Natural and Artificial Moral Agency

The main focus of my attention in this work will be moral responsibility, in both its natural and artificial manifestations. However, I wish to draw upon and distinguish my account of artificial moral responsibility from the notion of AMAs. Thus, a fruitful point of departure will be an examination of moral agency, again, in both the natural and artificial manifestations. So, what exactly does it mean, naturally speaking, to be a moral agent?

As theorists since Aristotle have established, when we speak of agency we typically have in mind the ultimate source of an action.Footnote ⁸ Beyond being a random movement, actions can be intentional and performed on the basis of reasons. Intentional actions follow from decisions, or might be said to include decision-making as a sort of mental action.Footnote ⁹ Importantly, for those with the capacity to act intentionally, it is often thought, we see an ability to form a judgment based upon consideration of a multitude of reasons—for and against bringing about some state of affairs, or for bringing about some state of affairs rather than another.Footnote ¹⁰ In this way, agents have the ability to display desires, at least very loosely, along with beliefs and plans for how to fulfill them.Footnote ¹¹ In sum, to be the source of an action entails a number of crucial capacities, many of which appear to require something like consciousness.Footnote ¹² It will be helpful to note, then, what it is about these capacities that makes an agent a moral agent.

Crudely, if an agent is one who is capable of being the source of an action, along with the coinciding capacities for acting intentionally (and so on), moral agents can be thought of as those who maintain capacities for initiating action bearing moral significance. As Michael McKenna aptly states, a “moral agent is a person who is capable of action that can be morally evaluated as good or bad, right or wrong, virtuous or vicious.”Footnote ¹³ What this conception provides is a line with which we can draw several fundamental distinctions. First, we see the beginnings of a separation between moral agency and moral responsibility; I will discuss this further in the following section. At present, what we also begin to see is that moral agency typically describes one’s membership in a special subset of a wider class, known as the moral ‘patients’ or, simply, moral subjects. We commonly think, for example, that children or the environment are proper subjects of moral concern; while, normally, they do not (yet) have the capacity to be morally concerned. That is, the broader moral community is made up of a great diversity of natural specimens, only some of which maintain and display moral concern.

I suggested above that two key capacities are widely associated with moral agency, and often held as necessary conditions, namely a certain sort of knowledge and the ability to act freely. While agency simpliciter can be fruitfully seen as involving intentional action on the basis of reasons, when it comes to moral agency it seems that one must be able to act intentionally on moral reasons. Indeed, for those who can act intentionally but cannot grasp or feel the force of moral reasons, we would be hesitant to ascribe moral agency.Footnote ¹⁴ One must know right from wrong, as it were, along with a host of moral and non-moral considerations pertaining to the immediate situation.Footnote ¹⁵ Here we see why children and persons with cognitive or emotional impairments are not thought to possess full moral agency. They don’t know any better. Likewise, even where one does understand and is motivated by moral reasons, moral agency is not ascribed to those who cannot freely control their behavior. In cases of coercion—whether facing a gunman or a psychological compulsion—coerced subjects simply cannot act in ways they otherwise would or in ways they know to be right. Without a sufficient degree of control, again, we see that moral agency is compromised or entirely absent. The question, then, is who or what else, if anyone or anything other than fully functioning adult human beings, can truly possess moral agency?

On the natural picture of moral agency, as I have in mind here, the short answer is no one and nothing. Although the moral community is broad, encompassing all kinds of creatures properly demanding moral concern, the class of moral agents is a rather exclusive one.Footnote ¹⁶ Granted, it may be that the especially intelligent non-human animals—dolphins or chimpanzees, for instance—are capable of the sort of empathy we associate with moral concern.Footnote ¹⁷ Such creatures are also natural, biological beings, of course. But what appears non-natural, I am suggesting, is the ascription of full moral agency to anything other than creatures like ourselves. No doubt, we occasionally consider non-human animals as possessors of something like moral agency. But when we think, for example, of dolphins as non-human persons, or of the family dog as happy to see us, we are artificially ascribing a sense of moral agency. Such practices are not necessarily incorrect. The claim here is simply that we employ non-natural ascriptions of moral agency when the concept is used to denote creatures other than those who can grasp and act freely upon moral reasons.

