publication by the Forum team.
I begin my discussion with a reconstruction of what I take to be the core of DiGiovanni's Unawareness Argument (UA) as I understand it. Let me begin one step back from his starting point, with the notion of a consequence-preference (c-preference) more generally defined:
C-preference: x c-prefers φ over υ iffdef x prefers φ wholly or largely in virtue of the possible consequences of those actions.[1]
So, c-preferences are just preferences determined largely by the consequences of whatever kind of an action. Thus, if I prefer giving to the ballet rather than the soup kitchen largely because it makes my peer group think I am sophisticated, my preference is a c-preference in this sense.
Given this notion, we can now inquire into the normative question of when a c-preference (for φ over υ) is altruistically justified (a-justified); that is, justified on the grounds of its altruistic consequences (ignoring non-altruistic consequences). What counts as an "altruistic consequence"? Before answering, let me first define the notion of acting altruistically:
Altruistic Action: x acts altruistically in φ-ing if, and to the extent that, x's purpose in φ-ing is to realize some objective goods for their own sake.[2]
Since acting altruistically, on this view, amounts to aiming at the production of objective goods for their own sake, it is natural to think of a consequence as an altruistic consequence as the production of an objective good. That is:
Altruistic Consequence: Some consequence Δ of x's φ-ing is an altruistic consequence iffdef Δ is an objective good or objective bad.
It might be objected that this definition of an altruistic consequence is too broad. It allows for the possibility that, even if my aims in acting are entirely non-altruistic, my act may nevertheless have many (indeed, only) altruistic consequences. But it might seem that altruistic consequences can only flow from acts that are themselves at least partly altruistic in aim. The definition, however, is stipulative and fits most naturally with the consequentialist background framing of the discussion. So, the correct reading is that an altruistic consequence is the kind of consequence an altruist (qua altruist) might aim at.
Now consider the normative question: When is a c-preference (for φ over υ) a-justified? DiGiovanni's most general (partial) answer appears to be this:
Norm: x's c-preferring φ over υ is a-justified only if x's reason for preferring φ is based on an impartial altruistic comparison of the possible consequences of φ and υ.
Unpacking the notion of an "altruistic comparison" in the vocabulary above we get:
Norm*: x's c-preferring φ over υ is a-justified only if x's reason for preferring φ is based on an impartial comparison of the possible altruistic consequences of φ and υ (viz., a comparison of their production of objective goods).
As stated, Norm* is silent about the scope of the altruistic consequences that must be taken into account or the manner of the comparison beyond impartiality. But the UA can only succeed if these notions are cashed out in a manner that gives the empirical unawareness premise teeth.
And this is what DiGiovanni does by unpacking the comparison in terms of the expected value calculations of idealized agents and the scope across the totality of altruistic consequences. That is, he commits to something like the following: An impartial EV comparison of some possible altruistic consequences of φ and υ a-justifies preferring φ over υ only if we have grounds for thinking that the EV comparison would be vindicated by an idealized aggregation of all possible altruistic consequences. This is a principle of idealized projectability.
IP: x's c-preference for φ over υ is a-justified only if x is epistemically justified in believing that EVI(φ) > EVI(υ), where EVI is an idealized expected value calculation over the totality of consequences.[3]
DiGiovanni offers a compressed argument for IP saying in his summary post that failing this our grounds for c-preferring would be "unacceptably arbitrary", explicitly ruling out heuristics or other similar grounds for the preference.
The remainder of the premises then establish projectability failure. Here, in summary, is the argument:
Two features of this reconstruction will matter in what follows. The first concerns the modal strength of (4). DiGiovanni's case for it does not rest on contingent limitations of our present evidence—on the thought that we happen to lack the relevant models, or that further inquiry might one day supply them. It rests instead on the structural coarseness of any finite agent's representation of the space of possible consequences: the catch-all is not a gap in a map we might eventually fill in, but a permanent feature of maps drawn by agents like us. Accordingly, (4) is asserted for arbitrary φ and υ, and it must be, since the UA is meant to hold generally rather than to report a passing predicament. So read, the unawareness at issue is unremediable rather than merely unremedied. I will hold DiGiovanni to that reading in §3, where it does work he may not welcome.
