II. SUCCESS-FAILURE PAIRS

Arguments for the orthodox view normally involve a pair of cases of a characteristic sort, and mine will be no exception. I call these pairs success-failure pairs. Success-failure pairs are supposed to make it salient that in treating successful and unsuccessful offenders unequally, we afford significance to a feature that, we should recognize, has no significance at all. The first member of the pair, the success case, describes the action of an agent, Alvarez, for which Alvarez is indisputably to blame. It might be that Alvarez kills her neighbor by shooting at her from the attic or burns down her enemy's house by throwing an incendiary bomb—the details of the case are unimportant; you may bring in those of your choice. For convenience, I use the variable α for the type of action that best describes Alvarez's doing in this context.

The only crucial feature to observe in devising the success case is that there must be some room for failure in the agent's α-ing. The α-ing must ultimately depend on conditions that are, as it were, “up to nature.”Footnote ⁶ Human actions in the actual world are always like that, so any realistic case will do. The reason for this condition is that the failure case, the second case of the pair, is to fall within the space of Alvarez's likely failure—the failure case actualizes a possible but unactualized failure of the success case.

The agent in the failure case, whom I will call Borges, is identical to Alvarez in every respect, displays the same skills, beliefs, and intentions as Alvarez does, engages in an identical course of action that is as likely to succeed as his counterpart's is, but unlike Alvarez, Borges fails. The particular conditions of α-ing that the agents neglect are again unimportant: the neighbor's window might happen to have bullet-proof glass, or an unexpected passing bird might deviate the shot; a sudden rainstorm might kill the fire, or a team of firefighters might happen to be partying in the house next door, with the fire truck and equipment. The pair of cases must be such, in other words, that Borges does everything Alvarez does except α-ing, and Borges's failure cannot be explained as the consequence of something one agent has done or omitted to do which the other has not. To simplify, I use the variable “β” for the type of action that best describes Borges's doing. Both Alvarez and Borges thus do β, but Alvarez's β-ing becomes an α-ing, while Borges's does not, in virtue of some nonactional difference between the cases.

Let me clarify a bit more what the relationship is between the types of actions (α and β) for whose instantiations in success and failure the agents (Alvarez and Borges) are to blame. First, β-ing is for the agents the actional means to α-ing, which entails the following two things: (i) β-ing is more basic an action than α-ing is (in the sense of “basic” in which we say, noncomparatively, that a given action is a “basic action”),Footnote ⁷ which means that for the agent to do α, she must do β as part of it, or that she can only do α by doing β; (ii) β-ing is practically sufficient for α-ing, which is to say that the agent does α just by doing β. This does not mean that β-ing entails α-ing—it does not; it is indeed practical rather than logical sufficiency that is at stake.

Second, the α label stands for the best description of Alvarez's behavior for which she is to blame; that is, the description picks out every feature of the agent's doing for which the agent is to blame, and these features, in turn, are the same as the features in virtue of which the doing is morally wrong. Particularly when wrongness is to do with doing harm, wrongness turns not only on the fact that harm is caused but also on the way the harm is produced. So the α label is meant to include not just the fact that a harm of a particular kind has been produced some way or another but also every aspect of the way in which the harm is brought about that affects the overall judgment of wrongness.

Suppose I intentionally kill Martha, my colleague next door, by throwing a grenade into her office. I do wrong, no doubt, in virtue of the fact that I kill her; and assuming I have no excuse, I am to blame for killing her. Killing Martha is certainly the dominant wrong-making factor in this story, but it is not the only one. There are also other features of my behavior that are intuitively relevant for an exhaustive judgment of wrongness. I think most people would agree on the following: that it is an intentional killing (rather than, say, an accidental killing) by a close colleague while the victim is working at the university (rather than, say, by a war enemy on the battlefield) and that it brings about death in a physically destructive way (rather than in a focused, painless, clean, and quick way). If I am right, then, to blame me just for killing Martha would understate the ground of my blameworthiness. A more accurate description of that for which I am to blame would be something like “intentionally killing Martha, the colleague next door, by throwing a grenade into her office.” Thus, if “I intentionally kill my colleague next door by throwing a grenade into her office” picks out every wrong-making aspect of my action, then “I α” would stand for it.

The features of the action that the β description picks out are in turn determined by the α description and the notion of actional means. Thus, in the example of my killing Martha, where “I α” equals “I intentionally kill my colleague next door by throwing a grenade into her office,” “I β” would stand for something like “I throw a grenade into my colleague's office with the intention of killing her.” For I do α just by doing that.

I mean the proposition “Alvarez does α” to entail the proposition “Alvarez does β”—though not the converse—so that when we place blame on Alvarez for her doing α, we are expressly reacting to the features of the case that make the β description true (as well as the α description). This point rules out the apparent oddity in the orthodox claim that Alvarez is no more to blame than Borges is, when Alvarez would be to blame both for α-ing and for β-ing, whereas Borges is to blame only for β-ing. My point ensures that although we can describe Alvarez's doing as an instance of α and an instance of β, we cannot place blame on Alvarez both for α-ing and for β-ing. Alvarez is to blame for α-ing. That is the relevant description under which she is to blame. For any other description would either understate the grounds of her blameworthiness or include irrelevant features. We can describe her behavior more narrowly by saying that she does β. Under that description, Alvarez would still be blameworthy (and no less blameworthy, according to the orthodox view), just as Borges is.

But this is not to say that Alvarez may be blamed for the conjunction of the two wrongdoings. This is a rather general point that applies to any action that is wrongful under more than one description when a description is entailed by another. If I intentionally kill Martha by throwing a grenade into her office, I also kill Martha, period. My behavior makes me blameworthy both under the description “intentionally killing Martha by throwing a grenade into her office” and under the narrower description “killing Martha.” Placing blame on me both for the conjunction of my killing Martha and my intentionally killing Martha by throwing a grenade into her office would be to engage in a kind of double-counting.

Success-failure pairs are expected to highlight the role of luck in the agents' doings. Alvarez succeeds, but she could as well have ended up as the agent of the failure case. Likewise, Borges fails, but he could as well have starred in the success case. Neither of them does anything that the other does not do to guarantee one particular role rather than the other. Their actions turn out the way they do (α-ing or just β-ing) in virtue of sheer luck.Footnote ⁸ In other words, Alvarez's β-ing happens to turn into an α-ing, in the sense that it could as well have remained a mere β-ing, as it is the case in Borges's case. With success, the world is such that the β-ing becomes an α-ing. With failure, the world is minimally different, so that the β-ing fails to become an α-ing.

Alvarez cannot be more to blame than Borges is, the orthodox claims, because such differences in the agents' luck as to how their actions turn out cannot by themselves ground differences in blameworthiness. It is in the nature of blameworthiness that they cannot.

III. BLAMEWORTHINESS, WRONGNESS, AND OWNERSHIP

Blameworthiness turns on wrongness, so that all things being equal, the more wrong an action is, the more blameworthy its agent will be. Now, if in success-failure pairs, Alvarez's conduct is more wrong than Borges's conduct is, then, all things being equal, Alvarez must be more to blame for her α-ing than Borges is for his β-ing. Things, I argue below, are however not equal.

Blameworthiness is a function of responsible wrongdoing, not just wrongdoing—one can do wrong without being to blame for it, as when wrong is done under insanity or other excusing conditions. Responsibility for a given action, as well as wrongness, comes in degrees. So one can be more or less to blame for a given wrongdoing depending on the degree of one's responsibility for that wrongdoing. Likewise, and this is the key to my argument for the orthodox view, one may be equally to blame for wrongs of different values if the corresponding degrees of responsibility are such that they counterbalance the difference in wrongness. Indeed, my argument is that although Alvarez does more wrong by α-ing than Borges does by β-ing, Alvarez's responsibility for α-ing is proportionally lower than Borges's responsibility for β-ing.

Being responsible here, as in the Strawsonian account of responsibility, is being appropriately susceptible to any of the reactive moral attitudes (e.g., love, admiration, gratitude, contempt, resentment, indignation, etc.).Footnote ⁹ Blameworthiness is a species of this more general notion of moral responsibility; being blameworthy is thus being appropriately susceptible to a particular reactive attitude—namely, blame. In this general sense, responsibility is, as David Copp puts it, response-worthiness.Footnote ¹⁰ For an agent S to be responsible for an action φ of hers is to be deserving of a moral response (any one) in virtue of her φ-ing—that is to say, her φ-ing makes it fitting that S is reacted to with some of the characteristic reactive moral attitudes.

