Introduction
Even a casual viewer or reader of science-fiction will doubtless have encountered the image of the maltreated machine. From the abused simulants of Blade Runner to the neglected child-robot “David” in A.I. Artificial Intelligence, we seem to have little difficulty in imagining artificial beings feeling pain and distress, and even empathizing with them. The kind of advanced artificial intelligence (AI) we see in fiction is of course far removed from the capabilities of real-world machines. However, as the capacities of AI continue to improve, interest has grown within the AI ethics community concerning the question of whether—and when—artificial beings may reasonably come to possess or demand some form of moral status.Footnote 1 This paper aims to engage with a specifically epistemological aspect of this question, namely what sort of evidence or criteria could or should be used to determine whether an AI possesses moral status, and specifically a form of moral status I term psychological moral patiency. In the next section, I briefly spell out this notion of a psychological moral patient, distinguishing it from other grounds of moral obligation. In Section “Evidence of Psychological Moral Patiency,” the main part of the paper, I consider five possible strategies we might adopt for identifying artificial psychological moral patients, namely (i) reliance on intuition and empathy, (ii) intelligence-based measures, (iii) the presence of autonomous desires, (iv) the presence of empathy, and (v) behavioral equivalence to beings to whom we already take ourselves to have moral obligations. Although all of these strategies have their merits, I go on in Section “The Cognitive Equivalence Strategy” to argue for a variant of (v) that I term the cognitive equivalence strategy, which suggests that we assign moral status to artificial beings based on whether—and how closely—they instantiate cognitive capacities that our best current science attributes to other beings we already take to qualify as moral patients.
Moral Relevance and Moral Patiency
In everyday ethical decisionmaking. we typically understand ourselves to have a variety of sources of ethical obligation, including those arising from the demands of friendship, family, personal integrity, political commitments, and religious or spiritual imperatives. Among these obligations, however, one important class comprises the obligations we have to other beings in virtue of their possession of intrinsic interests, such as the interest in not being made to suffer or not having their autonomy violated. These obligations extend to our fellow humans as well as to at least some nonhuman animals, but not—typically at least—to inanimate objects such as rocks, cars, or houses. I will use the term moral patients to refer to beings who are capable of exerting moral obligation on us in virtue of the possession of such intrinsic interests.Footnote 2 This account of a moral patient is still quite broad, and potentially includes beings like trees, flowers, and seeds that lack psychological states like attitudes or experiences yet may possess interests relating to their biological function such as growth and development. Although it is a matter of debate whether such interests alone justify the ascription of moral patiency to a being,Footnote 3 I will set the matter aside in what follows, and focus specifically on what I term psychological moral patiency, that is, a form of moral status that may arise in virtue of a possession of specific psychological capacities such as sentience, autonomy, desires, and so on. I would suggest that this notion of psychological moral patiency is of great relevance to much of our moral decisionmaking, insofar as capacities for undergoing suffering or exercising autonomy are relevant to many difficult decisions we make. The question of whether a being is a psychological moral patient is especially salient when interacting with nonhuman animals, as well as people in persistent vegetative states or comas whose residual capacity for psychological states may be uncertain. When David Foster Wallace famously wondered whether lobsters feel pain,Footnote 4 for example, or when neuroscientists assess whether a person in a persistent vegetative state is minimally conscious,Footnote 5 these inquiries can be understood as at least partly concerned with whether the being in question qualifies as a psychological moral patient. Although I take the notion of psychological moral patient to be important for ethical decisionmaking, I do not assume that only psychological moral patients exert moral obligations upon us. As noted, some living things might qualify as moral patients in virtue of their possessing biological functions that we should endeavor to respect. Likewise, one might think that some inanimate objects like rivers and mountains and sacred artefacts qualify as “final goods” (in the sense of Christine Korsgaard 1983Footnote 6) and matter in their own right despite lacking clearly definable interests and thus failing to be moral