The second concerns the argument for IP, which is an anti-arbitrariness argument. What is supposed to be wrong with grounds falling short of idealized projectability is not that they are unreliable, nor that they are false, but that they are arbitrary—and unacceptably so. IP is thus offered as the condition that rescues altruistic preference from arbitrariness. Whether it succeeds in that role, or is even required for it, is the question of §3.
In my view, both premises (2) and (3) can be challenged, though I think the case against (2) is stronger and more interesting. In the next section, I will sketch an argument against (3) and then turn in section 3 to the case against premise (2).
Consider premise (3). It says, roughly, that if we are massively unaware of the altruistic consequences of our actions over the long term, we lack sufficient justification (epistemic justification) for believing that the ordering of our actions based on near term EVs will be preserved over the long term. The motivating idea is that the massive number of long term effects of an action are ultimately going to swamp the small number of short term effects on which our near term EVs are based and on which (we may suppose) we have a relatively good handle.
But even if we grant that we are massively unaware of long-term consequences, it doesn't follow that this unawareness automatically severs the evidential connection between near-term and total value.
Let N(φ) and N(υ) be the near-term altruistic value of φ and υ, respectively, let D(φ) and D(υ) be the altruistic value of their more remote downstream consequences, and let T(φ) = N(φ) + D(φ), be the total value of φ (respectively, υ). The corresponding expectations we write EVN, EVD, and EVT. Now suppose our evidence supports the claim that the near-term expectation of φ exceeds that of υ:
EVN(φ) > EVN(υ).
DiGiovanni's epistemic conclusion requires us to accept that, because we are massively unaware of D(φ) and D(υ), we are not justified in believing that EVT(φ) > EVT(υ). But that does not follow from unawareness alone. It also requires the claim that our evidence provides no justified constraint on the comparative expected downstream values of φ and υ, not even a basis for believing that their downstream difference is unlikely to reverse the near-term expected-value ordering. Thus, the argument requires more than establishing unawareness of particular downstream consequences. It requires rejecting any defensible principle of value persistence or subjective non-reversal for these actions.
There is already a substantive discussion of this kind of response to the UA, and I'm inclined to think some versions of it may succeed. But here I want to make a narrower point: It is natural to understand the UA to be making a universal argument, for arbitrary x, φ, and υ. So, consider a principle like non-reversal:
NR: EVN(φ) > EVN(υ) ⇒ EVT(φ) > EVT(υ)
Defending the plausibility of NR for arbitrary actions and actors is going to be a daunting task. But that doesn't mean that specific instances of it can't be plausibly defended. Specifically, we may have reason to think that some classes of actions are more likely to validate it than others.[4] But because DiGiovanni's argument generalizes over actions, it does not take into account features of specific action types that could help justify accepting those instances of NR.
Consider, for instance, education. Education plausibly has a number of features that promote value-persistence including: recursivity (learning to read, e.g., increases the capacity for learning), option expansion (learning facilitates awareness of options[5]), complementarity (learning can serve to enhance the value of available resources), error correction (learning can facilitate awareness of error and corrective steps), and transferability (learning is portable across contexts). These features are not incidental to education, but arguably inherent in it to some degree. The point is not that education has downstream effects. Rather it is that the nature of education predicts downstream effects that are correlated with value production and harm mitigation. And, if that is so, then we have evidence that educational interventions are particularly likely to exhibit value persistence. This is a form of structural non-reversal because the argument for it turns on the specific "structure" of the action in question.
Call action types with a high degree of these kinds of built-in value-productive (or harm mitigating) properties "value-persistent actions" (VPAs, for short). Noting these features is not to be pollyannaish about the potential downstream negative impacts of education or other potential VPAs. In fact, these same structures can themselves amplify negative impacts. Knowledge can be used destructively no less than constructively. So, both bad actors and unforeseen consequences can counteract their value-persistent features. And if the structural features themselves are evaluatively neutral, they seem incapable of supporting any expectation of non-reversal.
This objection would be telling if these structures operated independently of human agency. But they do not. How these features play out downstream is not simply an impersonal unfolding of events but is filtered through the lens of human decision-making and agency that involve both evaluative commitments and powerful incentives to avoid at least many forms of severe disvalue. Even setting altruistic concern aside, agents generally have compelling reasons to preserve their own lives, health, security, and opportunities for flourishing. These motivations create systematic pressure to identify, mitigate, and eliminate the most severe negative consequences. By contrast, there is no corresponding pressure to eliminate beneficial outcomes simply because they are beneficial.