In this sense of response-worthiness, responsibility centrally depends on what we may call attributability or ownership: S is responsible for her φ-ing if and to the extent that the φ-ing is S's own in a strong sense.Footnote ¹¹ Attributability warrants only that some moral response is due. For a particular moral response to be warranted, further conditions need be met. Blameworthiness requires that that for which one is to blame makes the distinctively negative valence of the blaming reaction fitting. The natural view is that for S to deserve blame for her φ-ing, it must be that she ought not to φ, that it is morally wrong for S to φ. Thus attributability accounts for the fact that S deserves blame for her φ-ing; wrongness in turn accommodates that it is blame that she deserves.

I follow Robert Nozick in framing the interplay of ownership and wrongness in terms of the product of a variable representing a measure of the seriousness of the action's wrongness multiplied by a discounting coefficient representing the degree of responsibility of the agent for the wrongdoing (ranging from zero for no responsibility to one for full responsibility). Under this framework, an agent's degree of blameworthiness for an action φ of hers is a function of the product r _φ × W _φ, where W _φ stands for the measure of φ's wrongness and r _φ is the discounting coefficient representing the degree to which the φ-ing is the agent's own.Footnote ¹²

I contend that between the cases in success-failure pairs there is a difference in responsibility—one, moreover, that counterbalances the difference in the amount of wrong done by one and the other agent. To show this, I analyze attributability in terms of responsiveness to reasons. Reasons-responsiveness, as exhibited in S's φ-ing, has two dimensions, I argue: one negative (reactivity), which involves the capacity to not-φ in response to reasons to not-φ; and the other positive (reliability), which involves the capacity to φ in response to reasons to φ. In success-failure pairs, the reactivity variables are fixed at constant values. Reliability, in contrast, changes its value from one case to the other, and this variation accounts for the difference in attributability on which I base my argument for the orthodox view.

A simpler example may serve to illustrate the kind of difference in attributability that, I argue, obtains between cases in success-failure pairs. Suppose that on my first day of archery, with my first shot my arrow happens to hit the bull's-eye. And suppose that my archery instructor also shoots and hits the bull's-eye—as she normally does when shooting from that distance. There is an intuitive sense in which the instructor may claim more responsibility for her hitting the bull's-eye than I can claim for my own. The reason is that she was far more in control of her shot than I was of my own, and therefore her hitting the bull's-eye is due more to her than my hitting the bull's-eye is due to me. Her hitting the bull's-eye is more her own doing than mine is my own doing. For luck has played a larger role in my hitting the bull's-eye than in the instructor's, and the more an action owes to luck, the less it owes to the agent.

In the following three sections, I build my case for this part of my argument for the orthodox view. Then I complete the argument by showing that the difference in responsibility cancels out the difference in wrongness between Alvarez's α-ing and Borges's β-ing.

IV. ON OWNERSHIP

Attributability or ownership is a relation that links agents and actions under descriptions.Footnote ¹³ For one and the same act—understood as a particular event—there may be many things we may predicate of it as things that an agent does.Footnote ¹⁴ And responsibility varies depending on the description under which the action is evaluated. Suppose you phone your friend on his birthday. Thus described, your action is something nice for you to do, something that calls for a positive moral reaction—some sort of gratitude for your remembering him on his birthday. But suppose that it is also true of your action that your call is too late. Under the description “telephoning your friend late,” your action is a poor thing for you to do, calling therefore for some kind of negative response. Your friend may reasonably be resentful at your calling him so late, and yet he may still reasonably feel thankful for your calling on his birthday. The positive judgment that the friend expresses in his thankfulness is fixed to whatever it is that makes it true of your action that it is a “calling your friend on his birthday.” The negative judgment that he also expresses in his manifest resentment attaches to other aspects of your action, namely, those that make the description of it true as an inadequately late call.

As responsibility varies with action-description, so does ownership. The attribution of an action φ, where “φ” stands for an act description, to an agent S depends on its being the case that (i) there is an event e such that e is an action of S's; (ii) e is φ—that is, we may predicate with truth “φ” of e or, which amounts to the same thing, we may speak with truth of e as an instance of φ or a φ-ing; and (iii) it is not just an accident for S that e is φ, that is, S exerts some suitable degree of control over e's being φ. Conditions (i) and (ii) just ensure that it is the case that S did actually φ.Footnote ¹⁵ It is the condition (iii), the control condition, that calls for some elaboration.

The relevant notion of control here is that which Timothy O'Connor calls agent control.Footnote ¹⁶ Agent control involves the internal relation between agent and action, a relation such that the action is an “outflowing of” the agent,Footnote ¹⁷ it is something that she does and somehow bears her personal mark on it. The account of agent control that I favor is in terms of responsiveness to reasons: an agent S controls an action φ of hers if and only if S displays (some degree of) responsiveness to reasons in φ-ing.Footnote ¹⁸ On this account, very roughly, we are in control of something we do—where the “something” is our action being such that it satisfies a given description—when it is our response to the circumstances as we see them. What we thus do is our doing in the strong sense of attributability because and to the extent that it expresses what things we take to be reasons for action, how we weigh them, and how we translate them into action—it expresses, in other words, what may be called our practical identity.

Reasons-responsiveness is a modal concept—its meaning is not fixed by what the case is in the actual world. The attributability of S's φ-ing in the actual world as S's own depends on S's behavior in a suitable range of possible worlds or scenarios. In John Martin Fischer's and Mark Ravizza's rendition of the account, S's φ-ing is reasons-responsive only if, holding fixed the kind of action-producing mechanism that issues in φ-ing in the actual world, there exists some possible world in which there is sufficient reason to not-φ and the agent does not-φ for that reason.Footnote ¹⁹ Two remarks are in order.

First, Fischer and Ravizza restrict the set of possible worlds within which to check for reasons-responsiveness to those where the kind of “action-producing mechanism” at play in the agent's doing φ in the actual world is held fixed.Footnote ²⁰ The concept of an “action-producing mechanism” refers to the process that leads to the action, or “the way the action comes about.”Footnote ²¹ Suppose I decide to φ after careful deliberation and consequently φ intentionally. The mechanism that issues in my intentionally φ-ing is my normal practical reasoning, which involves some mental events—such as some sort of intention to φ, some beliefs as to how to φ, and so on—and some physical events associated with those mental events. Holding this action-producing mechanism fixed means considering only worlds in which if I φ, I do it as the output of the same kind of practical reasoning—rather than, say, as a consequence of the manipulation of my brain by a Frankfurtian neuroscientist. The mechanism concept is admittedly vague but it is precise enough for my argument.

Second, the kind of reasons-responsiveness just introduced is too weak. There are cases of indisputably nonresponsible agents that exhibit this weak sensitivity to reasons. What responsibility requires, Fischer and Ravizza then add, is a more demanding kind of responsiveness, one that is revealed in the range of relevant possible worlds in which the agent recognizes there is a reason to not-φ, and does not-φ for that reason. To warrant responsibility, that range of possible worlds should evince an intelligible pattern of reasons that the agent is receptive to—a pattern that includes moral reasons.Footnote ²²

The relevant criterion for reasons-responsiveness should then be adjusted as follows: S's φ-ing is reasons-responsive if and only if, holding fixed the operation of the kind of action-producing mechanism issuing in φ-ing in the actual world, (i) there exists a range of possible worlds in which there is sufficient reason to not-φ and the agent does not-φ for that reason; and (ii) that range of possible worlds evinces an intelligible pattern of reasons (some of which are moral reasons) that the agent is receptive to.

Fischer's and Ravizza's criterion of reasons-responsiveness captures only one aspect or dimension of ownership. This dimension of ownership is, in a sense, negative, as it is represented by the agent doing other than φ-ing in the relevant set of possible worlds. I call it, following Fischer and Ravizza, reactivity.Footnote ²³ There is also, I contend, a positive dimension of reasons-responsiveness that Fischer and Ravizza fail to explore and which is crucial to my argument for the orthodox view. This dimension reflects the agent's ability to φ out of reasons for φ-ing.

I use the term reliability to refer to this positive dimension of agent control.Footnote ²⁴ Take the following example from Alfred Mele's and Paul Moser's analysis of intentional action:

I take the case of Fred and the nuclear reactor to be a good example of a doing that is a poor expression of Fred's agency, that is, of a very low degree of attributability.Footnote ²⁶ I contend that we are inclined to believe that the shutting down of the reactor is not an intentional action of Fred's—as the proponents of the example suggest—because we are inclined to believe that under that description, it is not an action of Fred's—that is, that it is too coincidental to be his own doing in the strong sense of attributability.

