II. A Puzzle

Fei, an undergraduate student, signs up for a psychological experiment. The researchers give her a version of the Stroop task, a classic task used in countless studies.^{Footnote 3} On each trial, a color word is presented on a screen (for example, “red,” “blue,” “green”) and the word itself is presented in an ink color. Fei is instructed that on all the trials, she should respond with the ink color of the word and not the word meaning (she responds by pressing one of five buttons, each associated with a color). On some trials, the color word presented on the screen matches its ink color, making it easier to get the correct response. On other trials, they don’t match (for example, the word “red” is shown in yellow colored ink), and it is more challenging to produce the correct response.

The task involves 1,000 trials over the course of about forty minutes. Her performance on the task is incentivized: for each trial with a correct response, twenty cents is donated to the local branch of the Humane Society, a charity that helps stray animals in the community. For each incorrect response, twenty cents is deducted from the amount to be donated.

Fei cares very much about animals—over the years, she has fostered several animals without homes—and so she tries very hard, and equally hard, on every single trial of the task. She produces the correct response on 996 trials and makes just four errors. By “error,” what I mean is that she intentionally presses one of the five buttons for her response but the button she presses is incorrect.^{Footnote 4} Overall, her response accuracy is very impressive: she is in the top 0.01% of people who have taken this task.^{Footnote 5}

Now, consider this question: On each trial, does Fei’s action arise from a reasons-responsive mechanism? The answer seems to be: surely yes. On each trial, Fei performs an intentional action: reaching out for and pressing one of the buttons. A number of psychological processes contribute to her performing this action. As we shall see later in more detail, these include processes related to attention, decision-making, and intention formation. The mechanism that issues in Fei’s action is naturally understood in terms of the joint operation of these processes; this mechanism helps her to both track what the correct response is and produce the correct response.

Of course, this mechanism does not operate flawlessly, since she does make some errors. But reasons-responsiveness theorists almost never insist that to qualify as reasons-responsive, the relevant mechanism that issues in action must be unerring in tracking one’s reasons. Rather, it is typical to set some threshold for reasons tracking, often a fairly lenient one. For example, theorists typically say the relevant mechanism must issue in alternative actions in “at least one” or “some” of the worlds in which there is sufficient reason to do otherwise.^{Footnote 6}

In Fei’s case, the reasons that she has to produce one of the five responses change constantly throughout the task as she is shown different color words printed in different ink colors. The mechanism that issues in Fei’s actions in the task is extraordinarily adept at tracking these reasons, which is why she gets the correct response on more than 99.5% of the trials, one of the best performances on record. Thus, it certainly seems plausible that in her case, the mechanism that issues in action in this task meets the threshold for being reasons-responsive.^{Footnote 7}

Now, if Fei acts on a reasons-responsive mechanism on each trial of the task, and if we assume a reasons-responsiveness view of moral responsibility is correct (and there are no defeaters pertaining to the other criteria for moral responsibility), then it follows that on each trial, Fei is morally responsible for what she does on that trial. That is, she is morally responsible for what she does on trial 1, morally responsible for what she does on trial 2, morally responsible for what she does on trial 3, and so on for all 1,000 trials.

But now, it seems, we have a problem. Given all that we know from thousands of studies of performance on the Stroop (and Stroop-like) tasks, just about everyone will make at least a few errors on this task, and this isn’t something that they can avoid. Fei, in particular, due to her deep love for animals, tried very hard to get the correct answer on all the trials, and she performed extraordinarily well: she is in the top 0.1% of those who have taken the task. But even she did not perform flawlessly.

It appears to be a basic feature of human psychology, just the way that our minds are set up, that errors of the kind that Fei made occasionally occur. We can try harder to focus on a task and thus drive down the rate of errors somewhat. But for us humans, such errors cannot be eliminated (and this is for principled reasons, as I discuss in what follows).^{Footnote 8} If the occurrence of rare errors on Stroop-type tasks is not something that we humans have the power to prevent, then, for a person who doesn’t want such errors to happen and sincerely tries to prevent them, it seems plainly wrong to say she is nonetheless morally responsible for them.