patients in the sense given above. It is also not obvious that we have obligations toward a being just in virtue of its being a psychological moral patient. One might think, for example, that it is possible to recognize a being as sentient or autonomous while denying we have actual obligations to them. To give an extreme example, a moral community might believe it has no obligations towards outsiders while nonetheless recognizing them as fully capable of suffering and being otherwise harmed. Likewise, a retributive judge might determine that a heinous criminal has waived all right to moral treatment by others. More mundanely, we sometimes recognize ourselves as having additional obligations to individuals that do not simply arise as a function of their status as moral patients, for example, the special obligations that we commonly take ourselves to have toward family and friends. There is thus room for slippage between the notions of being a psychological moral patient and being an entity to whom we bear moral obligations. With these distinctions in place, I can clearly state the question to be addressed in what follows, namely what sort of evidence could lead us to reasonably believe that an artificial system should be considered to have moral status just in virtue of its possessing appropriate psychological states. This is distinct from the question of whether an artificial system could come to have moral relevance in a broader sense (as might be the case for a particularly beautiful or significant technological artefact), and at least somewhat distinguishable from the question of the specific kinds of moral obligations we might bear toward machines. In relation to this second issue, it would be remiss not to note in passing the rich “relation-based” account of moral obligation provided by David GunkelFootnote 7 and Mark CoeckelberghFootnote 8. This approach stresses the primacy of interactions and relations with beings in giving an adequate account of our moral obligations to them (see also Cora Diamond, 1978, for a similar account concerning animals7). In what follows, I will not pursue this relational strategy, instead loosely following what Coeckelbergh8 calls a “properties approach,” which is concerned with how a being’s possession of particular properties such as sentience or autonomy relates to their moral status. However, I leave open the possibility that a relational approach will be ultimately necessary for an adequately rich account of the nature and scope of our obligations to machines.
Evidence of Psychological Moral Patiency
I now turn to the primary concern of this paper, namely what sort of evidence might reasonably lead us to believe that an artificial being qualified as a psychological moral patient, and thus as an entity capable of exerting at least one sort of moral obligation upon us. I should note at the outset that some would affirm that no sort of evidence could license this conclusion, and that no artificial being could ever thus qualify as a moral patient. This position may even seem intuitive; as David Levy9 notes, “[t]o many people the notion of robots having rights is unthinkable.” It is worth distinguishing, however, between two negative positions regarding artificial psychological moral patiency. First, one might grant that any robot that was psychologically equivalent to a human would qualify as a psychological moral patient, but deny that such equivalence is in practice possible, perhaps on similar grounds to those famously raised by John Searle.10 Second, one might deny that even a robot that was psychologically identical to a human would qualify as a moral patient, perhaps because it lacked our evolutionary history or was not a living thing. Both of these positions are worth taking seriously, although I will not engage with them in any detail in the present context.Footnote 9 However, I will note that both face challenges and involve controversial commitments. The first position, for example, will likely require (or follow from) the rejection of functionalist and computational accounts of mentality, thus making it unpalatable for many contemporary philosophers of mind, and faces famous “slow replacement” objections in which parts of a human brain are gradually replaced by silicon components.Footnote 10, Footnote 11 For its part, the second position is vulnerable to the charge of “speciesism” (in the sense of Peter Singer, e.g., 2009Footnote 12). Setting aside such doubts, let us consider some different types of evidence that might lead us reasonably conclude that a given robot or computer program constituted a psychological moral patient. Specifically, I will consider five possible routes to such a conclusion, namely intuition and empathy, intelligence, autonomy and desires, sentience, and behavioral equivalence, noting problems that arise in each case. Note that although I identify limitations of each of these routes, these are not intended as knockdown objections; what follows is intended more in the spirit of a survey than a set of dispositive arguments.