More generally, I propose, whatever the minimum of reliability might be for a doing to be an expression of agency, differences along the dimension of reliability determine differences in attributability. Suppose a companion case that is as close to Fred's case as it can be but for the fact that the agent, Fred*, does know what the ten-digit code is—because, say, he is the person in charge of the reactor and has access to the secret document containing the code. In the companion case, Fred*'s shutting down of the reactor is certainly his own doing. It is also meaningful to say, I contend, that Fred*'s shutting down of the reactor is more Fred*'s doing than Fred's shutting down of the reactor is Fred's own. The shutting down of the reactor in the companion case is, as it were, more due to Fred* than the shutting down of the reactor in the original example is due to Fred. This difference in ownership may be accounted for as a difference in the relative degrees of reliability of the corresponding action-producing mechanisms.

Reliability is a variable property that may be understood, like the reactivity dimension of reasons-responsiveness, in terms of the agent's performance in a set of possible worlds. Something like the following criterion will do:

This criterion captures the intuitive view that whereas Fred*'s shutting down the reactor is Fred*'s own doing in the strongest sense, Fred's shutting down the reactor is hardly an expression of his agency. What accounts for this difference in attributability is that the mechanism issuing in Fred's shutting down the reactor in the actual world (i.e., typing in the first ten digits that come to mind) is such that Fred shuts down the reactor in too narrow a range of possible worlds—in those in which the ten-digit code happens to be the same as the first ten numbers that come to Fred's mind, whereas Fred*'s mechanism (i.e., typing in the ten digits that he knows figure as the code in the pertinent secret document) is such that Fred* shuts down the reactor in most possible worlds within the relevant set.Footnote ²⁷

Since all other things are equal between these two cases, such difference in the reliability dimension of attributability will warrant differential responsibility judgments.

V. SOME PRELIMINARY OBJECTIONSFootnote ²⁸

I propose to conceive of attributability as depending on agent control and this to be reducible to the properties I called reactivity and reliability. Reliability, in turn, is a variable property—it comes in degrees. And in my analysis, so does attributability. Recall the archery example above. In his first day of archery, the novice archer surprisingly hits the bull's-eye with his first shot. Aiming from the same spot, the skillful instructor also hits the bull's-eye, as she normally does. We say of the novice archer's shot that it is an instance of “beginner's luck,” implying that his hitting the bull's-eye is not as attributable to him as his own doing as the instructor's hitting the bull's-eye is attributable to her as her own. My criterion of reliability purports to articulate differences of this sort.

One may point out, however, that it is an odd implication of my analysis that actions may be partially attributable to an agent—that, for instance, my doing this or that might be, say, half mine. Attributability, this concern would suggest, appears to be a binary relation—actions either are attributable to an agent or are not actions of that agent. There seems to be nothing in between.

I concede that there is a meaningful binary use of the concept of attributability under which it is meaningless to say of an action that it is, say, 60 percent mine, or less mine than that other action is yours. However, since in my account the factor upon which attributability depends (i.e., agent control) comes in degrees, it is also meaningful to conceive of attributability as a scalar or quantifiable relation reflecting the variable degree of agent control—and to use “attributability” to refer to that relation. Under this use, nonattributability and full attributability are only the extremes of a continuous scale.

Though I do use the language of attributability in its scalar form throughout the paper, I do not claim any privilege for that use. My only commitment is to the substantive claim that differences in the degree of control an agent bears over a given wrongful action of his—more specifically, differences in the reliability dimension of that control—affect the agent's blameworthiness for that action by affecting the degree of responsibility for that action. Whether we should call this variable upon which the degree of responsibility depends “attributability” or something else is an issue I need not adjudicate here.

So the troubling question is whether the binary use of the language of attributability reflects a substantive view about responsibility according to which the degree of responsibility does not vary with the degree of agent control, or at least, not in the way I claim it varies. Consider the killing of an innocent victim committed with bow and arrow by my archery instructor, on the one hand, and a similar killing performed by a lucky novice archer, on the other. If, as I claim, responsibility varies along with the degree of reliability, we should find the instructor more blameworthy for the murder she has committed than we find the novice archer for his. If I am right, equally wrongful but differentially reliable actions should warrant differential judgments of blameworthiness. And yet it seems that we would not intuitively draw any salient distinction in our blaming reactions to these two assassins; and this suggests that as far as our intuitive reactions go, differences in the reliability dimension of responsibility do not impact on our blameworthiness judgments.

The objection to my account that such a finding may ground—that is, the finding that we would not to draw any significant distinction in our blaming reactions for equally wrongful but differentially reliable actions—is not as strong as it might appear. The finding conflicts with other findings pointing in the opposite direction. Indeed, those intuitive blaming reactions contrast with our reactions when what is at stake are instances not of killing but of hitting the bull's-eye in an archery practice. In the archery scenario, the intuitive reaction is that the lucky novice archer is less deserving of praise for his hitting the bull's-eye than the skillful instructor is for hers; and this suggests that responsibility does vary with the degree of agent control, much in the way articulated in the criterion of reliability proposed above. It may be that we should revise one of these two reactions if we cannot properly accommodate both—it might be, for instance, that the reliability variable is relevant when the reaction at stake is of positive valence (say, praise), whereas it is irrelevant when the reaction is of negative valence (as in the case of blame). In any case, adjudicating the matter requires an account of the relationship between responsibility and agent control. The finding of a conflicting intuition is just not enough.

There are two plausible alternatives to my account that responsibility for an action φ varies as a direct function of the degree of reliability exhibited in the agent's φ-ing. The first alternative results from denying a fundamental idea that undergirds my account. In effect, my account captures the idea that responsibility and luck are essentially in tension with each other—that is, the idea motivating the debate on moral luck. The straightforward rejection of moral luck is, as I claim in the introduction, too extreme an expression of this tension. It is too extreme, I posit, for it leaves no room for our ever being responsible for anything. The account of responsibility that I propose for the domain of resultant luck does better than the utter rejection of moral luck, for it accommodates the same fundamental idea without falling into the cumbersome conclusion that we are never responsible for anything. In my account, luck is indeed in tension with responsibility but it does not altogether thwart responsibility; it modulates responsibility by diluting our agency—by undermining, that is to say, the attributability relation linking agents with actions.

The first alternative to my account results from dropping that fundamental idea that luck and responsibility are mutually in tension, or, which amounts to the same thing, that responsibility is essentially connected with control. The absurdity of the implications of denying that there is something like moral luck should, one might think, lead us straightforwardly to embrace moral luck—though perhaps with some marginal qualifications. If thus responsibility for actions had nothing to do with agent control, then it would appear that the degree of responsibility would not depend on the degree of reliability.

My impression is that a conception of responsibility for actions that is not fundamentally connected with control is just meaningless. As Thomas Nagel maintains in his seminal paper on moral luck, the condition of control is not the result of a generalization over a number of clear cases, nor can it be refuted by a number of wayward counterexamples. Elucidating the idea of responsibility is in part elucidating its troublesome and perhaps paradoxical relationship with luck.Footnote ²⁹ But even if that move were not meaningless in itself—if there were something left after removing the condition of control that could be meaningfully recognized as responsibility—it would nevertheless not be a move I should make in the context of this paper. In effect, my goal here is to produce an argument for the orthodox view—that is, the view that resultant luck is of no significance to the issue of deserved blame. And a conception of responsibility such that luck and responsibility are mutually in tension is part and parcel of that orthodox view. So it is no argument against my account that it adopts a conception of responsibility such that control is an essential factor—or, conversely, one that is essentially in tension with luck.

This takes us to the second alternative to my account. This second alternative would concede that responsibility depends on control but would deny that degree of responsibility is a function of agent control in the way advanced in my account. Particularly, this alternative would assert that responsibility is sensitive to control up to a threshold point, so that below that point the agent is not responsible—that is, there is no responsibility for insufficiently controlled actions—and above the threshold point, variations in the control dimension do not determine variations in degree of responsibility. Let us call this the threshold account.