Turning to the notion of AMAs, it should be stated upfront that Allen and Wallach are concerned with ascriptions of moral agency specifically to machines, from robots with human-like structures to automated software. Still, with the account offered here, we gain a broader understanding of what it means, generally, to ascribe moral agency artificially. Allen and Wallach state that the complexity of today’s systems will require “the systems themselves to make moral decisions…[and] will expand the circle of moral agents beyond humans to artificially intelligent systems.”Footnote ¹⁸ On this picture, AMAs are a fairly straightforward concept. They can be known simply as AI systems, whether robots or software, within the exclusive class of moral agents. Note that my initial framing of this class, namely as a subset of the wider circle of subjects deserving moral concern, might imply that some machines are to be considered capable of both exhibiting moral concern as well as being the target of our moral concern. It is here we see room for reasonable resistance to this expansion of moral agency.

Some thinkers encourage us to be open to the idea of being cared for and caring for AI. Consider the tales of Isaac Asimov, creator of the famed Laws of Robotics.Footnote ¹⁹ More recently, the work of David Gunkel persuades us to think of machines as possessing capacities for action and as deserving protections with rights of their own.Footnote ²⁰ For Gunkel, when considering the social and legal status of machines, moral agency should indeed remain a subset of the wider class of moral subjects. Yet, it is not clear why we must retain this standard conceptual map when navigating our emerging regulatory mechanisms for novel technologies. Undeniably, it would be difficult to justify showing absolutely no moral concern for a creature like ourselves, that is, for a natural moral agent. But the notion of AMAs, I take it, accounts precisely for the fact that non-natural ascriptions of moral agency to AI systems are just that—artificial—and nothing more. In other words, there appear to be few reasons to retain our conceptual map of natural moral agency and patiency when dealing with artificially intelligent beings. Admittedly, some of those reasons will find support. It might be argued, for example, that exhibiting disrespectful behavior toward AI systems, treating them as if they lack moral patiency, might globalize by negatively affecting our behavior toward fellow human beings.Footnote ²¹

My main objective, as stated, is to examine moral responsibility as it might be applied to AI. Thus, without firmly committing myself to what we owe (or do not owe) to machines, I want to suppose that AMAs represent an outlying specimen, a deviation from our standard conceptual map of moral agency and patiency. When it comes to artificial moral agents, moral agency is no longer a proper subset; instead, it is likely a largely overlapping Venn diagram, with some AMAs occupying the space beyond the class of moral subjects. Indeed, this picture appears to resemble what Allen and Wallach have in mind with their deployment of the term AMAs.

An AMA is not yet capable of consciousness, Allen and Wallach admit. Still, “if its architecture and mechanism allow it to do many of the same tasks” as human consciousness, it can be considered functionally conscious. Footnote ²² Analogously, even if AI is not capable of full moral agency, the architecture and mechanisms of many emerging systems allow for the same sorts of behaviors as natural moral agents. Accordingly, many AI systems can be considered functionally moral, occupying a space somewhere between “operational morality”—where decisions concerning values remain in the hands of the designers and users—and full moral agency.Footnote ²³ In short, while AI cannot be said to possess capacities for understanding moral reasons and for freely acting upon decisions in the ways we enjoy, today’s technologies are capable of behaving in ways that appear morally significant.Footnote ²⁴