Intuition and Empathy
As noted at the outset of the present inquiry, we have little difficulty in empathizing with artificial beings in fictional contexts. Consequently, one might imagine that no grievous epistemological difficulties will arise if and when we construct artificial beings with genuine claim to moral status: we will simply perceive that they possess mental states.Footnote 13 It certainly seems true that in many cases, we immediately and automatically empathize with artificial beings. Gunkel,Footnote 14 for example, citing studies by Astrid Rosenthal-von der Pütten et al.Footnote 15 and Yutaka Suzuki et al.Footnote 16 notes that “researchers found that human users empathized with what appeared to be robot suffering even when they had prior experience with the device and knew that it was “just a machine.” This empathizing tendency is borne out by the actions and reports of those who interact with robots on a daily basis, such as soldiers and rescue workers serving alongside “Packbots,” military robots primarily used to identify and disarm improvised explosive devices. Gunkel notes that these AIs are frequently treated as “fellow combatants,” with soldiers “giving them names, awarding them battlefield promotions, risking their own lives to protect that of the robot, and even mourning their death.” However, as these examples suggest, to the extent that we rely on human empathy as our primary arbiter of whether an artificial system constitutes a psychological moral patient, we run the risk of “false positives,” misattributing moral patiency to relatively simple systems that lack any psychological basis for such a status. One might, of course, consider this in a positive light: given our disposition to over-attribute mentality to inanimate objects, it may be almost certain that we would recognize any genuine artificial psychological moral patients as such. False positives are not free from costs; however, as noted by Joanna Bryson,Footnote 17 inappropriate identification with robots involves “economic and human consequence of time, money, and possibly, other finite resources being given to a robot that would otherwise be spent directly on humans and human interaction.” Moreover, the fact that we will likely commit false positives in our empathetic identification with artificial beings does not rule out the possibility of false negatives. The specific traits and patterns of behavior that we naturally associate with the possession of morally relevant psychological states are likely to be somewhat parochial, tracking those forms of appearance and behavior that are readily assimilated to our own. As Eric Schwitzgebel and Mara Garza note, “human beings are much readier, from infancy, to attribute mental states to entities with eyes, movement patterns that look goal-directed, and contingent patterns of responsiveness than to attribute mentality to eyeless entities with inertial movement patterns and noninteractive responses. But, of course, such superficial features need not track underlying mentality very well in AI cases.”Footnote 18 This prompts the worry that we may fail to identify genuine artificial psychological moral patients in cases where they are too alien or too exotic to elicit our standard empathetic responses. These dangers may be particularly salient in the case of AIs that lack an embodied form or easy visualization.Footnote 19 Consequently, reliance on intuition and empathy alone seems like a risky strategy for ensuring appropriate moral responses to artificial beings.
Intelligence
A second source of information we might use to rationally ground attributions of psychological moral patiency to artificial systems is intelligence. This is clearest in the case of human-level AI, as demonstrated by the Turing Test.Footnote 20 Were we to build a system with the complete repertoire of our cognitive abilities and with the ability to converse indistinguishably from a human being, we would have few grounds for denying it substantial moral status (however, see Robert Sparrow 2004Footnote 21). Although human-level intelligence will satisfy many as a sufficient condition for psychological moral patiency, however, it is far less clear that it is a necessary one. For example, most of us consider many nonhuman animals to be psychological moral patients capable of exerting consequent moral obligations upon us, despite falling short of our cognitive abilities in many domains. It seems conceivable and perhaps even likely, then, that the first AIs with a genuine claim to moral consideration in virtue of their psychological capacities will nonetheless lack human-level intelligence. However, there are potentially useful intelligence-based measures of psychological moral patiency besides the Turing Test. In a future era of robotics, for example, we might administer “intelligence tests” to artificial systems, and use these to guide our attributions of moral patiency to them. Needless to say, this is a proposal fraught with difficulties. To begin with, the history of intelligence testing as a guide to ethical decisionmaking is famously dubious, beset with prejudice and racial bias, and we should not expect the use of intelligence testing in regard to artificial systems to be immune to such forms of prejudice. Moreover, it is worth noting the somewhat complex relationship between intelligence and psychological moral patiency. It certainly seems to be the case that we are more willing to extend moral status to animals with impressive cognitive abilities: while few would hesitate to grant moral status to dogs, dolphins, and chimpanzees, creatures with smaller and less complex nervous systems such as insects and crustaceans constitute a gray area in our folk morality. However, it is not clear that it is