I do not have an argument against the threshold account, as there seems to be no argument against my own account either. Both my account and the threshold account are sensible hypotheses about the way agent control impacts on responsibility. There is something, however, that may count in favor of my account, and that is its relative simplicity. Indeed, the threshold account needs to explain why the agent control variable, which is relevant to responsibility up to the threshold level, becomes irrelevant once the threshold is crossed. In the absence of such an explanation, it seems natural to assume that differences in the degree of agent control above the threshold level continue to be relevant, determining differences in degree of responsibility. My account is simpler than its alternative in the sense that it does not depend on whatever it is that might explain that behavior of the responsibility variable.

Now, even if I am right in maintaining that the degree of responsibility for an action is sensitive to the degree of agent control for that action, one may still want to raise the following objection. My account includes a measurement scale for the reliability dimension of agent control—the dimension, that is, in which the difference in responsibility between the cases in success-failure pairs manifests itself—according to which degree of reliability for an action φ is a directly proportional function of the frequency with which the agent does φ for a reason to φ within the set of worlds in which there is sufficient reason for that agent to φ. Combined with the idea that the degree of agent control determines the degree of responsibility, that measurement scale yields apparently counterintuitive conclusions, as illustrated in the following pair of cases.

Suppose that my archery instructor assassinates an innocent victim with her bow and arrow, and that her assassinating this victim exhibits an extremely high level of reliability—for she succeeds in killing the victim in almost every possible world within the pertinent set. Let us say that 0.99 expresses that ratio. And suppose, on the other hand, that the novice archer also assassinates an innocent victim in the same way but with an action exhibiting a reliability level of only 0.01—for he fails to kill the victim in almost every possible world within the relevant set. If everything else in these cases is equal, it would follow from my account that the instructor is ninety-nine times more blameworthy for the murder she committed than the novice archer is for his own act of murder. This seems counterintuitive. We may be willing to revise our intuitions in our blameworthiness judgments to adjust them to what we do in our praiseworthiness judgments—and so be prepared to have our blaming reactions be sensitive to the degree of agent control. But to be committed to finding the instructor ninety-nine times more to blame than we find the novice archer appears far too weird to be acceptable—or so the objection would go.

The objection, however, is only apparent. In my account, the level of responsibility of an agent for an action φ corresponds to a point on a continuous scale—the scale representing all the possible frequencies with which an agent can succeed in doing φ out of a reason to φ. We may thus say that the level of responsibility of the instructor for the murder she committed corresponds to a point close to the top of the scale, whereas the level of responsibility of the novice archer for his otherwise equivalent murder corresponds to a point near the bottom of the same scale. Now, to tell how many times more responsible than the novice archer the archery instructor is, we need to partition the continuous scale into a finite number of discrete intervals. In my rendition of the objection, the scale is partitioned into one hundred intervals, so that the level of responsibility of the novice archer falls within segment number 1, and the level of responsibility of the archery instructor falls within the segment number 99. Thus the instructor is ninety-nine times more responsible than the novice archer relative to that discrete distribution.

More generally, how many times more (or less) responsible one agent is with respect to another is relative to the granularity of the partition of the corresponding responsibility scale: the coarser the partition is, that is, the fewer the number of intervals into which we divide the continuous scale, the fewer the number of times more responsible than the novice archer the archery instructor will be—and conversely, the finer the partition is, the larger the number of times more responsible than the novice archer the instructor will be.Footnote ³⁰

Whether we should partition the responsibility scale—and in that case how finely we should partition it—depends on what we need the partition for. A salient reason to do it reveals itself when we are to express blameworthiness judgments through our behavior. We express judgments of blameworthiness by adjusting our behavior toward the blameworthy person—typically acting on her in harmful ways. But the repertoire of actions at our disposal through which we can express blame is finite and often pretty coarse—think of our standing regimes of imprisonment sanctions. So, for any given kind of wrongdoing, we will have a finite and perhaps rather coarse set of possible behavioral reactions through which to express an infinite set of possible differences in responsibility. Our blaming practices are thus inescapably approximative. The objection I am considering exploits this feature of our blaming practices. If we intuitively presuppose a coarse repertoire of blaming reactions, it would seem counterintuitive to be committed to make more distinctions than our repertoire of reactions allows.

But my account does not require any partitioning of the responsibility scale. Indeed, I offer a model that purports to help us understand how agents may be more or less blameworthy for their actions, a model, moreover, that makes room for measurements of differences in degree of responsibility that may compensate for differences in degree of wrongness. In my account, responsibility for actions depends on degree of attributability, which in turn depends on the variable I name reliability. Degree of reliability may be understood, I propose, as the frequency of success of an action of the pertinent kind within a suitably restricted set of possible worlds. We need not translate the scale that this criterion presupposes into a scale of discrete values in order to tell whether in success-failure pairs differences in responsibility determined by differences in degree of reliability compensate for differences in degree of wrongness. As is apparent below, the criterion I am proposing is just enough.

Furthermore, my account does not presuppose anything finer than the coarse repertoires of blaming reactions that are available to us—those under which we may draw no difference in our blaming responses for equally wrongful but differentially reliable actions. If it did, that would naturally count against my account. But it does not. My reducing the reliability dimension of responsibility to a frequency allows the comparison of variations in degree of responsibility with corresponding variations in degree of wrongness and therefore the comparison of degrees of blameworthiness in success-failure pairs. If my complete argument is right, it proves that agents in success-failure pairs are equally blameworthy. It follows from that conclusion that we have reason to blame agents in success-failure pairs the same way—and not to blame one more than we blame the other.

The point that I want to stress now is that this implication holds independently of the granularity of the available palette of possible blaming reactions. Indeed, for a palette of blaming reactions to be coarse is for it to clump together under equivalent reactions cases yielding different blameworthiness judgments—that is, cases that would deserve different reactions under an unconstrained, ideally perfect palette. But cases that yield equivalent blameworthiness judgments—that is, cases that would deserve the same reaction under the perfect palette—warrant equivalent blaming reactions whatever the length of the intervals into which the applicable repertoire of reactions divides the blameworthiness scale; hence my claim that agents in success-failure pairs, if equally blameworthy, are to be equally blamed even under a realistically coarse palette of blaming reactions.

VI. SUCCESS-FAILURE PAIRS: ASSESSING RELATIVE OWNERSHIP

Let us go back to success-failure pairs. Alvarez is blameworthy for α-ing, and her blameworthiness is determined by the amount of wrong done by her α-ing and the degree to which the α-ing may be attributed to Alvarez as her doing. The same goes for Borges, whose blameworthiness for his β-ing is a function of the amount of wrong done by β-ing and the attributability of this action as Borges's own doing. If, as I assume, α-ing is somehow more wrongful than β-ing is, then for the orthodox view to be true, there must be a proportional difference in the attributability variables such that it cancels out the difference in the amount of wrong done by one and the other agent.

In this section I argue that there is indeed a difference in the relevant attributability variables between the cases in success-failure pairs that is similar in kind to the difference in attributability between cases like Fred's shutting down the reactor and Fred*'s shutting down the reactor, and between the archery instructor's hitting the bull's-eye and the novice archer's hitting the bull's-eye. More specifically, I maintain that we may conceive of the degree of attributability of Alvarez's α-ing as her doing as a determinate fraction of the measure of the degree of attributability of Borges's β-ing as Borges's doing.

The degree of attributability of an action, I claim above, is determined by the degree of agent control exhibited by the agent in her action, which in turn has two aspects or dimensions—reactivity and reliability. There is no difference of attributability between the success case and the failure case as regards the dimension of reactivity. We measure reactivity, let us recall, by checking for possible worlds where the agent has sufficient reason to act differently from the way she does in the actual world, and where she does act differently for that reason. Reactivity does not vary from success to failure, for the set of possible worlds in which Alvarez has a reason to not-α and does not-α for that reason is identical to the set of possible worlds in which Borges has a reason to not-β and does not-β for that reason. This point is ensured by the following two features of success-failure pairs.

The first feature is the fact that Borges and Alvarez are only numerically distinct—they are in every other respect one and the same person and so they are identically prone to find motivation in the same sort of facts. More specifically, for any fact f and any act φ, if f is a reason for Alvarez to φ, it is also a reason for Borges to φ, and vice versa. Therefore, in every world in which Alvarez finds a reason to φ and does φ for that reason, Borges would identically find a reason to φ and would φ for that reason, and vice versa.