Taken separately, these considerations may be too weak to constitute a convincing defense of non-reversibility. In combination, however, they generate an agency–action structure that produces an asymmetry between good and bad downstream effects. The value-persisting features of VPAs, together with even the selfish and nepotistic motivational features of human agency (bracketing altruistic motivations) can justify belief in specific instances of NR, thereby undercutting the general case for incomparability. Call this structural non-reversibility. In cases of structural non-reversibility, the precise downstream value assignments may remain opaque, while the combined structure of the action and the agency through which its effects unfold nevertheless provides defeasible reason to expect the initial positive value differential to be preferentially preserved rather than reversed.
This also suggests a research program for altruists: identifying the features that genuinely underwrite VPAs and determining the extent to which those features can be strengthened.
Let us now turn our attention back to premise (2):
(If x's c-preferring φ over υ on the basis the calculation EV(φ) > EV(υ) is a-justified) → (x is epistemically justified in believing that EVI(φ) > EVI(υ))
The idea behind the premise is that c-preference is belief-dependent: a purely altruistic preference for one action over another based on the comparative consequences of two actions can be rational or justified only if the agent occupies the right kind of epistemic position with respect to those consequences. Since the preference is purely altruistic, only altruistic considerations can come into play in the comparison. A comparison of idealized expected-value calculations over the totality of consequences would clearly count as an impartial comparison of consequences. So, if x has reason to believe that near term value orderings of altruistic consequences would project to total altruistic consequences, then x would plausibly have a purely altruistic preference. But any way of truncating the comparison appears to be inadequate because any restriction on which consequences are included and which are excluded appears to appeal to some non-altruistic consideration; they are, from the purely altruistic point of view, arbitrary. Consider bracketing. It is not arbitrary, since it rests on a principled epistemic distinction. But the distinction is epistemic rather than altruistic; and so, the objection runs, bracketing imports non-altruistic (epistemically prudential) considerations into the preference. So understood, the question is whether there are any altruistic grounds for truncating in any particular way.
The answer, I believe, is surprisingly easy—Yes. To see why, consider the case of "cosmic ties" in which x has reason to believe that EVI(φ) = EVI(υ). Does it follow that in such cases that there is no purely altruistic reason for preferring φ over υ? Suppose in this case I prefer φ because it was the winning result of a coin toss. Here, I submit, my reasons for preferring φ remain entirely altruistic, including my reasons for accepting the result of the coin toss. I accept that result of the coin toss for the purpose of realizing some objective goods for their own sake. The preference is, of course, arbitrary, but not unacceptably so.
The case prises apart two things the argument for IP runs together. The selection of φ is arbitrary: nothing about φ, and nothing in my reasons, favors it over υ. But the preference is not thereby impure. No non-altruistic consideration has entered it. I did not toss the coin to spare myself the trouble of deliberating, or to secure the appearance of even-handedness before an audience. I tossed it because, having taken up everything altruistic reason had to offer, doing something was what my altruistic purpose required. The arbitrariness while real does not touch on the altruism or rationality of the final decision.
If that is right, the argument for IP rests on a mistaken criterion of altruistic purity. The criterion it requires is one about the content of considerations: a preference is purely altruistic just in case each consideration on which it rests is itself an altruistic consideration, one concerning the production of objective goods. Epistemic considerations, being about what can be known rather than about what is good, fail that test; so any truncation appealing to them is impure. But the conception of altruism with which we began supplies a different criterion, and one that fits the coin-toss case:
Purity: A consideration C figures altruistically in x's preferring φ over υ if, and to the extent that, x's taking up C is in the service of x's purpose of realizing objective goods for their own sake.[6]
On this criterion what makes a consideration altruistically deployed is not its subject matter but how it figures into an agent's goals and intentions. This is not a concession to save the case, but what altruistic action is—a matter of what an agent aims at. Altruistic action is a matter of the integrity of the aim, not of the pedigree of each consideration the aim recruits.