intelligence per se that motivates this association, so much as the inferred link between cognitive complexity and sentience (see section “Sentience”). Footnote 22 And certainly, most of us would not endorse the use of relative intelligence as a primary criterion for making decision between the ethical claims of different individual humans. A final worry for the use of intelligence as a criterion for psychological moral patiency concerns the challenge of even giving an adequate account of the notion. Although Robert Sparrow, for example, may be right to say that “we seem to have a firm intuitive grasp of what intelligence is,”Footnote 23 translating our pretheoretical concept of intelligence into one that can be operationalized and applied to systems quite different from us is deeply challenging, as reflected by the confusion about the very meaning of the term “artificial intelligence.”Footnote 24 Current artificial systems dramatically outperform humans in a variety of cognitively demanding tasks, from arithmetic to chess and (certain forms of) image recognition, yet none are plausibly as intelligent as a human or even most animals.Footnote 25 Quite apart from the normative challenges discussed above, then, any intelligence-based measure for assessing psychological moral patiency faces the unenviable if not impossible task of determining how to weigh and synthesize the myriad forms of cognitive proficiency into a unified measure of a system’s overall intelligence.Footnote 26
Autonomy and Desires
Rather than focus on intelligence tout court, we might instead attempt to use evidence of some specific aspect of a machine’s intelligence as a guide to its status as a psychological moral patient. There are of course many such aspects of intelligence we might appeal to, such as its flexibility, creativity, or memory, but for present purposes, I will consider whether autonomy might constitute an appealing criterion. This is a suggestion persuasively developed by Erica L. Neely,Footnote 27 who identifies autonomy as a measure for whether a system possesses genuine desires which in turn would ground moral status.
[C]onsider the case where the agent’s goals are not always determined by an outside source, i.e., where the agent is capable of determining its own goals at least some of the time. In this case, the agent is expressing a basic capacity for autonomy, which implies that these goals must be chosen by the agent itself.
The possession of such desires, she argues, gives us grounds for assigning a system some kind of moral status; “while the agent’s desires may be overridden, they may not simply be ignored.” There are interesting similarities between Neely’s position and some leading positions in animal welfare ethics. Marian Dawkins,Footnote 28 for example, argues that we can construct a science of animal welfare based on the satisfaction of preferences. If an animal has very strong preferences that are not being met—for food, mating, or a certain kind of environment, say—then their individual interests are not being satisfied. As she puts it, “‘having what they want’ is a shorthand way of covering a wide variety of way in which animals can, in a publicly observable way, show us whether the environment they are in is positive (something they like and want to continue with) or negative (something they dislike and want to escape from or avoid).”Footnote 29 The positions of both Neely and Dawkins are well-developed and sophisticated, and a detailed response lies beyond the scope of the present paper. Moreover, Neely acknowledges the difficulties involved in assessing whether a given system is truly autonomous or not, suggesting that we should treat such as assessments as probabilistic. Nonetheless, I would note that the use of autonomy as a criterion for moral patiency faces some key worries. For one, as Neely herself notes, it is not a necessary condition for patiency: a system may be sentient and thus capable of suffering despite lacking endogenous desires (perhaps having its attitudes directly “programmed in”). Perhaps a more fundamental issue, however, comes from the fact that there are many beings that at least superficially exhibit autonomy yet lack any claim to serious moral consideration. The aversive responses of nematode worms, the chemotaxis of bacteria, and the phototaxis of plants, for example, could all be loosely classified as autonomous desires in the broadest sense. Even if we take these behaviors as constituting some minimal claim for moral status, it is a claim that would be outweighed by even fairly trivial instrumental considerations. To use an example of John Basl,Footnote 30 were there to be some scientific interest to be gained from growing and destroying a million maple trees, few would balk at the experiment on moral grounds, assuming it generated no significant negative externalities. What is required for an account of moral patiency founded in autonomous desire to have real ethical force is some means of distinguishing between mere appetitive and aversive behaviors and robust preferences. Some philosophers in the preference utilitarian tradition (including Peter SingerFootnote 31 and Gary Varner25) limit moral patiency to conscious or sentient beings with preferences, but this would of course be at odds with Neely and Dawkins’ goal of providing a separate criterion for moral patiency.Footnote 32 It is of course possible that ongoing work in cognitive science may identify a principled criterion for distinguishing morally salient “thick” preferences from simpler forms of appetitive and aversive behavior and learning. However, absent such a criterion, an approach to moral patiency founded on the use of autonomy as a guide to possession of morally significant desires is at best incomplete.
