Christian Tarsney

The standard case for longtermism focuses on a small set of risks to the far future, and argues that in a small set of choice situations, the present marginal value of mitigating those risks is very great. But many longtermists are attracted to, and many critics of longtermism worried by, a farther-reaching form of longtermism. According to this farther-reaching form, there are many ways of improving the far future, which determine the value of our options in all or nearly all choice situations, and will continue to do so over the coming decades even if we make substantial investments in longtermist priorities. This chapter highlights the gap between the minimal form of longtermism established by standard arguments and this more expansive view, and considers (without reaching any firm conclusions) which form of longtermism is more plausible.

Researchers worried about catastrophic risks from advanced AI have argued that we should expect sufficiently capable AI agents to pursue power over humanity because power is a convergent instrumental goal, something that is useful for a wide range of final goals. Others have recently expressed skepticism of these claims. This paper aims to formalize the concepts of instrumental convergence and power-seeking in an abstract, decision-theoretic framework, and to assess the claim that power is a convergent instrumental goal. I conclude that this claim contains at least an element of truth, but might turn out to have limited predictive utility, since an agent’s options cannot always be ranked in terms of power in the absence of substantive information about the agent’s final goals. However, the fact of instrumental convergence is more predictive for agents who have a good shot at attaining absolute or near-absolute power.

Expected value maximization gives plausible guidance for moral decision‐making under uncertainty in many situations. But it has unappetizing implications in ‘Pascalian’ situations involving tiny probabilities of extreme outcomes. This paper shows, first, that under realistic levels of ‘background uncertainty’ about sources of value independent of one's present choice, a widely accepted and apparently innocuous principle—stochastic dominance—requires that prospects be ranked by the expected value of their consequences in most ordinary choice situations. But second, this implication does not hold when differences in expected value are driven by tiny probabilities of extreme outcomes. Stochastic dominance therefore lets us draw a surprisingly principled line between ‘ordinary’ and ‘Pascalian’ situations, providing a powerful justification for de facto expected value maximization in the former context while permitting deviations in the latter. Drawing this distinction is incompatible with an in‐principle commitment to maximizing expected value, but does not require too much departure from decision‐theoretic orthodoxy: it is compatible, for instance, with the view that moral agents must maximize the expectation of a utility function that is an increasing function of moral value.

Large language models now possess human-level linguistic abilities in many contexts. This raises the concern that they can be used to deceive and manipulate on unprecedented scales, for instance spreading political misinformation on social media. In future, agentic AI systems might also deceive and manipulate humans for their own purposes. In this paper, first, I argue that AI-generated content should be subject to stricter standards against deception and manipulation than we ordinarily apply to humans. Second, I offer new characterizations of AI deception and manipulation meant to support such standards, according to which a statement is deceptive (resp. manipulative) if it leads human addressees away from the beliefs (resp. choices) they would endorse under “semi-ideal” conditions. Third, I propose two measures to guard against AI deception and manipulation, inspired by this characterization: “extreme transparency” requirements for AI-generated content and “defensive systems” that, among other things, annotate AI-generated statements with contextualizing information. Finally, I consider to what extent these measures can protect against deceptive behavior in future, agentic AI systems.

Is the overall value of a world just the sum of values contributed by each value-bearing entity in that world? Additively separable axiologies (like total utilitarianism, prioritarianism, and critical level views) say 'yes', but non-additive axiologies (like average utilitarianism, rank-discounted utilitarianism, and variable value views) say 'no'. This distinction appears to be practically important: among other things, additive axiologies generally assign great importance to large changes in population size, and therefore tend to strongly prioritize the long-term survival of humanity over the interests of the present generation. Non-additive axiologies, on the other hand, need not assign great importance to large changes in population size. We show, however, that when there is a large enough `background population' unaffected by our choices, a wide range of non-additive axiologies converge in their implications with additive axiologies—for instance, average utilitarianism converges with critical-level utilitarianism and various egalitarian theories converge with prioritarianism. We further argue that real-world background populations may be large enough to make these limit results practically significant. This means that arguments from the scale of potential future populations for the astronomical importance of avoiding existential catastrophe, and other arguments in practical ethics that seem to presuppose additive separability, may succeed in practice whether or not we accept additive separability as a basic axiological principle.