The content-based criterion is in any case untenable on its own terms. An altruist who verifies that the organization she is about to fund exists, that her donation will reach it, and that the intervention is one she is positioned to assess deploys epistemic and merely prudential considerations at every step. On the content-based criterion she is to that extent less purely altruistic than a donor who consults none of these things, which makes the ideally pure altruist the one who never checks anything. Epistemic and instrumental considerations do not compete with an altruistic aim when they are enlisted by it; they are how a finite agent pursues any aim at all. So bracketing needs no apology. It is arbitrary in the harmless sense that it draws a line that altruistic considerations alone do not draw. The altruist brackets in order to act on the goods she can actually aim at.
Premise (2) now faces a dilemma. Either the condition it imposes is the anti-arbitrariness condition that motivates IP, or it is a bare epistemic condition detached from that motivation.
Consider the first horn. If what a-justification requires is that the agent's grounds not be arbitrary, then cosmic ties are a counterexample, since there the grounds are arbitrary and the preference is a-justified all the same. Nor is this a fringe case that some refinement might exclude. The anti-arbitrariness thought derives whatever plausibility it has from cases in which the arbitrary ground displaces altruistic reasoning in favor of something else. Ties are the pure case of an arbitrary ground that displaces nothing and the objection lapses.
So, consider the second horn. If (2) imposes an epistemic condition standing on its own—x must be justified in believing that EVI(φ) > EVI(υ), and that is that—then the tie case leaves it untouched, since in a tie the condition is met. But it has lost its argumentative support. The reason offered for IP was the anti-arbitrariness reason, and the first horn shows that reason does not support it. What remains is a bare assertion.
Still, the second horn deserves better than the charge of being unmotivated, because a natural thought stands behind it. The thought is that ties are special. In a tie the epistemic condition of (2) is satisfied: the agent knows how the idealized comparison comes out, and her coin toss operates after a completed altruistic assessment has been undertaken; it is harmless precisely because nothing of altruistic relevance turns on it. The unaware altruist is not in that position. Her assessment is not complete but interrupted, and her toss substitutes for altruistic guidance rather than following it. So (2) might fail as a fully general claim while holding of every case the UA is actually about.
But what, in the case of cosmic ties, does the licensing? The obvious answer is the equality of the idealized expectations. I think that answer is wrong. To see why, compare the following cases.
First, consider an agent who knows nothing about how φ and υ compare and has made no attempt to find out. She has not compared their near-term expectations; she has not asked whether either organization is solvent; she has not considered whether she is placed to assess either intervention. She tosses a coin. Her preference is clearly not a-justified. But what disqualifies her is not the bare absence of justified belief about EVI. Rather, this agent has declined guidance that was there for the taking. Her toss displaces altruistic reasoning, and that, rather than her ignorance of the totality, is the defect.
Now consider an agent who has done all of it. She has compared near-term expectations and found no difference she can credit; she has checked what could be checked; she has attended to whatever features of the two actions bear on value persistence in the sense of §2, and found no asymmetry. She remains unaware of the totality. She tosses a coin. Her situation differs from the tie case in what she knows and resembles it in what she has done, and it is the second that our verdicts track. Nothing distinguishes her from the tie-case agent in respect of her responsiveness to altruistic reasons. The difference lies entirely in how much altruistic reason could provide, and this is not up to her.
What licenses the toss, then, is not that the altruistic guidance came out even but that it ran out. Equality is one route to exhaustion; it is not what exhaustion consists in. And once the licensing condition is identified as exhaustion, the extension from ties to unawareness goes through, provided the guidance really has run out rather than merely been left unconsulted. Which is where §1's reading of (4) is cashed out. On DiGiovanni's own understanding the unawareness is structural: it is not that the idealized comparison is available and unsought, but that agents constituted as we are cannot reach it. Guidance no agent of the relevant kind can reach is not guidance the agent is epistemically accountable for. It is therefore no part of what she is required to take up, and its absence leaves her exactly where the tie-case agent stands, namely, at the end of what altruistic reason can say, with a choice still to make.
DiGiovanni can resist this only by weakening (4). If the unawareness arises from a failure to seek what is within our grasp, then the guidance has not run out, and the extension fails. But then the UA loses its generality: it becomes a report on a predicament we might grow out of, and (4) can no longer be asserted for arbitrary φ and υ. The modal strength that makes the argument interesting is the modal strength that makes it self-undermining.