Sentience
Perhaps the most obvious criterion for artificial psychological moral patiency is sentience, defined in the present context as a capacity for undergoing conscious states like pleasure and pain that feel good or bad to them; in the terminology of contemporary psychology, a capacity for undergoing positively and negatively valenced conscious states. Even if we do not wish to claim that this is “all there is” to psychological moral patiency, certainly it is natural to think that if a being can feel pain, it has at least some potentially morally salient interests, namely not feeling pain. Applying this to artificial entities, then, the notion of sentience might provide a conceptually straightforward pathway to moral patiency: if and when we build machines capable of consciously experiencing pleasure, pain, or other valenced states, we will have grounds for recognizing them as moral patients. There are several attractive features of this approach. First, as noted, the claim that sentience grounds at least some minimal form of moral patiency is relatively uncontroversial; even if we wish to allow there is more to well-being than pain and distress, it is hard to deny that such experiences are often detrimental to our interests and those of nonhuman animals. Second, a sentience-based approach to establishing moral patiency can accommodate quite simple forms of sentient being. Thus, although we may not wish to consider fish or lobsters persons or as exhibiting autonomy in a rich sense, if it could be established that they could feel pain this would give us some grounds for potentially taking their interests into account. And as noted above, it seems antecedently possible or likely that long before we have achieved human-level AI with claims to more robust moral notions such as personhood or self-determination, we will along the way build systems with the cognitive complexity equivalent to that of nonhuman animals that might potentially qualify as moral patients in virtue of capacities to experience pain, fear, or other simple forms of negatively valenced state. Finally, a sentience-based approach to psychological moral patiency seems at least in principle empirically tractable, insofar as it identifies this status with the capacity to undergo specific kinds of psychological state which we might expect to be amenable to assessment via the tools of cognitive science. Despite this superficial simplicity, however, sentience-based approaches to moral patiency face serious challenges of both a philosophical and practical character. The most daunting is of course the challenge of how we can ever establish whether a given system is conscious. This is an area of intense debate within comparative psychology, but expert opinions vary wildly; some philosophers and scientist extend consciousness to all vertebrates and some invertebrates,Footnote 33 while others have suggested that organisms as relatively cognitive sophisticated as fish may lack the capacity to feel conscious pain.Footnote 34 There are even those (including Peter CarruthersFootnote 35) who believe the very question of nonhuman consciousness to be ill-posed. However great the challenges are for assigning consciousness to animals, the corresponding challenges for assigning it to machines will be far greater. In the case of animals, we can at least appeal to broad similarities in cognitive architecture and material composition, as well as a shared evolutionary history. By contrast, most artificial systems possess—and will continue to possess—wildly divergent forms of cognitive architecture and processing substrate from those found in human beings. Such controversies lead some including GunkelFootnote 36 to a pessimistic conclusion about the prospects of a science of artificial consciousness. As he puts it, the present science of consciousness is “unable to demonstrate with any certitude whether animals, machines, or other entities are in fact conscious (or sentient) and therefore legitimate moral persons (or not), [and] we are left doubting whether we can even say the same for other human beings.” Such extreme pessimism may not be warranted; considerable progress has been made in the last few decades on the development of better scientific theories of consciousness, with practical implications including better tools for the assessment of consciousness in patients in vegetative states.Footnote 37 Nonetheless, in light of the fundamental theoretical disputes ongoing among the many quite different scientific approaches to consciousness,Footnote 38 even the optimist must grant that a settled consensus in the field remains a distant goal. Consequently, and in light of the rapid progress in the development of AI, we should not too much hope in the idea that we will be able to rely on scientific measures of sentience to establish psychological moral patiency in artificial systems.