The second feature is the relation between actions α and β. Since β is the actional means to α (i.e., it is more basic than and practically sufficient to α), they are identical at the level of reasons. This identity at the level of reasons for α and β stands for the conjunction of the following four propositions: For any fact f and any agent S, (i) if f is a reason for S to α, it is also a reason for S to β; (ii) if f is a reason for S to β, it is also a reason for S to α; (iii) if f is a reason for S to not-α, it is also a reason for S to not-β; and (iv) if f is a reason for S to not-β, it is also a reason for S to not-α. Propositions (i) and (iv) follow from the fact that β is a constitutive part of α, which is part of the meaning of its being more basic than α; and (ii) and (iii) follow from the relation of practical sufficiency. Particularly, proposition (iii) follows from the fact that if S is to not-α, she must avoid β-ing. For by definition, S cannot do β without making it likely that the β-ing turns into an instance of α-ing.

Proposition (ii) is a little trickier. For β is a part of α, and it seems that we may have reason to do something that happens to be part of a more complex action for which we have no reason—or for which the former reason is no reason at all. For example, I went out for a morning walk today, and the path I took led me to the riverside. My need for some daily exercise and fresh air was the reason I set out for my morning walk. But I went down to the riverside—rather than up to the park—for no particular reason. Although my walking this morning was part of my going to the riverside, the reason for my walking was not also a reason to go to the riverside—I could have honored my reason for walking by heading to the park instead. But the relationship between α and β in success-failure pairs is not like the relationship between “taking a morning walk” and “going to the riverside.” Recall that the β label refers to a rather peculiar action description, a description we get at by subtracting only a nonactional feature from the action described by the α label, so that when we say of Borges that he does β, we imply that there is no description of what Alvarez does other than α that is not true as well of Borges.

If, say, “I α” stands for “I intentionally kill my colleague next door by throwing a grenade into her office with the intention of killing her,” my actional means to that action will be my throwing a grenade into my colleague's office with the intention of killing her—so in this example, “I β” stands for “I throw a grenade into may colleague's office with the intention of killing her.” The relationship between α and β is thus so tight that it ensures that the agent cannot rationally choose to β without thereby also choosing to α (at least with some probability). Reasons are facts that ground considerations in favor of our choosing what to do. Given that tight connection between α and β, if a fact gives reason to β, it must be that it gives reason (at least probably) to α—hence proposition (ii) above.

Those two features of success-failure pairs—Alvarez's and Borges's personal identity, and identity at the level of reasons between actions α and β—ensure, then, that the set of worlds in which Alvarez finds a reason to not-α and does not-α for that reason is identical to the set of worlds in which Borges finds a reason to not-β and does not-β for that reason; hence my point that there is no difference to draw between success and failure along the reactivity dimension of attributability.

Let us turn now to the reliability dimension of attributability. Unlike reactivity, reliability does mark a difference in attributability of Alvarez's α-ing as her doing, and Borges's β-ing as his doing. I analyze reliability of the agent's doing a given action, let us recall, as the frequency of success in doing that action across a range of possible worlds. The measure of reliability is thus to be expressed as a value ranging from one for full reliability to zero for total unreliability, depending on the frequency with which the pertinent action obtains within the relevant set of possible worlds.

Let r _β express the degree of attributability (as a function of reliability) of Borges's β-ing with failure. It will be at some point between full attributability (where r _β = 1) and nonattributability (where r _β = 0), depending on (i) the number of relevant possible worlds in which there is sufficient reason for Borges to β and Borges does β for that reason, and (ii) the number of relevant possible worlds in which there is sufficient reason for Borges to β. The value of r _β represents the ratio of the number of worlds in (i) to the number of worlds in (ii).

Whatever the actual value of r _β is, the value of the corresponding variable expressing the degree of attributability (as a function of reliability) of Alvarez's α-ing with success (or r _α) is equivalent to a fraction of r _β. The following three features of success-failure pairs support this claim.

First, r _α stands for the ratio of (a), the number of relevant possible worlds in which there is sufficient reason for Alvarez to α and Alvarez does α for that reason, to (b), the number of relevant possible worlds in which there is sufficient reason for Alvarez to α. Now, it follows from the fact that β is the actional means to α, that Alvarez does α only in those worlds in which she does β. Moreover, since it is part of the definition of success-failure pairs that there is some room for failure in Alvarez's α-ing—for it is within that space where the failure case lies—then it follows that the set of worlds in which Alvarez does α is a subset of the set of worlds in which she does β. There will be some possible world (or worlds), that is to say, in which Alvarez does β without doing α as well. So the measure of variable r _α will be lower than the corresponding measure of the frequency of Alvarez's β-ing within the reference set of possible worlds relevant to reliability for Alvarez's α-ing—that is, the set of relevant possible worlds in which there is sufficient reason for Alvarez to α.

Second, the identity at the level of reasons of α and β allows us to move from that observation to the claim that the measure of the attributability of Alvarez's β-ing as her doing is higher than the measure of the attributability of Alvarez's α-ing as her doing. Indeed, we assess the reliability variable for Alvarez's β-ing with success by estimating the frequency of Alvarez's β-ing within the set of relevant possible worlds in which there is sufficient reason for Alvarez to β. The identity of α and β at the level of reasons warrants that this reference set is identical to the reference set within which to measure the variable r _α (i.e., the set of possible worlds in which there is sufficient reason for Alvarez to α). Since the frequency of Alvarez's β-ing within that set is higher than the frequency of her α-ing within that set (for there are some worlds in which she does β without thereby doing α), it follows that the value of the variable expressing the attributability of Alvarez's β-ing with success is higher than r _α.

Third, cases in success-failure pairs are as close to each other as to ensure that the β-ing with success is as much Alvarez's doing as the β-ing with failure is Borges's doing. Indeed, we may conceive of success and failure as describing two members of the reference set of possible worlds within which we measure the frequency of Alvarez's α-ing (and β-ing) to establish attributability (“Alvarez” and “Borges” are the names with which we refer to one and the same agent, depending on the world she/he is in, a success world or a failure world). Hence the degree of attributability of the β-ing with success to Alvarez as her doing is equivalent to r _β, which is in turn, then, higher than r _α.

Success-failure pairs illustrate a general thought about attributability. The thought is that attributability decreases as the complexity of the action increases, all things being equal. More specifically, for any agent S and any nonbasic action φ of hers, there will be some more basic (or less complex) action that S does as a part of φ by which S does φ—for example, S turns the light on by flipping the switch; her flipping the switch is more basic than her turning the light on. The more basic the action is, the fewer the external conditions that need be met for it to succeed—for example, S succeeds in flipping the switch even if the bulb fails to light, but she needs the bulb actually to work for it to be the case that she turns the light on. Hence, if there is room for failure in the more complex action—for example, if it is possible that the bulb fails—the set of the possible worlds in which the agent does the more complex action will be smaller than the set of the corresponding possible worlds in which she achieves the more basic action. (In the extreme, attributability would reach its maximum level with the corresponding noncomparatively basic act.) The difference in the attributability degrees in success-failure pairs instantiates this general thought—that is, r _β is higher than r _α because β is more basic than α, and there is no difference between Alvarez's β-ing with success and Borges's β-ing with failure.

This argument shows that the level of attributability of Alvarez's α-ing is necessarily lower than the corresponding degree of attributability of Borges's β-ing. Let me now state more clearly how to measure such difference.

Let r _α be the coefficient representing the degree of Alvarez's responsibility for her α-ing, and r _β be the corresponding coefficient representing the degree of Borges's responsibility for his β-ing, and reduce both coefficients (r _α and r _β) to the corresponding reliability variables.

According to what I am arguing, r _β may be understood as the measure of the frequency with which the agent does β within the relevant reference set of possible worlds in which there is sufficient reason to β. Coefficient r _α, in turn, represents the measure of the frequency with which the agent does α within the same reference set—that is, due to the identity at the level of reasons between α and β, the set of worlds in which there is sufficient reason to α is identical to the set of worlds in which there is sufficient reason to β. Since the agent can only do α by doing β as a part of it, the set of worlds in which the agent does α is in turn a subset of the set of worlds in which she does β.

Let variable r _α|β represent the ratio of the number of α-ings to the number of β-ings within the reference set of possible worlds, or the fraction of β-ings that are also α-ings within that set. Variable r _α|β thus represents the degree to which the attributability of Alvarez's α-ing is lower than the corresponding attributability of Borges's β-ing. Indeed, if r _β is the frequency of the β-ings obtaining within a given set of possible worlds, r _α is the frequency of the α-ings obtaining within the same set, and the frequency of α-ings depends on the frequency of β-ings (for an instance of α can obtain only if an instance of β has obtained, but not the converse), then, applying elementary probability calculus, we may calculate r _α as the product of r _β multiplied by the relative frequency of α-ings to β-ings, or r _α|β. That is, r _α = r _β × r _α|β. Variable r _α|β is therefore a sort of discounting coefficient that multiplied by r _β yields the value of r _α. In that sense I claim that r _α|β represents the degree to which the attributability of Alvarez's α-ing with success is lower than the corresponding attributability of Borges's β-ing with failure.