If the foregoing is right, IP is not the condition that saves altruistic preference from arbitrariness, and premise (2) is false. Something needs to take its place. Here is a suggestion:
EX: x's c-preferring φ over υ is a-justified only if (i) x's preference is formed on the basis of the altruistic guidance reachable by x, and (ii) x's grounds for preferring φ (including whatever settles any residue left by (i)) are taken up in the service of x's purpose of realizing objective goods for their own sake.
EX is epistemically demanding where it matters and permissive only where epistemic access runs out. It rules out the donor of §1 who prefers the ballet because it makes his peer group think him sophisticated: his preference may be a c-preference, but it fails (ii), and no amount of information about consequences will repair it. It rules out the agent who tosses a coin without looking, who fails (i). And it does not license tossing a coin between feeding people and burning money, since there the near-term comparison is both reachable and decisive.
Clause (i) is where almost all real altruistic deliberation happens: comparison of near-term expectations, local evaluative judgment about the goods at stake, attention to tractability and to whether one is placed to assess the intervention at all, and attention to the structural features canvassed in §2. This is the point of contact between the two arguments of this paper. The features that make an action value-persistent are reachable in exactly the way EVI is not: they are features of the action's structure and of the agency through which its effects unfold, and an altruist can find out about them. Where §2 argued that such features can support belief in particular instances of non-reversal, §3 argues that even their absence leaves the altruist justified; both locate the grounds of altruistic preference among things an altruist can get at.
EX also assigns near-term expected value a role that DiGiovanni's framing obscures. On his account the near-term calculation matters only as evidence about the limit, so that if the inference to the limit fails, the calculation is idle. This is why the UA can look like a wholesale indictment of expected-value reasoning in altruistic contexts. On EX it matters directly: it is part of the guidance that must be taken up, and taking it up is a condition on a-justification whether or not anything about the totality follows.
It may be objected that EX is too easily satisfied, since the residue left by (i) will typically be vast, and EX then permits settling a vast question by a coin toss. But the size of the residue is not the agent's doing, and a condition met only by agents for whom the residue is small is a condition finite agency cannot satisfy. Nor does EX say that anything goes once the reachable is exhausted: what settles the residue must itself be adopted in the service of the altruistic aim, so an agent who settles it by consulting her own convenience, or her standing in her community, violates (ii) having satisfied (i).
The general moral is that there is no special epistemology of altruism. The unaware altruist is not in a distinctive predicament requiring distinctive normative treatment; she is a finite agent acting under partial information, subject to the requirement that governs finite agency everywhere. What the UA presents as a gap between altruistic rationality and our epistemic position is better understood as a gap between altruistic rationality and a theory of what it demands built for agents we are not.
Two hedges. EX turns on a notion of reachability that it leaves unanalyzed; reachability is agent-relative, and an account is owed of how the boundary is drawn without letting it slide toward whatever the agent happens to have thought of.[7] And EX is silent on which objective goods an altruist should aim at, and so on the positive metric of altruistic rationality. A fuller account of that metric is available, I think, and belongs in a treatment of altruistic rationality as a whole; but nothing above turns on it, and I set it aside.
DiGiovanni's data are real. We are massively unaware of the downstream consequences of what we do, and the idealized comparison over the totality of consequences is out of reach. The question is what those facts show. On the reading defended here they do not show that altruistic preference is unjustified; they show that a condition requiring the idealized comparison cannot be the condition on altruistic justification, since finite agency cannot meet it while altruistic reason makes demands on finite agents nonetheless.
Two conclusions, thus, are offered of different strengths. Section 2 argued that the projection claim is not uniformly hopeless: for value-persistent actions, the structure of the action together with the agency through which its effects unfold supplies defeasible reason to expect the near-term ordering to be preserved rather than reversed. Section 3 argued that even where projection fails altogether, altruistic preference can be altruistically justified. Neither conclusion requires optimism about the long run. Both require only that altruistic rationality be a standard that finite altruists are able to rationally meet.
On this view, therefore, an action counts as altruistic in virtue of its aims, not in virtue of its consequences.
On this view, therefore, an action counts as altruistic in virtue of its aims, not in virtue of its consequences.