Behavioral Equivalence
Faced with the daunting challenges of assessing the occurrence of states like consciousness in other beings, a final recourse may be to appeal to some notion of behavioral equivalency. This is the approach developed by John DanaherFootnote 39 in the context of what he calls a theory of ethical behaviorism. Danaher claims that in our ordinary ascriptions of moral status to other beings (including other humans), we rely fundamentally on behavior as a source of evidence: even if someone takes moral status to involve capacities such as sentience or a soul, their evidence for ascribing these capacities, he argues, will ultimately derive from their observations of behavior. On this basis, Danaher argues for a criterion of performative equivalence in the assessment of artificial moral patiency: “if a robot consistently behaves as if it is in pain, and if the capacity to feel pain is a ground of moral status, then a robot should be granted the same moral status as any other entity to whom we ascribe moral status on the grounds that they can feel pain.” Behavioral equivalence as an approach to artificial psychological moral patiency has attractive features. It is not hostage to the fortunes of any future advances in cognitive science, but is instead immediately applicable: for any candidate artificial moral patient, we can simply ask whether its behavioral repertoire is equivalent to that of an entity to which we already grant the status of moral patiency. Likewise, it avoids some of the pitfalls of a purely intuitive approach, insofar as it relies not on our immediate empathetic responses to artificial beings, but on the (in principle) more rigorous criteria of behavioral similarity. Danaher’s position is again complex and well-developed, and the view I will ultimately defend below draws heavily upon it. Nonetheless, I would suggest that it has features that make it at least somewhat problematic. For one, it is limited in application to systems whose behavior is at least somewhat similar to that of beings to which we already attribute moral status: when no behavioral equivalency can be found, it must at best remain agnostic about questions of psychological moral patiency. For another, it is at risk of being “gamed”: given some criterion for behavioral equivalence (say, avoidance behavior), a dedicated computer scientist may be able to produce a robot or system with the capability of producing the relevant behavior yet doing nothing else.Footnote 40 In such a case, we might have little reason to consider the system to have any moral status, yet it would have satisfied our earlier criteria. These problems relate to a deeper one that Danaher himself notes, namely the challenge of determining the appropriate “performative threshold” for behavioral equivalency, in other words, how to determine the appropriate degree of generality or specificity to adopt when seeking to establish equivalence. For example, if a given stimulus applied to a robot served the function of negatively reinforcing a given behavior, but did not produce any outward signs of distress, should we consider it equivalent to a “punishment” or other negative state undergone by a human or animal? As Danaher notes, an excessively liberal threshold risks including systems within our moral circle inappropriately, while an excessively conservative one carries the danger that we will fail to identify morally significant harms. This challenge for the theory illustrates perhaps its key limiting factor, namely its commitment to reliance on behavior alone. Danaher stresses that he intends the term to be understood broadly, and to include “all external observable patterns, including functional operations of the brain.”Footnote 41 However, what it presumably cannot include is the theoretical vocabulary of cognitive science; things such as episodic memory, metacognitive representation, and affective states are not themselves entities that we can directly observe, but are instead mechanisms posited within the context of specific psychological theories. Yet I would suggest that it is by precisely by reference to these mechanisms that we can begin to adjudicate questions about the appropriate performative threshold to adopt in assessing behavioral equivalence. To give a toy example from comparative psychology, imagine that we find a deep-sea worm whose “pain behavior” is at least superficially similar to that of a creature we already consider to be a psychological moral patient—let us say a fish. The worm, we can suppose, recoils from damaging stimuli, and learns to avoid them in future. Is this enough to warrant ascribing it moral status? The answer may not be readily settled with appeal to behavior alone, but instead depends on our best theories of the creature’s internal cognitive architecture. If we were to determine, for example, that the behavior was accomplished entirely by peripheral mechanisms in the skin of the worm that did not (again, according to our best theory of the cognitive structure of the organism) communicate with its central nervous system, this might push us toward a negative answer. If, by contrast, the relevant stimulation was encoded in central short- and long-term memory stores, we might take more seriously the possibility that the creature underwent some morally significant negative event. Footnote 42 Addressing such interpretative challenges is likely to be particularly important if we wish to determine the occurrence of potentially morally significant events in artificial systems that differ from us considerably in respect of their behavioral repertoire. To give another toy example, imagine that an AI system responded to externally damaging stimuli by dumping the contents of its short-term memory. Could this constitute a form of distress or some other morally significant mental event? To address this issue, we would again surely need to ask questions about the structure and dynamics of the system’s cognitive architecture. For example, does the system have a centralized process for registering external events and responding to them? Does it possess a system of representations allowing it to “weigh up” the predicted occurrence of such events with other possible positive and negative inputs? Although such questions would of course not settle the issue of whether the specific event had any moral significance, in the context of a broader theory of the cognitive function of the system they could certainly serve to inform our considered judgment of the matter.