Picture, for example, personal AI assistants programmed to encourage courteous interactions from their users.Footnote ²⁵ Clearly, this level of technological sophistication is not enough to warrant full moral agency, nor does it call for such devices being included within the sphere of our moral concern. The reasons to encourage courteous interactions are, of course, to support a certain developmental pattern in the users, particularly with children.Footnote ²⁶ Likewise, for a host of AI systems, the concerns for potentially negative interactions globalizing to fellow humans does not show that we should treat AI with moral concern for the sake of AI. It is because we wish to foster healthy moral development in humans and maintain moral concern for each other that we see reasons to treat some AI systems as if they were moral subjects and perhaps as something like moral agents.Footnote ²⁷

In what follows, I want to continue supposing that many AI systems already possess—and will increasingly exhibit—the sort of “functional morality” that Allen and Wallach outline. AMAs are surely not natural moral agents. Yet, they are capable of behaving in ways that are more significant, morally speaking, than machines that simply execute the commands of their designers and users. And although the notion of AMAs allows us to redraw our standard conceptual map of moral agency and patiency, we are nonetheless left without a firm grasp on how we might hold machines responsible. In order to reveal these processes, we must, to some degree, put aside questions of natural and artificial moral agency.

How Agency Does and Does Not Matter

I claimed above that moral agency is not coextensive with moral responsibility. In many cases, we might easily grant that one is a moral agent but that she is not morally responsible for some action performed, even one bearing great moral significance. Genuine moral dilemmas and cases of moral luck, for instance, begin to elicit intuitions supporting this division. But how do we draw the line in theory? What exactly is moral responsibility apart from moral agency? Addressing these inquiries will help to clarify the extent of agency’s significance in our attempt to locate responsibility in AI.

With his widely influential essay, “Freedom and Resentment,” P.F. Strawson moved the debate over the nature of moral responsibility beyond a strict focus on the conditions for moral agency.Footnote ²⁸ According to Strawson’s observations, there is something much more fundamental to what it means to be morally responsible than simply assuring that one is sufficiently informed and capable of acting freely. The moral community, he pointed out, is built upon relationships with one another. As members of the community and within our personal relationships, we hold expectations and we demand, at least implicitly, that we are treated by others with good will or, at minimum, without ill will. Accordingly, our concerns for the conduct of others are focused both on the actions themselves and on how they are performed. To illustrate, consider that we respond quite differently to the fellow subway passenger who steps on our toe to avoid tripping onto an elderly person, versus the fellow passenger who steps on our toe deliberately and with a mischievous smirk. Clearly, in the latter case, we are likely to respond with something like contempt or resentment. These sorts of responses—along with more positive sentiments, like admiration and gratitude—are commonly known as the reactive attitudes. For Strawson and numerous followers, being morally responsible is a function of our “natural human reactions to the good or ill will or indifference of others towards us.”Footnote ²⁹ We resent others for insulting us. We are grateful when others go out of their way to help us. In these ways, our attitudes as well as our corresponding practices of blaming and praising, punishing and rewarding, make up our responsibility responses. Footnote ³⁰

On its surface, Strawson’s development may appear rather commonplace, for indeed, the theory is constructed consistently with our everyday reflections and interactions. Yet, for many, its significance is hard to overstate. On an extreme reading, moral responsibility turns out to be a highly subjective property of its object, if it can even be called a property. On this view, moral responsibility is a process—we hold others responsible and, importantly, holding is conceptually prior to being responsible. Here we see what looks like a reversal of traditional concepts of responsibility. That is, the facts of responsibility are determined primarily by our responses, our attitudes and practices, and not the other way around.Footnote ³¹ In other words, moral agency does not really matter for our determinations of responsibility. If anything, considerations of agency are secondary, serving to affirm or call into question our existing attitudes and practices. For example, when I respond to the subway toe-stepper with anger, I might do so before knowing exactly who or what stepped on my toe. Upon learning that it was the smirking misfit, my anger is surely vindicated and I may well act accordingly, say, by communicating the offense. By contrast, were I to learn that it was purely an accident, that the toe-stepper was a child who knew no better, or that it was a lost bowling ball rolling down the aisle, I would likely work to quell my anger—or at least shift it from the immediate object to, say, the Fates. Thus, considerations of agency still matter, but are supplemental. We do not continue to hold ignorant children or bowling balls responsible, at least not appropriately. In any case, a key takeaway is that being morally responsible, on this view, is clearly something more than being a moral agent. It is to be held morally responsible by being targeted with a moral agent’s attitudes and practices. I’ll refer to this as the ‘process view’ of responsibility.