How should you decide what to do when you’re uncertain about basic normative principles? A natural suggestion is to follow some "second-order:" norm: e.g., obey the most probable norm or maximize expected choiceworthiness. But what if you’re uncertain about second-order norms too—must you then invoke some third-order norm? If so, any norm-guided response to normative uncertainty appears doomed to a vicious regress. This paper aims to rescue second-order norms from the threat of regress. I first elaborate and defend the claim some philosophers have made that the regress problem forces us to accept normative externalism, the view that at least one norm is incumbent on all agents regardless of their normative beliefs. But, I then argue, we need not accept externalism about first-order norms, thus closing off any question of how agents should respond to normative uncertainty. Rather, we can head off the threat of regress by ascribing external force to a single second-order norm: the enkratic principle.

This article presents a new argument against many forms of moral and prudential value incompleteness. The argument relies on two central principles: (i) a weak "negative dominance" principle, to the effect that Lottery 1 is better than Lottery 2 only if some possible outcome of Lottery 1 is better than some possible outcome of Lottery 2, and (ii) a weak form of ex ante Pareto, to the effect that, if Lottery 1 gives an unambiguously better (stochastically dominant) prospect to some individuals than Lottery 2, and equally good prospects to everyone else, then Lottery 1 is better than Lottery 2. Given modest auxiliary assumptions, these two principles rule out incompleteness in the prudential ranking of individual lives, and many forms of incompleteness in the moral rankings of outcomes and lotteries.

Should you be willing to forego any sure good for a tiny probability of a vastly greater good? Fanatics say you should, anti‐fanatics say you should not. Anti‐fanaticism has great intuitive appeal. But, I argue, these intuitions are untenable, because satisfying them in their full generality is incompatible with three very plausible principles: acyclicity, a minimal dominance principle, and the principle that any outcome can be made better or worse. This argument against anti‐fanaticism can be turned into a positive argument for a weak version of fanaticism, but only from significantly more contentious premises. In combination, these facts suggest that those who find fanaticism counterintuitive should favor not anti‐fanaticism, but an intermediate position that permits agents to have incomplete preferences that are neither fanatical nor anti‐fanatical.

Grill (2023) defends the sum of averages view (SAV), on which the value of a population is found by summing the average welfare of each generation or birth cohort. A major advantage of SAV, according to Grill, is that it escapes the Egyptology objection to average utilitarianism. But, we argue, SAV escapes only the most literal understanding of this objection, since it still allows the value of adding a life to depend on facts about other, intuitively irrelevant lives. Moreover, SAV has a decisive drawback not shared with either average or total utilitarianism: it can evaluate an outcome in which every individual is worse off as better overall, even when exactly the same people exist in both outcomes. These problems, we argue, afflict not only Grill's view but any view that uses a sum of subpopulation averages, apart from the limiting cases of average and total utilitarianism.

Longtermists claim that what we ought to do is mainly determined by how our actions might affect the very long-run future. A natural objection to longtermism is that these effects may be nearly impossible to predict — perhaps so close to impossible that, despite the astronomical importance of the far future, the expected value of our present actions is mainly determined by near-term considerations. This paper aims to precisify and evaluate one version of this epistemic objection to longtermism. To that end, I develop two simple models for comparing ‘longtermist’ and ‘neartermist’ interventions, incorporating the idea that it is harder to make a predictable difference to the further future. These models yield mixed conclusions: if we simply aim to maximize expected value, and don’t mind premising our choices on minuscule probabilities of astronomical payoffs, the case for longtermism looks robust. But on some prima facie plausible empirical worldviews, the expectational superiority of longtermist interventions depends heavily on these ‘Pascalian’ probabilities. So the case for longtermism may depend either on plausible but non-obvious empirical claims or on a tolerance for Pascalian fanaticism.

The case for longtermism depends on the vast potential scale of the future. But that same vastness also threatens to undermine the case for longtermism: If the universe as a whole, or the future in particular, contain infinite quantities of value and/or disvalue, then many of the theories of value that support longtermism (e.g., risk-neutral total utilitarianism) seem to imply that none of our available options are better than any other. If so, then even apparently vast effects on the far future cannot in fact make the world morally better. On top of this, some strategies for avoiding this problem of “infinitarian paralysis” (e.g., exponential pure time discounting) yield views that are much less supportive of longtermism. In this chapter, we explore how the potential infinitude of the future affects the case for longtermism. We argue that (i) there are reasonable prospects for extending risk-neutral totalism and similar views to infinite contexts and (ii) many such extension strategies will still support the case for longtermism, since they imply that when we can only effect (or only predictably affect) a finite, bounded part of an infinite universe, we can ignore the unaffectable rest of the universe and reason as if the finite, affectable part were all there is.