Yet, because Fei’s actions arise from a reasons-responsive mechanism throughout the task, the reasons-responsiveness view says that Fei is morally responsible for what she does in all the trials including her rare errors. The chain of reasoning that leads to this inconsistency is what I call the “fallibility paradox.”^{Footnote 9}

III. Relation to the Lottery Paradox

Some readers will have noticed that the fallibility paradox bears some similarities to the lottery paradox, first set out by Henry Kyburg,^{Footnote 10} one version of which goes as follows: Let us suppose that it is rational to accept a proposition if its probability of being true is very high, say at least 99.5% likely. Now consider a fair lottery with 1,000 tickets and a single winner. By our criteria, it is clearly rational to accept that the first ticket will not win. The same is true of the second ticket. And the third ticket. And so on for all 1,000 tickets. Given these observations, it seems we are entitled to conclude something stronger, based on the following inference:

This, of course, raises a problem. You already know that the fair lottery has a single winner, and so it is clearly rational to accept that one ticket will win. How can it be rational to accept both that a single ticket will win and that no ticket will win?

The core similarity between the lottery paradox and the fallibility paradox is that they challenge the use of seemingly arbitrary thresholds in philosophical accounts of rational acceptance and moral responsibility, respectively. They show that no matter how the thresholds are set, counterexamples emerge.

In the case of the lottery paradox, the threshold concerns rational acceptance. Some theorists might insist on a stringent threshold, say 0.995 probability, before one should rationally accept a proposition. Others might propose a more lenient standard. In the case of reasons-responsiveness views of moral responsibility, the threshold concerns the quantity of possible worlds in which the action-issuing mechanism does otherwise where there is sufficient reason to do so. As noted earlier, Fischer says the mechanism must do otherwise in at least one of these possible worlds, while other theorists have proposed more stringent thresholds.

Regardless of where the thresholds are set, for both accounts, once the thresholds are crossed, that is good enough: the proposition counts as rationally accepted and the mechanism counts as reasons-responsive, respectively.

Given the role of thresholds in these accounts, the two paradoxes produce their puzzling conclusions by applying the respective concepts (that is, rational acceptance and moral responsibility) in “iterated contexts.” These are contexts that involve a large number of highly similar cases. We antecedently know that the relevant concept does not apply to all the cases, yet in each individual case, the relevant concept’s associated threshold is exceeded. So the preceding accounts tell us to apply the relevant concept to all the cases. This results in the puzzling inconsistencies that lie at the heart of both paradoxes.

There are also differences between the fallibility paradox and the lottery paradox. Let me highlight one particularly important one. The chain of reasoning in the lottery paradox relies on a principle of agglomeration, one version of which says: for propositions p ₁, p ₂, . . . p _n, if it is rational to accept that p ₁, and if it is rational to accept that p ₂, and so on up to p _n, then it is rational to accept (p ₁ & p ₂, . . . p _n). This principle underlies the inference laid out in (A) above, which is an essential step in setting up the lottery paradox.

Agglomeration principles are problematic, and Kyberg’s strategy for addressing the lottery paradox involved disallowing the agglomerative step.^{Footnote 11} The fallibility paradox, however, does not rely on agglomeration at all. You can see this by looking closely at this step of the argument:

This premise is not agglomerative. An agglomerative version of this premise would say that if Fei is morally responsible for what she does on trail 1, morally responsible for what she does on trial 2, morally responsible for what she does on trial 3, and so on, then she is morally responsible for (what she does on trial 1 & what she does on trial 2 & what she does on trial 3, and so on . . . ). The thing in the parentheses is what we might call a “compound action”; it is the conjunction of multiple actions, and it is analogous to the conjunction of beliefs that plays a key role in the lottery paradox.

Importantly, the fallibility paradox does not rely on attributing moral responsibility for compound actions. Indeed, it is not even clear if such attributions are sensible. Leaving this question aside, the point I want to emphasize is that the fallibility paradox does not rely on agglomeration in any form. This observation significantly strengthens the fallibility paradox as a counterexample to reasons-responsiveness views; one cannot challenge the paradox by pointing to the problematic nature of agglomeration.

IV. The Different Mechanisms Response

Consider the following reply from a reasons responsiveness theorist: The mechanism that issues in action in Fei’s 996 correct trials is a reasons-responsive mechanism and she is morally responsible for what she does on those trials. But on the four trials in which she produces an incorrect response, a different type of mechanism operates. Furthermore, this mechanism is not reasons-responsive. In this way the reasons-responsiveness theorist can capture the intuition that she is not, after all, morally responsible for what she does on these four trials.