On the opposite extreme, we must—and we often do—account for the agential status of the objects to which we respond. On this line of thought, responsibility is an objective property, and our determinations of it will be based largely upon our awareness of fundamental features of the object. Is the source of harm human? Did he or she (or they) know what they were doing, and did they do so freely? Notice that the conditions of moral agency play a more prominent role here. Our attitudes and practices can certainly aid us when it comes to holding a given target responsible. For example, blame is often seen paradigmatically as a form of communication, and as we can imagine, communicating an offense may well be assisted by expressions of anger.Footnote ³² Still, on this picture, our responses are merely “epistemic markers” of responsibility, where being responsible remains conceptually prior to being held responsible.Footnote ³³ Again, on this view as well, we see that moral responsibility is something other than moral agency. It is to be identified as the appropriate target of our attitudes and practices. I’ll refer to this as the ‘property view’.

No doubt, the question of responsibility as a process or a property contains layers of complexity that cannot be fully addressed here. But again, I want to avoid coming down definitively in favor of either extreme. This is in part due to spatial limitations, but also because I can grant that both views have their merits. Specifically, the process view holds that moral responsibility depends upon our actual attitudes and practices, which seems highly plausible. As Strawson and many followers helped us to understand, without the interactions and responses that occur within our interpersonal relationships, it would be far from clear what we mean when we ascribe responsibility. At the same time, the property view illuminates the fact that we naturally modify and refine our responses according to observable features of the target.Footnote ³⁴ For these reasons, in what remains of this section, I will return to the issue of AMAs with the broadly Strawsonian framework, assessing where and why some theorists argue against locating moral responsibility in machines.

At the outset, I introduced Sparrow as a staunch advocate for the position that we lack responsibility when machines act autonomously. Considering that responsibility is implied by the jus in bello principles, our deployment of AI systems in warfare appears morally indefensible. For Sparrow, neither the programmers, the users, nor the machines themselves are plausible loci of responsibility. I take this line of argumentation to be generalizable to many public domains where AI is being developed or is already deployed.Footnote ³⁵ And while locating responsibility in programmers, users, or machines may indeed present difficulties, I want to clarify and challenge the argument, and thereby call into question any generalizations to other domains. With this task, my goal is not to promote AI in military and public contexts, but only to demote those who find it absolutely unjustified.

Sparrow states of autonomous systems that “it is not possible to hold anyone else responsible for its actions.”Footnote ³⁶ Note both the descriptive, absolute nature of this claim and that already he is loading the dice by singling-out machines themselves as the only potential loci of responsibility. Shortly thereafter, it is claimed “the more these machines are held to be autonomous the less it seems that those who program or design them, or those who order them into action, should be held responsible.”Footnote ³⁷ Note here two points. First, this is a normative claim; it tells us not that it is impossible to hold others responsible for a machine’s actions, only that doing so goes against some norm. Second, although it was not likely intended, with both claims cited here (among others) Sparrow displays support for a process view of responsibility. In other claims, however, the focus is less on holding and more on being responsible: “The question of who is responsible,” he says, for example, “remains crucial.”Footnote ³⁸

What is clear is that Sparrow does not consistently maintain a view on the nature of responsibility, which seems peculiar considering that the key to his overarching argument is to negate our chances of locating it. At times he deploys talk of responsibility as a process that can be subject to normative guidelines, and at other times as a property which simply can or cannot be found. Nevertheless, rather than settling on charges of inconsistency on a distinction that was likely unnoticed, I want to take his premises on the terms invoked. Doing so, I will show, serves to render the overarching argument unconvincing.