To complete my argument for the orthodox view, then, I offer in the next section a conception of wrongness such that the value of the wrong done by Borges's β-ing is lower than the wrong done by Alvarez's α-ing in a degree that matches the difference in attributability as represented by variable r _α|β.

VII. DISCOUNTED WRONGNESS

I contend that the most natural account of the wrongness of the actional means to a wrongful action—that is the wrongness of an action that stands to a wrongful action as β stands to α, involves the following two claims. First, the wrongness of the actional means is of a derivative nature: the actional means is wrongful, that is, because the less basic action to which it is a means is wrongful. Second, the path through which this derivative wrongness is derived is some sort of likelihood or probability of the actional means becoming the less basic nonderivatively wrongful action. For the actional means β and the wrongful action α to which β is the actional means, this natural view may be put in the following terms.

If W _α represents the amount of wrong done by α-ing, then the wrongness of β-ing, or W _β, is a function of W _α and the conditional probability of the agent achieving α once she has done β. Let p represent the measure of this conditional probability (ranging between 0, where β-ing creates no probability of α-ing, to 1, where β-ing necessarily ends in α-ing), and therefore W _β = p × W _α. Under this view, then, the wrongness of an actional means β to a wrongful action α varies with the probability of the agent doing α once she has done β. The more likely it is that the β-ing turns into an α-ing, the more wrongful it is for the agent to do β. In the extremes, where doing β has no chance whatsoever to bring about the wrongful action α, it is not derivatively wrongful for the agent to β; and where doing β necessarily brings about α, it is no less wrongful for the agent to β than it is for her to α.

In this account of derivative wrongness, probability is neither objective physical probability nor purely subjective probability understood as degree of confidence. Indeed, we may speak in terms of probability, as this term figures in the account, even if the world is deterministic. We can say, for instance, that I create a probability p (where p is at some point higher than 0 and lower than 1) of killing you by playing Russian roulette on you or that shooting at you after checking that the gun is fully loaded creates a higher probability of killing you than just playing Russian roulette on you, whatever the actual outcome happens to be and even if it is fixed by the physical history of the world and its natural laws. To claim in this sense that a particular event makes it probable that another particular event ensues is to express a kind of uncertainty about how the world is or how it functions.

This is not to say, on the other hand, that the language of probability, as it is used in this account of wrongness, is a purely subjective matter. By playing Russian roulette on you, I act derivatively wrongly because by playing Russian roulette on you, I may very likely kill you, and killing you is nonderivatively wrong. Now, by playing Russian roulette on you, I wrong you in this sense even if, being unreasonably confident in my good luck, I believe that you will not die as a consequence of the shot. By the same token, I do no wrong in trying to kill you by sticking pins in a voodoo doll because there is no chance whatsoever that I can kill you by thus sticking pins in such a doll, even though I unreasonably believe that by sticking pins in this doll, I make it likely that you will die.

Particularly, we may understand the probability variable p as a measure of the frequency with which the nonderivatively wrongful action α obtains within a reference set of possible worlds in which the agent does β as she does in the actual world. The reference set should be so restricted as to reflect the relevant kind of uncertainty that judgments of probability are meant to express. Suppose that S punches you in the nose and your nose starts bleeding, and we want to estimate the probability of S's making your nose bleed by punching you in the nose. If we consider what the frequency is of S's making your nose bleed by punching you in the nose within the set of worlds that are identical to the actual world in every causally relevant respect up to the time when S punches you in the nose, then we will come up with a frequency that matches the objective, physical probability or, if the actual world is deterministic, with the finding that S makes your nose bleed by punching you in the nose in every possible world within the set. So in order to capture the kind of probability in force in discounted wrongness, we should adjust the restriction of possible worlds, moving from such causally relevant identity with the actual world to something like perceived identity from a given agential point of view.

Let us call this latter set of possible worlds the set of possible worlds practically equivalent to the actual world. Practical equivalence is relative to a given kind of action and an agential standpoint. Thus possible world w is practically equivalent to the actual world as regards the action “making your nose bleed by punching you in the nose” and from S's point of view if and only if S finds in w as propitious an opportunity to make your nose bleed by punching you in the nose as she finds in the actual world. Put another way, w is practically equivalent to the actual world with respect to that kind of action and from S's point of view because if S wanted to make your nose bleed by punching you in the nose, she would be totally indifferent as to whether she is in w or in the actual world.

Practically equivalent worlds may be very different from one another. They may differ both in features epistemically unavailable to the reference agent—such as low-level physical properties or the contents of secret documents—and in any accessible feature bearing no influence on the prospects of the action from the point of view of the reference agent.

By restricting the reference set of possible worlds to the worlds that are practically equivalent to the actual world, we make the probability judgment reflect the epistemic limitations of the reference agent (i.e., the agent whose epistemic abilities we take into account to determine practical equivalence), so that the more epistemically able the reference agent is, the closer to the physical, objective probability (if there is such a thing) the resulting probability measure will be.

It is unclear to me who the relevant reference agent should be: the agent whose action is under evaluation or someone else—perhaps something like an ideal, “reasonable” agent—though I am inclined in favor of the former. The choice of the actual agent's stance as the relevant agential point of view is in keeping with the theory that moral reasons in general and wrong-making factors in particular must fall within the cognitive capacity of the agent whose conduct they purport to guide. In general, judgments of rightness or wrongness, Jonathan Dancy has argue, are relative to a certain body of facts; and judgments of moral rightness or wrongness in particular are relative to the body of facts available to the agent. The reason is that if something is to count to us as a moral reason, it must be “capable of being practically relevant for us.”Footnote ³¹ By including the agent's point of view in the criterion of practical equivalence of possible worlds, we ensure that the resulting wrong-making factor—that is, the probability of the action turning out to make your nose bleed—is accessible to the agent and therefore may play a role in guiding her conduct.

VIII. CLOSING THE ARGUMENT

I wish now to complete my case for the orthodox view by arguing that in success-failure pairs, the degree to which the attributability of Alvarez's α-ing is lower than the corresponding attributability measure of Borges's β-ing is equivalent to the degree to which β-ing is less wrongful than α-ing is—or why, formally, p = r _α|β.

Both r _α|β and p may be understood as expressing the ratio of the number of α-ings to the number of β-ings within a suitably restricted set of possible worlds. Variable p expresses the ratio of the number of times the agent does α to the number of possible worlds in the set of possible worlds that are practically equivalent to the actual world with respect to α and from the agent's point of view (the actual world being that in which failure obtains) and in which the agent does β. Let us call this reference set P. The r _α|β variable, on the other hand, stands for the ratio of the number of times the same agent does α for a reason to α to the number of possible worlds in the set of possible worlds in which there is sufficient reason for the agent to α and the agent does β for that reason. Let us call this latter set R.

These two sets, P and R, are identical to each other—and hence r _α|β and p state identical ratios—if the features of the agent's doing that the α label picks out are such that (i) the agent can only do α for a reason; and (ii) doing α entails that the agent (in doing β) entertains causal beliefs on the prospects of her achieving α of some given kind—so that if acted with relevantly different beliefs, it would not be α (nor β) that she does, but something else. Indeed, condition (i) ensures that in every world in which the agent does α (or just its actional means, β), she does so in response to a reason to α and so that every world in P belongs as well in R. Condition (ii), in turn, warrants the converse, that is, that every world in R also belongs in P, for every possible world in which the agent does α (or just β) for a reason is a world that is practically equivalent to the actual world with respect to α and from the agent's point of view.

Conditions (i) and (ii) are met, I contend, in every case of intentional wrongdoing that may play the role of a success case. This turns on two assumptions. The first assumption is that when an agent does something intentionally, she does it for a reason.Footnote ³² So, if “α” describes an intentional action (i.e., an action that is intentional under the α description), there is no world in which the agent does α (or its actional means β) without doing it for a reason. Hence condition (i) is met.