All else being equal, most of us typically prefer to have positive experiences in the future rather than the past and negative experiences in the past rather than the future. Recent empirical evidence tends not only to support the idea that people have these preferences, but further, that people tend to prefer more painful experiences in their past rather than fewer in their future (and mutatis mutandis for pleasant experiences). Are such preferences rationally permissible, or are they, as time-neutralists contend, rationally impermissible? And what is it that grounds their having the normative status that they do have? We consider two sorts of arguments regarding the normative status of future-biased preferences. The first appeals to the supposed arbitrariness of these preferences, and the second appeals to their upshot. We evaluate these arguments in light of the recent empirical research on future-bias.

People are ‘biased toward the future’: all else being equal, we typically prefer to have positive experiences in the future, and negative experiences in the past. Several explanations have been suggested for this pattern of preferences. Adjudicating among these explanations can, among other things, shed light on the rationality of future-bias: For instance, if our preferences are explained by unjustified beliefs or an illusory phenomenology, we might conclude that they are irrational. This paper investigates one hypothesis, according to which future-bias is explained by our having a phenomenology that we describe, or conceive of, as being as of time robustly passing. We empirically tested this hypothesis and found no evidence in its favour. Our results present a puzzle, however, when compared with the results of an earlier study. We conclude that although robust passage phenomenology on its own probably does not explain future-bias, having this phenomenology and taking it to be veridical may contribute to future-bias.

Empirical work has lately confirmed what many philosophers have taken to be true: people are ‘biased toward the future’. All else being equal, we usually prefer to have positive experiences in the future, and negative experiences in the past. According to one hypothesis, the temporal metaphysics hypothesis, future-bias is explained either by our beliefs about temporal metaphysics—the temporal belief hypothesis—or alternatively by our temporal phenomenology—the temporal phenomenology hypothesis. We empirically investigate a particular version of the temporal belief hypothesis according to which future-bias is explained by the belief that time robustly passes. Our results do not match the apparent predictions of this hypothesis, and so provide evidence against it. But we also find that people give more future-biased responses when asked to simulate a belief in robust passage. We take this to suggest that the phenomenology that attends simulation of that belief may be partially responsible for future-bias, and we examine the implications of these results for debates about the rationality of future-bias.

Is the overall value of a world just the sum of values contributed by each value-bearing entity in that world? Additively separable axiologies (like total utilitarianism, prioritarianism, and critical level views) say 'yes', but non-additive axiologies (like average utilitarianism, rank-discounted utilitarianism, and variable value views) say 'no'. This distinction is practically important: additive axiologies support 'arguments from astronomical scale' which suggest (among other things) that it is overwhelmingly important for humanity to avoid premature extinction and ensure the existence of a large future population, while non-additive axiologies need not. We show, however, that when there is a large enough 'background population' unaffected by our choices, a wide range of non-additive axiologies converge in their implications with some additive axiology -- for instance, average utilitarianism converges to critical-level utilitarianism and various egalitarian theories converge to prioritiarianism. We further argue that real-world background populations may be large enough to make these limit results practically significant. This means that arguments from astronomical scale, and other arguments in practical ethics that seem to presuppose additive separability, may be truth-preserving in practice whether or not we accept additive separability as a basic axiological principle.

Philosophers have long noted, and empirical psychology has lately confirmed, that most people are “biased toward the future”: we prefer to have positive experiences in the future, and negative experiences in the past. At least two explanations have been offered for this bias: belief in temporal passage and the practical irrelevance of the past resulting from our inability to influence past events. We set out to test the latter explanation. In a large survey, we find that participants exhibit significantly less future bias when asked to consider scenarios where they can affect their own past experiences. This supports the “practical irrelevance” explanation of future bias. It also suggests that future bias is not an inflexible preference hardwired by evolution, but results from a more general disposition to “accept the things we cannot change”. However, participants still exhibited substantial future bias in scenarios in which they could affect the past, leaving room for complementary explanations. Beyond the main finding, our results also indicate that future bias is stake-sensitive and that participants endorse the normative correctness of their future-biased preferences and choices. In combination, these results shed light on philosophical debates over the rationality of future bias, suggesting that it may be a rational response to empirical realities rather than a brute, arational disposition.