The Cognitive Equivalence Strategy
The five epistemic strategies described above all constitute possible sources of evidence that a given artificial system might qualify as a psychological moral patient. Although I have noted problems with each, I will reiterate that I do not take anything I have said to be decisive: all could be useful components of a broader toolkit for assessing the moral status of future artificial systems. However, of the five strategies considered, I would suggest that Danaher’s notion of behavioral equivalence provides the most directly promising path forward. In this final part of the paper, I wish to suggest how we might amend Danaher’s proposal so as to answer some of the worries just raised. Specifically, I wish to argue for a heuristic that I will term the cognitive equivalence strategy. In short, this states that we should we treat an artificial system as morally significant to the extent that our best science of its cognitive structure and dynamics attributes to it psychological capacities present in other beings to which we already assign moral status. This proposal clearly draws heavily upon Danaher’s view, but as I describe below, I believe it has some advantages. Before that, however, I would make some general observations about the cognitive equivalence strategy. The first is that it should be seen very much as a moral heuristic or rule-of-thumb: it is not a theory of psychological moral patiency, but a procedure we can apply that will allow us to make fallible but justified positive identifications of the moral status of an artificial being. Second, the theory is deliberately ecumenical and (relatively) theoretically neutral. It does not single out any one cognitive capacity—such as sentience, autonomy, self-awareness, and so on—as uniquely relevant for moral status, but considers them all as potential sources of evidence for psychological moral patiency. It should thus hopefully be acceptable to theorists with a broad range of positions. Finally, I would suggest that it is a strategy that should be seen as tightly coupled to contemporary cognitive science, and it is intended to accommodate new developments in our understanding of potentially morally significant psychological capacities in different biological or artificial beings. To spell out the strategy in a little more detail, then, imagine that we are considering whether some intelligent robot qualifies as an artificial psychological moral patient. The first question we would ask in this regard is what cognitive capacities we can attribute to the system. Does it possess working memory, a theory of mind, or metacognition? Can it engage in “mental time travel” and creative problem-solving? Does it possess an internal representational system for registering “desirable” and “undesirable” events? Answering these questions will be far from trivial, particularly if the system has not been fully programmed per se, but instead developed “organically” using machine learning techniques that render its internal structure somewhat opaque.Footnote 43 Nonetheless, just as comparative psychologists can make defeasible but informed judgments about the cognitive capacities of nonhuman animals, so too can we hope to make similar estimates about the capabilities of the artificial system. With such estimates in hand, we can now compare the system’s capacities to those of biological organisms, asking how its cognitive abilities compare to those of a rat, pigeon, or insect.Footnote 44 If and to the extent that the artificial system possesses relevantly similar cognitive capacities to those of such beings, we can then tentatively assign it moral status comparable to that we assign to them. In practice, of course, such comparisons will be awkward and incomplete. It is very unlikely that any future artificial system will possess the exact same cognitive mechanisms of, say, a dog, unless deliberately constructed to do so. Instead, we might expect it to exhibit greater cognitive abilities in some domains and less in others. This prompts the awkward question of which capacities matter and to what extent. Although I will not attempt to adjudicate this issue in any detail here, I would suggest that in practice such difficulties may be overcome via careful experiment and theory. Imagine, for example, that we believe chickens to be psychological moral patients. We can then ask which of their cognitive capacities, if absent in a specific chicken, would lead us to reasonably reduce our credence in their possession of this status. If a brain-damaged chicken were not to have mating instincts, for example, or lacked normal drives for hunger and thirst but was otherwise behaviorally identical to its conspecifics, we would not intuitively consider its moral status to be significantly different from that of other chickens. If, by contrast, it did not possess capacities for social reasoning, goal-directed agency, or had a dramatically diminished capacity for learning, we might revise downward our credences concerning its status as psychological moral patient. We could also use the study of the range of behaviors available to decapitated or decerebrated animals (who are unlikely to be conscious or otherwise qualify as moral patients) to constrain the relevant cognitive parameters of comparison. For example, it was demonstrated in 1886 by David FerrierFootnote 45 that a decapitated frog will still rub its thigh in