The second assumption is that if “α” describes an intentional action, it implies that the agent acts (in doing β) with certain beliefs as to the prospects of her achieving α.Footnote ³³ So every possible world in which the agent does α (or just β) is a world in which she entertains (in doing β) those same beliefs as to the prospects of her achieving α and is also a world in which the grounds of those beliefs are relevantly the same. By grounds I mean the features of the world on the basis of which the agent believes that she will (or may with a given degree of likelihood) achieve α once she has done β. This ensures that every feature of the actual world that from the agent's point of view is causally relevant to her achieving α must obtain in every possible world in which the agent does α (or just β). In other words, every possible world in which the agent does α (or just β) is practically equivalent to the actual world (in which the agent has done α) with respect to α and from the agent's point of view. Hence condition (ii) is satisfied.

If my contention is right, variables r _α|β and p in success-failure pairs necessarily represent equivalent ratios, at least when that for which the agents are to blame—that is, the actions that “α” and “β” describe—are intentional actions. In the domain of those kinds of actions, then, the orthodox view that successful offenders are no more to blame than their failing counterparts is true on the basis of the argument I am deploying.

Let me finally summarize my argument. I start with the assumptions that in success-failure pairs, the successful agent does more wrong than the unsuccessful counterpart does and that the magnitude of wrong done impacts on the agent's degree of blameworthiness. I argue next that blameworthiness is a function not only of the magnitude of the wrong done but also of the degree to which the wrongdoing is attributable to the agent as her own doing. Thus, applying the Nozickian framework r _φ × W _φ,Footnote ³⁴ we can understand the orthodox view in terms of the following equation:

(1)

The left-hand side of this equation expresses the magnitude of Alvarez's blameworthiness for his α-ing with success determined by the value of the wrong done by α-ing (W _α) multiplied by the coefficient representing the corresponding degree of attributability (r _α), or r _α × W _α. On the right-hand side, we find the expression of Borges's degree of blameworthiness for his β-ing with failure, as it is correspondingly determined by r _β × W _β. The orthodox view is true, I claim, because the difference between attributability coefficients (r _α and r _β) cancels out the difference in wrong done by one and the other agent.

I interpret the attributability coefficients (r _α and r _β) in equation (1) to reflect only the measures of reliability of each agent's doing—the other dimension of attributability, that is, reactivity, being constant from one case to the other. Thus they are understood to express a measure between 0 and 1 of the frequency with which the pertinent action obtains within a given set of possible worlds. Coefficient r _α in equation (1), I argue, is a fraction of coefficient r _β and may be reduced, I propose, to the product of r _β × r _α|β, where r _α|β represents the ratio of the number of actions of the type α to the number of actions of the type β within the relevant set of possible worlds. Together with the claim that r _α = r _β × r _α|β, equation (1) yields:

(2)

which in turn yields the simpler equation:

(3)

I maintain that the wrongness of Borges's β-ing should be conceived as a derivation of the wrongness of α-ing and its measure as a fraction of the value of the wrong done by α-ing. More specifically, I argue that the value of wrong done by β-ing equals the value of wrong done by α-ing discounted by the probability of the agent achieving α once he has done β. If we let p stand for a discounting coefficient representing the measure of that probability, then we may state the wrongness of Borges's β-ing as W _β = p × W _α. Replacing W _β in equation (3) with its equivalent expression leads to:

(4)

which in turn yields:

(5)

As (5) makes explicit, my argument ultimately depends on there being an argument that the frequency represented by coefficient r _α|β is equivalent to the measure p of the probability of the agent achieving α once she has done β. I argue that such is the case at least when the α label stands for an intentional wrongdoing.

IX. SOME FURTHER OBJECTIONS

Before concluding, I would like to discuss another group of possible objections. I start by taking into account opposition stemming from disagreement with my account of wrongness. Consider first the view that denies that there is any wrong done by the actional means to a wrongful act: that the wrongness of an action α, wherever it comes from, is not conveyed to its actional means. Take a version of the success-failure pair in which Alvarez and Borges attempt to kill Victor by shooting him; Alvarez succeeds, and Borges fails. If, as I am assuming, blameworthiness entails wrongness, this view implies that agents like Borges, who attempt to do serious wrong but fail, are blameless. I take this conclusion to be highly implausible and therefore not a reason to abandon my account in favor of this alternative view.

One may, however, escape the implausible conclusion by revising the assumption on the relation between blameworthiness and wrongdoing, thereby making room for judgments of blameworthiness in the absence of wrongdoing. Though possible,Footnote ³⁵ this move is not without costs for it breaks the identity between that which we would say an agent is to blame for and that in virtue of which he is to blame.

To illustrate, take again the version of the success-failure pair in which Alvarez and Borges attempt to kill Victor, Alvarez succeeds, and Borges fails. The cases describe an episode in the agents' lives—their respective attempts to kill Victor—that leads us to the judgment that both agents are to blame. In judging an agent blameworthy, we are expected to state what it is that the agent is to blame for. In my cases, the natural answer, I take it, will be that Alvarez is to blame for killing Victor intentionally, and that Borges is to blame for attempting to kill Victor. The account that best accommodates what we do in filling in these “for” clauses is that we are stating the ground or basis of the judgment of blameworthiness we have just passed; that is to say, we indicate the features of the case in virtue of which the agent is blameworthy. Under this account, what we offer as the basis of the judgment in the “for” clause should thus explain why the agent deserves blame.

We honor this expectation if, in “S is to blame for φ-ing,” we imply that it is morally wrong for S to φ. For the fact that S's φ-ing is morally wrong would indeed account for the fact that it is blame that S deserves. (The fact that the φ-ing is S's own doing would account for the fact that she deserves blame.) Were S's φ-ing not wrong—if it were, say, an action of the highest moral standard—it would be misleading, if intelligible at all, to say that S is to blame for φ-ing.

My point, in sum, is that it makes sense to give our natural answer “Borges is to blame for attempting to kill Victor” precisely because it is wrong for Borges to attempt to kill Victor.Footnote ³⁶ By denying that actions such as Borges's attempt to kill Victor can be wrong, we render the natural answer “Borges is to blame for attempting to kill Victor” meaningless.

I conclude that if we hold to the intuitively plausible view that agents such as Borges are blameworthy in failure, the natural way to account for this is to argue for the wrongness of what these agents do—that is, attempting to do what would be wrong if done. The alternative, consisting in freeing blameworthiness from the wrongdoing requirement, fails to accommodate a feature of our moral discourse (the natural “for” clauses of our judgments of blameworthiness).

Another alternative to my account of the wrongness of actional means would be to concede that attempts to do what would be wrong if done are wrongful themselves, but to claim that what makes such attempts wrong is the agent's acting on an intention she ought not to act upon—rather than the probability that the nonderivatively wrongful act obtains. Alec Walen argues that there are what he calls “illicit intentions,” that is, intentions we should not form or act upon, so that when we act on an illicit intention, we do wrong even though the act itself that we produce would be permissible under another (licit) intention.Footnote ³⁷ The wrongness of attempts, in Walen's view, may be accounted for in these terms. As he puts it, “[i]t is because an agent attempting to commit a crime is acting on an illicit intention that he is acting impermissibly.”Footnote ³⁸

An intention is illicit in this sense, according to Walen, if it makes the agent “disposed to perform actions that can independently be determined to be impermissible.”Footnote ³⁹ In a case of failure (under the killing-Victor interpretation), Borges's intention is illicit because it is such that Borges would carry out such intention by killing Victor (in the absence of justifying circumstances), and killing Victor (in the absence of justifying circumstances) is wrong independently of what the intention of the killer is. Borges's attempt to kill Victor is wrong, then, in virtue of his acting on an illicit intention and thus quite independently of what he does to Victor other than intending to kill him.

The illicit-intention approach to the wrongness of attempts fails to capture what appears to be an important dimension of attempts, namely, the imposition of risk. Surely it makes a great deal of a difference whether Borges attempts to kill Victor by playing Russian roulette on him or rather by sticking pins on a voodoo doll, even though in both cases Borges exhibits the same intention to kill Victor.Footnote ⁴⁰ At any rate, it does make a lot of a difference in the criminal law,Footnote ⁴¹ and I see no reason for this legal difference not to be a reflection of a moral difference. At the very least, it is a counterintuitive claim that, all things being equal, it is permissible for an agent to create for a nonconsenting victim a one-in-six chance of an instantaneous death, as in the Russian roulette example. The plausible view, as Judith Thomson convincingly argues, is that one ought not do that, quite independently of one's actual beliefs and intentions.Footnote ⁴² I do not mean that this plausible answer is without problems, as Thomson acutely remarks.Footnote ⁴³ But the same goes for the opposite view that denies the wrong-making nature of the imposition of risk.