Decision-making under normative uncertainty requires an agent to aggregate the assessments of options given by rival normative theories into a single assessment that tells her what to do in light of her uncertainty. But what if the assessments of rival theories differ not just in their content but in their structure -- e.g., some are merely ordinal while others are cardinal? This paper describes and evaluates three general approaches to this "problem of structural diversity": structural enrichment, structural depletion, and multi-stage aggregation. All three approaches have notable drawbacks, but I tentatively defend multi-stage aggregation as least bad of the three.

Even if I think it very likely that some morally good act is supererogatory rather than obligatory, I may nonetheless be rationally required to perform that act. This claim follows from an apparently straightforward dominance argument, which parallels Jacob Ross's argument for 'rejecting' moral nihilism. These arguments face analogous pairs of objections that illustrate general challenges for dominance reasoning under normative uncertainty, but (I argue) these objections can be largely overcome. This has practical consequences for the ethics of philanthropy -- in particular, it means that donors are often rationally required to maximize the positive impact of their donations.

ABSTRACT Average utilitarianism and several related axiologies, when paired with the standard expectational theory of decision-making under risk and with reasonable empirical credences, can find their practical prescriptions overwhelmingly determined by the minuscule probability that the agent assigns to solipsism—that is, to the hypothesis that there is only one welfare subject in the world, namely, herself. This either (i) constitutes a reductio of these axiologies, (ii) suggests that they require bespoke decision theories, or (iii) furnishes an unexpected argument for ethical egoism.

Longtermists claim that what we ought to do is mainly determined by how our actions might affect the very long-run future. A natural objection to longtermism is that these effects may be nearly impossible to predict -- perhaps so close to impossible that, despite the astronomical importance of the far future, the expected value of our present actions is mainly determined by near-term considerations. This paper aims to precisify and evaluate one version of this epistemic objection to longtermism. To that end, I develop two simple models for comparing "longtermist" and "neartermist" interventions, incorporating the idea that it is harder to make a predictable difference to the further future. These models yield mixed conclusions: if we simply aim to maximize expected value, and don't mind premising our choices on minuscule probabilities of astronomical payoffs, the case for longtermism looks robust. But on some prima facie plausible empirical worldviews, the expectational superiority of longtermist interventions depends heavily on these "Pascalian" probabilities. So the case for longtermism may depend either on plausible but non-obvious empirical claims or on a tolerance for Pascalian fanaticism.

How should you decide what to do when you're uncertain about basic normative principles (e.g., Kantianism vs. utilitarianism)? A natural suggestion is to follow some "second-order" norm: e.g., "comply with the first-order norm you regard as most probable" or "maximize expected choiceworthiness". But what if you're uncertain about second-order norms too -- must you then invoke some third-order norm? If so, it seems that any norm-guided response to normative uncertainty is doomed to a vicious regress. In this paper, I aim to rescue second-order norms from this threat of regress. I first elaborate and defend the suggestion some philosophers have entertained that the regress problem forces us to accept normative externalism, the view that at least one norm is incumbent on agents regardless of their beliefs or evidence concerning that norm. But, I then argue, we need not accept externalism about first-order (e.g., moral) norms, thus closing off any question of what an agent should do in light of her normative beliefs. Rather, it is more plausible to ascribe external force to a single, second-order rational norm: the enkratic principle, correctly formulated. This modest form of externalism, I argue, is both intrinsically well-motivated and sufficient to head off the threat of regress.

In ‘Normative Uncertainty as a Voting Problem’, William MacAskill argues that positive credence in ordinal-structured or intertheoretically incomparable normative theories does not prevent an agent from rationally accounting for her normative uncertainties in practical deliberation. Rather, such an agent can aggregate the theories in which she has positive credence by methods borrowed from voting theory—specifically, MacAskill suggests, by a kind of weighted Borda count. The appeal to voting methods opens up a promising new avenue for theories of rational choice under normative uncertainty. The Borda rule, however, is open to at least two serious objections. First, it seems implicitly to ‘cardinalize’ ordinal theories, and so does not fully face up to the problem of merely ordinal theories. Second, the Borda rule faces a problem of option individuation. MacAskill attempts to solve this problem by invoking a measure on the set of practical options. But it is unclear that there is any natural way of defining such a measure that will not make the output of the Borda rule implausibly sensitive to irrelevant empirical features of decision-situations. After developing these objections, I suggest an alternative: the McKelvey uncovered set, a Condorcet method that selects all and only the maximal options under a strong pairwise defeat relation. This decision rule has several advantages over Borda and mostly avoids the force of MacAskill’s objection to Condorcet methods in general.