To make some progress in evaluating the “different mechanism” proposal, we need a detailed specification of what are the psychological mechanisms that operate in the Stroop task.^{Footnote 12} Here, we are in luck. The Stroop task is one of the most extensively studied in psychology and neuroscience. Based on this body of work, I want to now present a model of Stroop task performance with an eye toward illuminating whether the “different mechanism” response is viable.^{Footnote 13}

V. Other Responses by Reasons-Responsiveness Defenders

VI. What the Fallibility Paradox Tells Us about Moral Responsibility

One way to get a deeper handle on what drives the fallibility paradox is to consider a family of views of moral responsibility that don’t appear to be susceptible to the paradox.

So-called valuationist views^{Footnote 20} are an important family of views that contrast in major ways with reasons-responsiveness views. These views say that for an agent to be morally responsible for an action, the action must flow from and express her evaluative point of view. Valuationist views are typically formulated in terms of two parts, with different theorists filling in these parts in different ways. First, there is an account of which subset among an agent’s attitudes constitutes her evaluative point of view: values, cares, policies, and so on. Second, there is an account of what it means for an action to flow from or express an agent’s evaluative point of view. I will here focus on valuationist views that analyze the expression relation in terms of a type causal contribution: an action expresses the attitudes that constitute the agent’s evaluative point of view if these attitudes causally contribute (in the right way) to the production of the action.^{Footnote 21}

Valuationist views—in particular the causal contribution variety (I drop the qualifier going forward)—appear to make the right predictions about the case of Fei used to formulate the fallibility paradox. Recall that Fei cares very much about animals; animal welfare is something she deeply values. On the 996 trials in which she gets the correct response, it is natural to say that her responses flow from her values; it is in virtue of valuing animals that she directs top-down attention to strongly favor the correct response, which in turns strongly causally contributes to her producing the correct response. On the four trials in which she produces the incorrect response, her actions don’t flow from her values. On each of these trials, she—again motivated by her values—directs top-down attention to favor the color naming response. Nonetheless, the reading response wins out in the decisional competition due to the effects of various stochastic factors. The incorrect response that occurs thus flows from other features of her psyche; her values in fact “push against” the response.^{Footnote 22}

We can home in further on what exactly allows valuationist views to avoid the fallibility paradox while reasons-responsiveness theorists fall prey to it. The key difference is that valuationist views can make fine-grained distinctions where the reasons-responsiveness cannot.

Consider an agent S performing a series of actions that arise from a set of psychological processes P. Call the set of these actions A, and further suppose that P has some low rate of noise-based errors. As we have seen, reasons-responsiveness views employ a threshold to assess S’s moral responsibility for the elements of A. If P meets the threshold, then S is morally responsible for every action in A; if P fails to meet this threshold, then S is not morally responsible for any element of A. That is, so long as it is P that is operative in producing the actions, it is not possible for a reasons-responsiveness view to say S is morally responsible for some of the actions in A and not morally responsible for others.^{Footnote 23} This is what I refer to as the “course-grain” constraint on reasons-responsiveness views: reasons-responsiveness views must treat all elements of A identically.

It is worth emphasizing that, for the reasons that I noted earlier (see Section V.C), the problem of course-grainedness isn’t specific to mechanism-based versions of reasons-responsiveness views. There I noted that the most natural way to understand agent-based views is that they say the relevant P is not a subset of the agent’s psychological processes, but rather is all the psychological processes of the agent. Having selected an expansive P, agent-based views still face the exact same issue: they must next place a threshold for when P is sufficiently responsive for moral responsibility. Thus they too have the problem of being fundamentally course-grained: given their selection of P, they too must treat all elements of A identically.

Valuationist theories, in contrast, don’t rely on arbitrary thresholds on the responsivity of P, and this enables them to make the critical fine-grained distinctions needed to address the fallibility paradox. Even when P is operative across a large number of cases, valuationist views say the agent is morally responsible for the relevant actions in all and only those cases where P produces an action that expresses the agent’s evaluative point of view. In the rare instances in which P produces an action that does not express the agent’s evaluative point of view (for example, due to noise-based errors), the agent is not morally responsible for the action. Thus we are able to say the agent is morally responsible for most elements of A, but not all—precisely what is needed to avoid the fallibility paradox.

The fallibility paradox, then, isn’t just a counterexample to reasons-responsiveness views of moral responsibility. It is also the basis for a strong argument in favor of valuationist views. Reasons-responsiveness views are inherently course-grained, but the fallibility paradox highlights that our intuitions about moral responsibility don’t follow a course-grained pattern. Valuationist views, in contrast, are extremely well positioned, perhaps even uniquely positioned, to capture the fine-grained pattern with which we ordinarily attribute moral responsibility.