In his survey of possible sources of responsibility for autonomous weapons, Sparrow begins with the design and programming. Holding designers or programmers responsible, it is said, “will only be fair if the situation described [war crimes] occurred as a result of negligence.”Footnote ³⁹ Notice again the normativity at work, namely in the notion of fairness. The basic idea is not that we cannot hold designers and programmers responsible, but only that we should not. We have reasons not to, which are grounded in the broken chain of control or predictability in cases of fully autonomous systems. Even for the users, on Sparrow’s picture, the claim is not that we cannot locate responsibility, but that once machines choose their own targets “it will no longer be fair to hold the Commanding Officer responsible.”Footnote ⁴⁰ Aside from showing his earlier claims regarding the impossibility of holding others responsible to be misleading, a potential solution is being overlooked.

If responsibility is a procedural matter, a fair or unfair process of holding responsible, it may still be that we can and do exercise our usual attitudes and practices. We can blame, say, Microsoft for the racist remarks generated by its autonomous AI bot, Tay.Footnote ⁴¹ We can praise doctors for their successful use of AI-enabled diagnostic tools.Footnote ⁴² Often, our responsibility responses will seem inappropriate, but largely because we maintain a natural conceptual link, namely moral responsibility falling directly upon those who exercised moral agency in the action being evaluated. I will have more to say on this in the following section. At present, the point to be made is simply that the direct link between agency and responsibility, however natural it may be, is not necessary. We can and often do assign responsibility to individuals other than the agent who brought about the consequences in question. Importantly for weaponized AI systems and other high-stakes applications, responsibility can (and likely should) be taken by individuals associated with the effects.Footnote ⁴³ This is precisely what Marc Champagne and Ryan Tonkens aptly argue with their notion of “blank check” responsibility.Footnote ⁴⁴ Of course, blaming those who take responsibility does not appear as natural as blaming the agent identified as the causal source of an action. Still, for conceptual and practical purposes, when dealing with artificial moral agents, we must move beyond our reliance upon natural moral agency and responsibility.

When examining whether or not responsibility might be found in machines themselves, understandably, Sparrow adopts the more objective view, wherein responsibility is a property. Adopting this position here allows him to argue, on a purely conceptual basis, that autonomous weapons “could never be held morally responsible.”Footnote ⁴⁵ Yet, with the concept invoked—namely, punishment—we again see a plausible solution overlooked. For someone to be held morally responsible, Sparrow notes, they must be capable of receiving blame or praise, punishment or reward. Since AI systems cannot be punished or rewarded, they cannot be held responsible. But here it is explicitly assumed that the most plausible account of the nature and justification of punishment is retributivist or deontological in spirit. For punishment to be punishment, he says, its target “must be capable of suffering.”Footnote ⁴⁶ However, there are other highly plausible accounts of the nature and justification of punishment. In short, according to a consequentialist account, the reasons we punish are grounded in the goods to be achieved and not in any type of treatment supposedly deserved by the target. In everyday situations, and in many legal cases, the target’s capacities for suffering are simply irrelevant to why we must, say, imprison those who cause great harm. Sparrow claims that “to serve as punishment” our treatments “must evoke the right sort of response in their object.”Footnote ⁴⁷ Yet, with a moment’s consideration of a smiling psychopath behind bars, we know this to be false. In other words, some who are punished suffer the consequences, even if they cannot experience the suffering of a natural moral agent.Footnote ⁴⁸

To conclude the present section, several points are worth reiterating. First, we punish individuals, very often because of the goods to be achieved. Punishment, indeed, serves as a key mechanism by which we hold individuals responsible. At times, locating responsibility is a matter of identifying a property in the target—say, the capacity for grasping and acting freely upon moral reasons—which turns out to resemble the search for moral agency. At other times, responsibility is a process, whereby we respond with attitudes and practices, and where a direct link to moral agency is unnecessary.Footnote ⁴⁹ What is certain, at least, is that moral responsibility is not a singular nor unified enterprise. Luckily, the complexities can help us to locate responsibility in novel ways and in novel beings.