The principle that rational agents should maximize expected utility or choiceworthiness is intuitively plausible in many ordinary cases of decision-making under uncertainty. But it is less plausible in cases of extreme, low-probability risk (like Pascal's Mugging), and intolerably paradoxical in cases like the St. Petersburg and Pasadena games. In this paper I show that, under certain conditions, stochastic dominance reasoning can capture most of the plausible implications of expectational reasoning while avoiding most of its pitfalls. Specifically, given sufficient background uncertainty about the choiceworthiness of one's options, many expectation-maximizing gambles that do not stochastically dominate their alternatives "in a vacuum" become stochastically dominant in virtue of that background uncertainty. But, even under these conditions, stochastic dominance will not require agents to accept options whose expectational superiority depends on sufficiently small probabilities of extreme payoffs. The sort of background uncertainty on which these results depend looks unavoidable for any agent who measures the choiceworthiness of her options in part by the total amount of value in the resulting world. At least for such agents, then, stochastic dominance offers a plausible general principle of choice under uncertainty that can explain more of the apparent rational constraints on such choices than has previously been recognized.

In the growing literature on decision-making under moral uncertainty, a number of skeptics have argued that there is an insuperable barrier to rational "hedging" for the risk of moral error, namely the apparent incomparability of moral reasons given by rival theories like Kantianism and utilitarianism. Various general theories of intertheoretic value comparison have been proposed to meet this objection, but each suffers from apparently fatal flaws. In this paper, I propose a more modest approach that aims to identify classes of moral theories that share common principles strong enough to establish bases for intertheoretic comparison. I show that, contra the claims of skeptics, there are often rationally perspicuous grounds for precise, quantitative value comparisons within such classes. In light of this fact, I argue, the existence of some apparent incomparabilities between widely divergent moral theories cannot serve as a general argument against hedging for one's moral uncertainties.

Defenders of deontological constraints in normative ethics face a challenge: how should an agent decide what to do when she is uncertain whether some course of action would violate a constraint? The most common response to this challenge has been to defend a threshold principle on which it is subjectively permissible to act iff the agent's credence that her action would be constraint-violating is below some threshold t. But the threshold approach seems arbitrary and unmotivated: what would possibly determine where the threshold should be set, and why should there be any precise threshold at all? Threshold views also seem to violate ought agglomeration, since a pair of actions each of which is below the threshold for acceptable moral risk can, in combination, exceed that threshold. In this paper, I argue that stochastic dominance reasoning can vindicate and lend rigor to the threshold approach: given characteristically deontological assumptions about the moral value of acts, it turns out that morally safe options will stochastically dominate morally risky alternatives when and only when the likelihood that the risky option violates a moral constraint is greater than some precisely definable threshold (in the simplest case, .5). I also show how, in combination with the observation that deontological moral evaluation is relativized to particular choice situations, this approach can overcome the agglomeration problem. This allows the deontologist to give a precise and well-motivated response to the problem of uncertainty.

How should an agent decide what to do when she is uncertain not just about morally relevant empirical matters, like the consequences of some course of action, but about the basic principles of morality itself? This question has only recently been taken up in a systematic way by philosophers. Advocates of moral hedging claim that an agent should weigh the reasons put forward by each moral theory in which she has positive credence, considering both the likelihood that that theory is true and the strength of the reasons it posits. The view that it is sometimes rational to hedge for one's moral uncertainties, however, has recently come under attack both from those who believe that an agent should always be guided by the dictates of the single moral theory she deems most probable and from those who believe that an agent's moral beliefs are simply irrelevant to what she ought to do. Among the many objections to hedging that have been pressed in the recent literature is the worry that there is no non-arbitrary way of making the intertheoretic comparisons of moral value necessary to aggregate the value assignments of rival moral theories into a single ranking of an agent's options. This dissertation has two principal objectives: First, I argue that, contra these recent objections, an agent's moral beliefs and uncertainties are relevant to what she rationally ought to do, and more particularly, that agents are at least sometimes rationally required to hedge for their moral uncertainties. My principal argument for these claims appeals to the enkratic conception of rationality, according to which the requirements of practical rationality derive from an agent's beliefs about the objective, desire-independent value or choiceworthiness of her options. Second, I outline a new general theory of rational choice under moral uncertainty. Central to this theory is the idea of content-based aggregation, that the principles according to which an agent should compare and aggregate rival moral theories are grounded in the content of those theories themselves, including not only their value assignments but also the metaethical and other non-surface-level propositions that underlie, justify, or explain those value assignments.