Are We Asking the Right Questions About Artificial Moral Agency?

by Fabio Tollon

Human beings are agents. I take it that this claim is uncontroversial. Agents are that class of entities capable of performing actions. A rock is not an agent, a dog might be. We are agents in the sense that we can perform actions, not out of necessity, but for reasons. These actions are to be distinguished from mere doings: animals, or perhaps even plants, may behave in this or that way by doing things, but strictly speaking, we do not say that they act.

It is often argued that action should be cashed out in intentional terms. Our beliefs, what we desire, and our ability to reason about these are all seemingly essential properties that we might cite when attempting to figure out what makes our kind of agency (and the actions that follow from it) distinct from the rest of the natural world. For a state to be intentional in this sense it should be about or directed towards something other than itself. For an agent to be a moral agent it must be able to do wrong, and perhaps be morally responsible for its actions (I will not elaborate on the exact relationship between being a moral agent and moral responsibility, but there is considerable nuance in how exactly these concepts relate to each other).

In the debate surrounding the potential of Artificial Moral Agency (AMA) this “Standard View” presented above is often a point of contention. The ubiquity of artificial systems in our lives can often lead to us believing that these systems are merely passive instruments. However, this is not always necessarily the case. It is becoming increasingly clear that intuitively “passive” systems, such as recommender algorithms (or even email filter bots), are very receptive to inputs (often by design). Specifically, such systems respond to certain inputs (user search history, etc.) in order to produce an output (a recommendation, etc.). The question that emerges is whether such kinds of “outputs” might be conceived of as “actions”. Moreover, what if such outputs have moral consequences? Might these artificial systems be considered moral agents? This is not to necessarily claim that recommender systems such as YouTube’s are in fact (moral) agents, but rather to think through whether this might be possible (now or in the future).

It should be clear that based on the Standard View of agency above, conceiving as it does of agents as entities capable of intentional action, the bar for inclusion in this set is quite high. The kind of intentional action required for agency is often argued to depend on other necessary conditions that only we humans seem to possess. This often includes reference to capacities such as phenomenal consciousness, rationality, free will, autonomy and moral responsibility, all of which seem to fall under various “intentional” descriptions. However, there are also those who argue that the requirements listed above do indeed set the bar too high. Broadly speaking, these are the two methodological approaches taken by authors in the debate surrounding AMA: those who use “intentional” criteria and those who propose we make use of “functionalist” criteria.

On the intentional side of the debate are those who argue that the above requirements (phenomenal consciousness, rationality, etc.) are necessary for moral agency. Functionalists, however, argue that these criteria set the bar too high. In a recent paper that provides a fantastic orientation to the topic, Behdadi and Munthe claim that

“The functionalist view is that agency requires only particular behaviours and reactions which advocates of the standard view would view as mere indicators of the capacities stressed by the standard view.”

Functionalists, therefore, loosen the requirements that seem to rest on criteria that are not straightforwardly amenable to third-person verification (such as internal mental states). They claim that there are certain epistemic issues surrounding the use of internal mental states to ground the concept of moral agency. An argument often put forward here is the claim that the requirement of internal mental states runs up against the problem of other minds: yes, it is true that we all experience ourselves as conscious beings, but how can we justify the universal belief that others have minds like ours? The best we can do, in the presence of such problems, is to focus on observable features in our “grounding” of moral agency. The functionalist alternative, therefore, seems to lack the metaphysical baggage associated with the “intentional” accounts.

In light of the above, while it seems as though the functional account is more likely to accommodate the emergence of a genuine AMAs, this does not necessarily say much about its philosophical tenability. As Behdadi and Munthe argue,

“our impression is that there is a multitude of (often underexplained) concepts of moral agency and that many proposals are therefore much less in conflict than debaters assume.”

The reason for this is that in the existing literature on moral agency, one finds authors using a variety of novel concepts in order to argue for or against the possibility of AMA. This might be done using either functionalist or intentional frameworks, but often involves adding other conceptual variants. Examples are arguments for hierarchies of agents, degrees of agency, different kinds of agents, surrogate agency, and virtual agency, to name but a few. These attempts at conceptual precision, when taken together, make navigating the conceptual terrain incredibly difficult. On the one hand, it seems authors cannot agree on the correct metaphysical framework to adopt for identifying agency, which makes fruitful discussion difficult. On the other hand, it seems as though the conceptual variations introduced do not edify our understanding of agency, but rather make an already complex terrain untraversable.

It therefore seems that the current debate, at least in a descriptive sense, is at an impasse. Here Behdadi and Munthe offer a useful suggestion: that we move the AMA debate into straightforwardly normative territory. Instead of asking about the necessary and sufficient conditions required for agency (which both the functionalist and intentional accounts above attempt) we should look towards the extent to which such artificial systems should be incorporated into human social practices. In other words, instead of arguing about what the correct theory of moral agency might be, we should look towards the practical instantiations of artificial systems. This would involve, for example, evaluating the potential consequences of AI systems in relation to real world applications such as sex robots, robotic caregivers, military drones, and the nature of work. It is perhaps possible to perform such normatively grounded analyses without a prior determination of whether AI systems are genuinely artificial agents or not. We can look at the way in which they interact with the world, and make our moral evaluation of their behaviour based on this. This might involve looking towards the value of the task the machine is performing, the potential consequences of this task, and the implications of further integrating the system with human social practices. In other words, we can look towards the social and political consequences of systems that act with increasing levels of independence from human control. Such an analysis can be performed without even using the term “moral agent”.

This change in methodology also changes the kinds of claims we can make about artificial systems. Such a normative orientation means definitional squabbles about whether or not a given system is a “moral agent” or not become unimportant. Our focus can, therefore, shift to the ways in which responsibility  is distributed across our moral landscape: humans often work in close collaboration with artificial systems, and so agency becomes a product of both biological and material processes. Artificial systems are often used to inform human decision making (for better or worse), and there is an important sense in which we need to ensure that humans remain in the loop when it comes to such decisions. By moving the debate into straightforwardly normative territory, we can focus on ensuring that artificial systems are designed to promote socially beneficial values.

This is not to say that this methodological shift is not without its problems. While it does sidestep a number of issues overtly descriptive accounts face, it might be argued that there is a loss of conceptual precision. For example, it seems that even when talking in strictly normative terms, we would still need a coherent conception of agency, especially if we want to make claims about moral responsibility (even in a distributed sense). It matters, in this distributed and systemic sense, which entities are agents and which are merely tools used by agents. In order to properly cash out responsibility ascriptions (especially of the moral kind), it needs to be the type of responsibility that only agents can have. But to do this, perhaps we would first need to figure out what exactly agents are (taking us back into the descriptive domain).

However, this criticism does not fully capture what shifting the AMA debate into the normative domain would actually entail. Just because we are asking normative questions, this does not necessarily imply that we cannot indulge in conceptual discussion. Instead, what is actually going on when we approach AMA from this practical standpoint, is that we inquire why and when moral agency and moral responsibility are important concepts. Subsequently, we can then make a determination as to what requirements ought to be meted out with respect to these concepts.

It is perhaps possible for us to work on both issues at once: attempts at providing detailed accounts of agency need not always be in opposition to normative investigations. Those engaging in normative research (as this relates to AMA) could perhaps adopt an “as-if” approach, looking at the various ways artificial systems affect human social and political life, without necessarily needing a fully developed notion of agency. Those engaging in descriptive ethics, however, should be mindful of the issues raised here, and seek to  engage with the normative suggestions listed above more fully.