How We Can and Cannot Hold Machines Responsible

With this final section, my suggestion is not that the various ways of understanding responsibility necessarily apply to AI systems. Nor is it my goal to show that the process view of responding naturally to one another can be neatly mapped onto AMAs. Instead, just as AMAs provide new conceptions of moral agency, I aim to help us move beyond natural conceptions of moral responsibility by showing that, often, our attitudes and practices are adaptable to AI systems.

In his renowned essay “Two Faces of Responsibility,” Gary Watson articulated our ambivalence toward agents that appear responsible but are not appropriately held responsible.Footnote ⁵⁰ We can, for example, attribute a harmful action to a person’s character while stopping short of engaging in other, more overt blaming practices. This might be due to a recognition that the potential blamee was raised in a world (an abusive family, perhaps) where the harmful action (say, insulting others) was a normal part of life. As a result, he may be seen as a bad person in light of his propensity to insult others, but outwardly resenting him for it seems inappropriate. What these sorts of cases show is that locating responsibility comes apart into two “faces.” On the one hand, we can assess an agent’s character by thinking some conduct is attributable; on the other hand, we can hold the agent accountable by openly blaming, praising, or otherwise maintaining a “readiness to respond.”Footnote ⁵¹

Expanding this pluralistic approach, David Shoemaker shows, first, that one can be responsible in the attributability sense when an action stems from his character. Here one reveals the things that truly matter to him, what Shoemaker calls “care-commitment clusters.”Footnote ⁵² Given this process of revealing one’s cares and commitments, an agent can be the object of others’ admiration or disdain, since these sorts of responses target one’s character. Secondly, one is responsible in the accountability sense when he is appropriately targeted with blame as a result of exhibiting poor regard for others. Typically, he has ‘slighted’ someone, namely by being inconsiderate of others’ perspectives or insensitive to others’ fortunes, both representing failures of empathy. Finally, one is responsible in the answerability sense when he is called upon to defend or ‘answer for’ some exercise of judgment. Here one must be capable of maintaining and citing reasons. Where one has this capacity but fails to properly utilize it, those who demand answers will naturally disapprove.

Although I have given only a very rough depiction of contemporary, pluralistic views of responsibility, we can nonetheless see how such accounts allow us to make sense of a range of questionable cases, from psychological abnormalities to many commonplace interactions. We might, say, disapprove of a friend’s decisions when she fails to show up for a weekly meeting; but if her failure is a rare occurrence, surely we do not hold her in contempt. Naturally, we might say ‘This isn’t like you!’ as we demand to hear her reasons. In other words, we take her to be answerable for her conduct, yet we would not likely attribute the action to her in a deep sense, given that her failure to show up does not reflect her underlying cares or commitments. Still, considering that she has the capacity to maintain cares and commitments, our friend is seen as responsible in terms of attributability. Because she is able to entertain and act upon various reasons, she is responsible in terms of answerability. And given that she could be appropriately targeted with blame based on her regard, she is responsible in terms of accountability.

Where one is lacking in any of the usual natural capacities, we find exemptions from moral responsibility.Footnote ⁵³ Depending upon the exemption in question, our responses will likely display a sense of ambivalence. Psychopathy, for example, is given a robust explanation on a pluralistic theory. According to Shoemaker, a person with disorders characterized under psychopathy is often exempted from our accountability responses (like anger and resentment) given that they are typically incapable of empathizing. At the same time, this person might effectively display cares (namely for themselves) and may be capable of considering reasons (at least in non-moral decisions); thus, they can be considered responsible in the attributability and answerability senses.Footnote ⁵⁴

Pluralistic views of responsibility make sense of our common practices while providing theoretically rich understandings of individuals at the margins of moral agency, in Shoemaker’s terms. My suggestion, then, is that with this framework we might come to better understand moral responsibility in artificial moral agents. Note that my aim here is not only to locate responsibility for AMAs. This mechanism can be seen in the consideration of others—like the designers or users—taking responsibility for a machine’s behavior, as discussed above. Instead, my focus is also on what it would look like to hold AMAs themselves responsible.