Of course, DiGiovanni is not saying we have to have the idealized EV calculations, but only that we are justified in believing that the value-ordering would hold for them. I will, in general, gloss over this caveat for simplicity unless it is needed.
An additional issue that needs careful consideration is how finely-individuated evaluated actions are. Yoaav Isaacs in discussing the structurally similar debate around transformative experience notes that transformative problems become more or less pervasive depending upon how finely possible outcomes are individuated ("The problems of transformative experience", Philosophical Studies, 177, 1065-1084: 2020).
The canonical statement of this is Mill's discussion in On Liberty of the epistemic liberties.
Purity is a claim about what makes a consideration altruistically deployed, not about degrees of altruistic motivation more generally; it inherits the "to the extent that" clause from Altruistic Action, so an agent whose deliberation serves mixed purposes counts as altruistic in proportion to the share her altruistic purpose has in it.
The obvious constraint is that reachability be indexed to what the agent is in a position to find out rather than to what she has in fact considered, so that negligent ignorance does not shrink the domain of guidance she is required to take up. Spelling that out is a substantive task, and it is the same task that any account of epistemic obligation faces; that it is not special to altruism is, on the present view, the expected result.
Ok so you've got two arguments. First:
Maybe, but what makes you think that this value persistence is net positive, considering long-term effects? Education (your example) may benefit AI accelerationists as much as, or more than, AI safety researchers. It may make humans better (for better or for worse) at creating digital minds, colonizing space, spreading wild animal suffering, or intensifying the farming of small animals off-Earth. We can't just assume education is going to solve all the problems you're worried about, rather than worsen them. What makes you think cluelessness does not bite here if it bites everywhere else?
Then, I don't fully understand your second argument (section 3) or how it rescues impartial action guidance. But it seems to take it for granted that DiGiovanni's P2a is false.
Hey Jim, Thanks for the comment.
RE your first comment, there are two points. Suppose, first, that VPAs merely transmit value forward without being sensitive to its valence (good or bad). This is the worst-case scenario for my proposal. My thought was two part: (i) even in the worst-case scenario, the downstream consequences aren't merely causal but are filtered through human agency; (2) human agency even for non-altruists has an inherent bias against eventuating the absolutely worst consequences of our actions even if we just look narrowly at things like self-interest, nepotism, and the like. So, by and large, if EV(A) > EV(B) near term, value-persistence + bias provides some defeasible reason for thinking EV(A) > EV(B) long term. The magnitude of that effect will be susceptible to the cases. But this is why I suggested that there is a research program here: What can we do as actors to maximize these features so the bias is strongest?
But (second point) let me also note something that didn't come out in the paper. Consider the education case: If one thinks of education in a coarse-grained way (e.g., merely imparting information to people), then the consequences you mention are live alternatives. But if one thinks in a fine-grained manner about the nature of the education, then we have reason to think its value tends to be positive and that this positive value persists. In other words, education isn't simply value neutral. For example, explicitly educating about things like intellectual virtues (epistemic humility, non-dogmatism, etc) and rational biases (confirmation bias, recency effects, etc.) might themselves help ensure positive downstream value. How that plays out in other cases, isn't something I have thought through. But, again, that was why I suggested a research program here, using education as illustrative.
The second argument is aimed at how one articulates the rationality of belief-dependent action, which is what DiGiovanni's normative premise turns on. As stated, this is a really strong premise, implausibly strong (I think) if one were just articulating a principle of rational action per se (not altruistic rational action). So, the goal was, first, to try to charitably explain why the strong principle might be reasonable when restricting to "impartial, altruistic" action. The argument was then designed to show that neither the "impartial" or "altruistic" side of that restriction warrants the strong, normative principle. One can act as an impartial altruist in spite of cluelessness (i.e., without justifiably believing that "if we were idealized agents who could aggregate all of A’s and B’s possible consequences into literal EVs, then we’d say A has higher EV"). I don't think that argument presupposes that P2a is false (Against precise EVs: We shouldn’t represent actions’ degrees of c-preferability with literal precise expected values). But it can be construed as an argument that impartial altruism doesn't require precise long-term EVs, even if (we can grant) it requires precise short term (reachable) EVs.
Hope that helps clarify!