response to the application of a drop of acetic acid. Similar findings in more recent years include that of Michael GentleFootnote 46 who showed that decerebrate chickens would still modulate pecking behavior in response to an injured beak if the injury preceded the removal of their forebrain, suggesting such behaviors are accomplished by relatively low-level mechanisms. Via a combination of empirical and reflective methods such as these, we might reasonably attempt to establish a set of cognitive dimensions relevant for moral patiency. I recognize that the cognitive equivalence approach introduces complications not present in Danaher’s behavioral equivalence account, and in particular, relies to a greater extent on the explanatory success of cognitive and comparative science. However, I would suggest that by licensing us to use broader forms of scientific explanation, it makes some otherwise opaque questions tractable. For example, as suggested by the arguments given in the preceding section, it allows us to move beyond behavior as a guide to moral equivalence, and instead to look for fine-grained comparisons at the level of cognitive mechanisms. This in turn may make it easier (though still far from trivial) to settle issues concerning appropriate equivalence standards. Thus, we may be able to appeal to the best current scientific theories and tests of things like memory, desire, agency, and motivation to determine if a system really has the relevant capacities, or simply exhibits behavior superficially indicative of possession of these. For example, episodic memory is defined by some psychologists as the ability to generate unified representations encoding the identity of an object together with the spatial location and time at which it was encountered, also known as the “what, where, when” criterion.Footnote 47 Such definitions (and associated experimental measures) can be utilized by the cognitive equivalence strategy, thus giving us at least an initial way of determining the appropriate degree of specificity to adopt for the purposes of determining moral commensurability. Additionally, the cognitive equivalence strategy may allow us an entry point into assessing moral patiency in quite exotic forms of AI whose behavior is radically different from that of humans and animals. As noted in the example of the “memory dumping” robot above, these cases may be hard or impossible to adjudicate on the basis of behavior alone, but by extending our theoretical vocabulary to include posited psychological mechanisms, we may be able to identify the internal processes of exotic forms of intelligence as falling under a broader cognitive category such as the registration of positive and negative valence that is realized in beings we regard as moral patient, enabling fruitful comparisons of their respective moral standing. I recognize, of course, that the cognitive equivalence strategy has important limitations. In particular, it relies on our existing judgements about whether various nonhuman animals constitute psychological moral patients. There is no reason to think that these assessments are infallible, or even particularly reliable, as suggested by the varied norms and legal protections enjoyed by different species in different jurisdictions. However, optimistically, one might hope that as comparative psychology progresses and greater attention is paid to animal welfare issues, these assessments will improve and become reliable. Via the cognitive equivalence strategy, such progress could be immediately and directly applied to our consideration of the moral status of machines. A final question worth noting is what the cognitive equivalence strategy would say about existing AIs. After all, it is not uncommon to read in machine learning journals of systems that exhibit capacities like agency, creativity, and even theory of mind. Should such ascriptions lead us to assign moral status to systems already in existence, as suggested by Tomasik?Footnote 48 Alas, a detailed response to this worry lies beyond the scope of the present paper, but I have argued elsewhere that such ascriptions should not necessarily be taken at face value,Footnote 49 and that in practice, the cognitive capabilities of current artificial systems pale in comparison to those of even simpler animals.Footnote 50 Nonetheless, it does not seem outlandish to speculate that the current gulf between artificial and nonhuman biological intelligence may recede quickly in light of new developments, and that the cognitive equivalence strategy might soon give us reason to give defeasible and perhaps probabilistic weightFootnote 51 to the claim that some artificial beings qualify as psychological moral patients.
Conclusion
In this paper, I have sought to address an epistemological question that I take to be of great relevance to our ethical treatment of future artificial systems, namely what sorts of evidence and which criteria we should use to determine whether they qualify as psychological moral patients. I suggested that none of the methods considered—including reliance on intuition, on intelligence, and on demonstrations of autonomy and sentience—offers an uncomplicated answer to the question. I argued that Danaher’s behavioral equivalence strategy offered a promising path forward, but offered a somewhat different formulation of the view in terms of cognitive equivalence that I believe overcomes some of its limitations.