Recall Sparrow’s hasty assumption that punishment, by its nature, aims at retribution. As I claimed, we can also punish those who harm—and reward those who help—in order to achieve certain results. Hence, incarceration is often referred to as rehabilitation, particularly when we want to emphasize a hope and belief that the individual in question can progress toward a more desirable condition. We want to improve upon the past, help others to learn from mistakes, and assure that similar behaviors are not repeated. In our more proactive efforts, we might educate early and encourage desirable behaviors before mistakes are made. The point here is that, in some sense, these practices are ways of holding others to account. We are ascribing a sort of responsibility when we punish (or reward) with the expectation that the target of our treatment will improve upon (or maintain) existing behavioral patterns.

Under the present description, the process of holding others to account is not the type of responsibility defended by Shoemaker, as his notion was grounded in the empathic capacities of the target. Indeed, his was a notion of accountability for natural moral agents. Nonetheless, the basic mechanism can be effectively adapted for our responses to AMAs. Contrary to Sparrow’s arguments, we can punish AMAs. We can impose sanctions on an AMA’s domain of application, restrict its previously authorized behaviors, or work to rewrite any deviant or undesirable lines of code.Footnote ⁵⁵ No doubt, these efforts will not cause it to suffer in anything like the ways we are capable of suffering. But, as I have argued, searching for the target’s capacities for suffering or consciousness is a means by which we locate agency and not necessarily the means by which we hold others responsible. While AMAs cannot suffer like us, they can and should suffer the consequences of carrying out harmful behaviors. AI systems capable of functional morality might one day learn from and improve upon their unique mistakes, as a sort of reinforcement learning.Footnote ⁵⁶ Importantly, it will be largely up to us—natural moral agents—to respond and signal when something like a moral transgression has occurred, and for this reason we cannot sit idly with the conviction that AI cannot be held responsible. Considering the sophistication of existing sensory and computing capacities in today’s AI systems, there are reasons to think that machines soon to be among us will be capable of recognizing and learning from our moral attitudes and practices—our anger and blame, gratitude and praise—perhaps just as effectively as from our simpler, non-moral commands.Footnote ⁵⁷

To summarize, what I have suggested is not only that it is possible to locate a sort of responsibility in AMAs—albeit, a non-natural sort of responsibility—but that we already possess mechanisms by which we can resolve the search. In particular, looking to recent pluralistic accounts, we see a range of types of responsibility and unique ways of holding others responsible, even those at the outermost margins of moral agency. Granted, some sorts of responsibility apply much less coherently to AMAs than others. It might seem odd, for example, to think we could attribute some action to an AI system’s underlying character. Still, our moral attitudes and practices are adaptable and will likely continue to evolve. For the time being, we can hold some machines to account, namely by engaging in consequential-based forms of punishment or reward. Likewise, we can demand that an AI system’s associates—its programmers or users—take responsibility, even where these individuals could not have controlled or foreseen the machine’s behavior.Footnote ⁵⁸ Undoubtedly, as an objective property, moral responsibility is most aptly identified only in natural moral agents, namely ourselves. That is to say, on a strict property view of responsibility, it may be that we cannot hold machines responsible. Nonetheless, as a procedural matter, we can coherently interact with AI systems—particularly those that are being developed to respond accordingly—in ways that assign a sort of responsibility. And although these mechanisms surely do not indicate the presence of our usual, natural sense of moral responsibility, they can function in very similar ways. If we are willing to grant, as some now are, that certain AI systems are artificial moral agents in virtue of engaging in functionally moral behaviors, we must likewise recognize that our moral attitudes and practices are adaptable in ways that serve to locate a sort of artificial moral responsibility.