Machines with Faces: Robot Bodies and the Problem of Cruelty

The Reverse Turing Test

Recent popular culture texts such as the film Ex Machina (2014) and the HBO TV series Westworld (2016–) reflect a fascination with the ethical conundrums that would be produced by artificial personhood. How should we treat an entity that is simultaneously a manufactured object and a sentient being? The topicality of such a question can be explained by the often unjustifiably optimistic claims made by and for artificial intelligence research; the fact that the meaning of the term ‘artificial intelligence’ itself has become less and less clear over the course of its history; and the ways in which current software agents and robots are designed to encourage a misattribution of human-like capacities and motivations in observers. In reality, it is extremely unlikely that we will confront such an ethical challenge at any time in the foreseeable future, or perhaps at all; but our tendency to misrecognise machines as human-like is itself potentially significant.

There already exist arguments to the effect that even a robot with no claim to personhood could indirectly cause harm to persons (e.g. Darling, 2012; Sparrow, 2017; Whitby, 2008). Even if a robot cannot suffer, it might habituate people to behaviour that would cause suffering if directed towards a human being. It is not difficult to imagine a future in which some men use realistic robots to indulge violent sexual fantasies, for example, raising the question of what impact such a practice would have on their behaviour towards real women (Sparrow, 2017).

We could take such concerns further and imagine a reconfiguration of Alan Turing’s famous imitation game, or Turing test (Turing, 1950). Foundational to Turing’s test is a claim that, from the perspective of an individual person, the line separating human personhood from machine personhood is less clear than it might first seem. This is due to the fact that, even when we credit another human being with an interior life, we are ultimately giving them the benefit of the doubt in the absence of any physical proof. Personhood therefore can be considered to exist ‘in the eye of the beholder’ rather than as any objective attribute of the putative person, resulting in the imitation game where the strength of a machine’s claim to personhood is measured simply by the reactions of observers. ¹ But if our belief in the personhood of any other body – human or machine – ultimately results from a subjective attribution arising from our perceptions, could exposure to a sufficiently human-like robot alter this attribution in such a way that we feel less empathy or shared humanity with real human beings? By casting doubt on a posited equivalence between human appearance and personhood, this could create an inversion of Turing’s imagined scenario, in which some human beings become obligated to substantiate their claims to personhood, given that external identifiers would no longer be considered sufficient in themselves. Ironically, rather than expanding the boundaries of what we consider to be entities deserving of human empathy and dignity, then, such human-like robots could cause a shrinking of those boundaries such that even some human beings are no longer encompassed by them.

But could a robot ever appear sufficiently human-like to blur the boundaries of the human? The field of social robotics is founded on a belief that machines can be designed to sustain human-like relationships. In one study of young children’s interactions with a robot, when the robot laid itself down on the floor because its batteries were flat, the children would cover it with a blanket and say ‘Night-night’ (Tanaka et al., 2007: 17956); in another study, test subjects experienced a genuinely negative affective response to the ‘torture’ of a robotic dinosaur despite knowing that it was incapable of suffering (Rosenthal-von der Pütten et al., 2013). MIT roboticist Cynthia Breazeal’s Kismet robot was able to evoke seemingly genuine social engagement from human beings despite being ultimately nothing more than a collection of largely reactive external expressions and babbled gibberish (Breazeal, 2002), and Breazeal has more recently sought to introduce social robots into the home. ²

The human tendency to anthropomorphise is well-known, and there are abundant examples of one-sided social and emotional investment in non-sentient objects. This is possible because, as noted by Sherry Turkle, ‘With robots, people are acting out “both halves” of complex relationships, projecting the robot’s side as well as their own’ (2007: 504–5). However, while it might be possible for robots to evoke such responses to a particularly high degree (Darling, 2012: 4–7), no robot currently in existence can be considered to genuinely threaten the boundaries of the human; most robot designs do not even attempt to look persuasively human, as doing so would only highlight their inability to succeed. What would it take for a machine to make a genuine appeal to the kind of empathy we feel towards another human being?