My account of the wrongness of attempts like Borges's attempt to kill Victor is that it is wrong for Borges to attempt to kill Victor because he ought not to kill Victor, and attempting to kill Victor is Borges's actional means to killing him. Conceiving of the wrongness of actional means, and attempts in particular, along these lines is in keeping with the wrongness of risk imposition. Indeed, it is part of what makes of an action β the actional means to an action α that by β-ing, the agent makes α-ing likely. When α entails that some evil outcome obtains, such as harm to somebody else, we talk of that likelihood in terms of risk. Borges's attempting to kill Victor is wrong, then, because killing Victor is wrong, and by attempting to kill Victor, Borges makes it likely that he will kill Victor—that is, because by attempting to kill Victor, he imposes on Victor the risk of getting killed.

One may agree with me on the account of the wrongness of β and yet deny that α is any more wrong than β is. In the introduction I call this the equal-wrongness claim. I am assuming from the outset that the equal-wrongness claim is defective. Indeed, its being a defective view of wrongness is (in part) what has motivated my inquiry. For if the equal-wrongness claim were not defective, then it would provide a solid ground for the orthodox view. The objection I would like to consider now is not so much that there is this alternative view of wrongness as it is the possibility that my argument for the orthodox view be just a version of an argument based on the equal-wrongness claim.Footnote ⁴⁴

The worry may be grounded on the fact that my argument, after assuming that the discounting of wrongness applies only to the failing member of the success-failure pair, replicates the discounting maneuver in the attributability measure for the successful member of the pair in a way that looks as though it does in two steps what the equal-wrongness argument achieves in only one. The worry is thus whether my argument is nothing but a needlessly complicated way of putting the same idea that is expressed by the equal-wrongness argument—in which case I would have failed to produce the new argument I promised. I would have only introduced unnecessary complexity in one of the existing arguments—one, moreover, that I had assumed to be defective.

My answer to this worry puts the issue the other way around, for I think that it is not that my argument is unnecessarily complex; rather, it is the equal-wrongness argument that is too simple. In its distinctive simplicity, it misses something important—something, moreover, that my argument adequately accommodates.

The equal-wrongness argument suffers from the same kind of failure as views that deny that attempts or risk impositions might ever be wrong. For it renders meaningless our natural view of the grounds of our judgments of blameworthiness in cases like those in success-failure pairs. Since what is wrong under the equal-wrongness claim is, say, the imposition of risk rather than the actual bringing about of bad consequences, it leaves no room for stating meaningfully that Alvarez, who has intentionally killed Victor, is to blame for having killed Victor. Alvarez's wrongdoing, and therefore that for which she is to blame, would be something like imposing risk of death on Victor. This is a problem because a theory of wrongness is expected to accommodate our well-considered judgments of that in virtue of which agents are to blame. And it seems to me that one of those well-considered judgments is that in Alvarez's intentionally killing Victor, it is the intentional killing of Victor (and not merely the imposition of deadly risk) for which Alvarez is to blame.

Moreover, the case for the equal-wrongness claim seems less than conclusive. As mentioned in the introduction, the equal-wrongness claim is premised on the theory that wrong-making features need to lie within the cognitive reach of the agents whose conduct they purport to guide, together with the fact of the agents' characteristic epistemic bounds. The fact is that agents can always figure out what the expected consequences of their actions are, whereas the actual consequences of such actions are often inaccessible to them. Accordingly, risk (or expected harm) appears better suited than actual harm to serve as wrong-maker. However, if there are consequences of our actions that we can know of—as it seems hard to deny—why can they not figure among the set of wrong-making features?

The equal-wrongness argument, in sum, rescues the orthodox view at the expense of forcing us to modify our moral discourse in an unwarranted way. My argument, in contrast, reflects the natural view that Alvarez is to blame for killing Victor, that Borges is to blame just for imposing risk of death on Victor, and that the wrongness of the latter kind of wrongdoing somehow derives from the wrongness of the former.

One might still think that my move yields an argument for the orthodox view that is just apparently different from the equal-wrongness argument. After all, it accounts for the natural view of wrongness by removing the discounting variable from the measure of wrongness in success, but then it lets it in again, through the back door, as it were, when measuring the degree of responsibility for the succeeding action.

There are two aspects to this worry that I would like to distinguish. The first focuses on my reliance on the idea that responsibility for the more complex action is comparatively reduced with respect to the responsibility for its actional means. The second aspect is to do with the identity of the measurement scales with which I establish the diminished degree of wrongness of the actional means and the diminished degree of responsibility for the more complex action.

I take the idea that, all things being equal, responsibility decreases as the complexity of the action increases to be a rather safe finding. In this article, I have extensively defended the idea that attributability is at least part of what moral responsibility depends upon and that attributability turns on the degree of reliability of the action-producing mechanism through which the agent acts. Equally uncontroversial, I think, should be my contention that attributability decreases with the complexity of the action performed. And the reason is simple: the more complex the action is, the more conditions that need be met for it to obtain. And the more conditions necessary for an action to obtain, the more likely it is that the action fails to obtain, all things being equal. It follows that for any given action-producing mechanism, the more complex the action is for which it is an action-producing mechanism, the more likely it is that it fails, and therefore the lower the corresponding degree of reliability.

If this finding is right, then there is nothing to be suspicious about in my introducing a discounting variable in the measure of attributability for the more complex action that may track the discounting variable that applies to the measure of wrongness of the actional means. I just put together two substantive features of our intuitive moral judgments—to wit, the natural view of wrongness, and the notion of attributability as a decreasing function of action complexity. And putting together those two features is what makes room for the argument for the orthodox view I am pursuing.

What might remain suspect is my choice of the particular measurement scales for the wrongness of the actional means and for the relative attributability for the more complex action. I am willing to concede that the necessary identity of the values of those variables that my proposed measurement scales yield may appear cooked up to ensure the orthodox conclusion. For in a way that is correct. My goal in this paper is to expound an argument for the orthodox view—a principled argument, that is, that both warrants the view and accommodates those features of our moral thought that established arguments fail to accommodate. So, arriving at the point where the argument needs measurement scales to produce comparable values of discounted wrongness and diminished attributability, I eventually opt for two scales that can do the trick of securing precisely the orthodox conclusion. But this does not vitiate my argument if, as I expect to have shown, my choice is not ad hoc.

The key to my analysis of the degrees to which β-ing is less wrong than α-ing is and r_α is lower than r_β has two parts—one pretty strong and the other concededly weaker but certainly not ad hoc. The first part is my reduction of both measures to a frequency of α-ings resulting from β-ings in a suitably restricted set of possible worlds. The second is the restriction of the set of possible worlds such that when α is an intentional action, the relevant frequencies are equivalent. I have already spent enough space—even in this very section—defending the view that we should (and do) think of the wrongness of the actional means (β) to an independently wrongful action (α) in terms of the likelihood of the β-ing turning into α-ing, which may plausibly be read as a frequency.

As to my understanding of attributability (as a function of reliability) in terms of a frequency of the same sort, I can only add, first, that I cannot envision a better way of measuring it—I cannot see another way of explicating our sense that, say, my archery instructor is more responsible for her hitting the bull's-eye than I am for my hitting the bull's-eye with beginner's luck. And second, my analysis of the reliability dimension of responsibility falls within the framework of Fischer's and Ravizza's philosophically compelling account of moral responsibility as reasons-responsiveness—it is indeed an extension of such an account. That is why I claim that this part of my analysis not only is not ad hoc but rather is as strong as one can expect a philosophical explanation to be.

The restriction of the reference sets of possible worlds within which to measure the relevant frequencies of α-ings resulting from β-ings is the weakest link in my argument, as I concede, for the following reason. I am able to find a necessary coincidence in the relevant sets due to the controversial move of taking the agent's point of view in selecting the relevant set from which to measure the degree of wrongness of the actional means. Only thus may this set be shown to match, when “α” describes an intentional action, the set of worlds in which the agent finds a reason to α and does β for that reason—that is, the set within which, according to my analysis, we should measure the degree of reliability. But, as I am arguing, there are reasons to adopt the agent's stance that are independent of the securing of the orthodox conclusion—particularly the idea that for a feature to count as a wrong-maker (or more generally as a moral reason), it should be epistemically accessible to the agent whose conduct it purports to guide. Taking the agent's point of view in determining the degree of risk imposition tracks that general idea about moral reasons. That is enough, I believe, to warrant that my choice is not ad hoc.