There is, presumably, a broad spectrum of anthropomorphisation, running from swearing at a malfunctioning car to being genuinely distressed by the damaging of a robot. At what point could our attribution of human qualities to a robot be said to approach parity with their attribution to actual human beings? I would argue that the face is a Rubicon in the humanisation of machines. This claim might at first seem unpersuasive; after all, all manner of robots already have faces and yet do not seem convincingly human-like. But possession of a face requires far more than a plastic head moulded into an approximation of certain human anatomical features.

How to Make a Face

While non-human animals can use tools and even employ at least the rudiments of language, the face – in the fullest meaning of the word – is arguably the only uniquely human attribute. In this fullest sense, a face means more than simply an animal’s possession of a mouth or nose, or the painting or sculpting of these things onto the surface of a puppet or the chassis of an android. A face is a tremendously complex but also unified phenomenon synthesised from both physiological and cognitive components. It is anatomically both specific and sophisticated – only the human skull has its forward surface entirely covered by a dense web of musculature able to communicate through the deformation of the skin – but it does not simply inhere in physical features. A face is an interpersonal phenomenon; the purpose of a face is signification, and signification can only exist between actors (Black, 2011: 21–2). For a face to be a face, it must be meaningful for an external observer: whatever the musculature on the front of a particular head might do, it cannot function as a face unless it produces some response from another body.

The face is therefore a product partly of biology, partly of perception, and partly of culture. It is for this reason, presumably, that it is difficult to give a fulsome account of its nature. Deleuze and Guattari’s account of faciality treats the face as something entirely produced by cultural forces: ‘“Primitives” may have the most human of heads, the most beautiful and most spiritual, but they have no face and need none’ (1987: 176). But in doing this, they mistake an extensive system of subjectificatory structures that leverage the power of the anatomical and perceptual components of the face for the face itself (Black, 2011); faces are a part of the organisation of our bodies and minds that pre-dates these structures. For his part, Peter Sloterdijk criticises Deleuze and Guattari for their untenable assertion that the face is nothing more than a cultural construction (2011: 167–8), but ultimately treats the physical reality of the face as little more than an evolved adornment added to the exterior of the human body, like a flower on a plant. Founding his discussion in a consideration of artistic representations of the face, his understanding is no less confined to static, culturally specific representations than that of Deleuze and Guattari. A painting or photograph of a face is not a face, just as two camera lenses positioned like eyes on the ‘head’ of a robot are not a face; a face is an endlessly mobile, dynamic entity animated by a reciprocity between living bodies. This is not only because of faces’ dependence on observers who perceive them, but also the sympathetic interplay between faces, where expressions are constantly being shared or passed back and forth.

There is compelling evidence that humans (and primates) are born with an innate attraction to and capacity to recognise faces (Black, 2011: 1), ³ highlighting the degree to which faces appear through the interaction of specialised anatomical attributes of the faced body and specialised perceptual attributes of the facing body. But this innate orientation towards faces quickly develops into more than just a fascination with a particular configuration of shapes. Babies less than an hour old have been documented imitating the facial movements of adults (Meltzoff and Moore, 1983), suggesting that, more than simply being able to make sense of faces, humans possess an innate sympathetic link between our perception of other faces and our experience of having faces of our own. In addition, within weeks of birth the nature of faces’ visual attraction for infants seems to change in important ways (Powell et al., 2018).

By four months of age, while a baby will be powerfully engaged by the sight of her mother’s face delivered via a live, two-way video link, if that link becomes asynchronous her fascination will start to fade, such that it will ultimately come to exert no stronger claim on her attention than nearby inanimate objects (Stormark and Braarud, 2004). The baby’s fascination does not arise simply from the sight of the mother’s face, but rather from that face’s interplay with the baby’s own – the way facial expressions and their associated affective states move back and forth between the two faces and are produced collaboratively by them. When the video feed between baby and mother is out of sync, this interplay is short-circuited and something fundamental to the baby’s perception of her mother’s face vanishes.

This suggests that, while human beings are born with an innate attraction to the form of the human face, this is only the first part of a scaffolded process by which faces come to be perceived in a more complex way: the special significance of the face becomes dependent on direct, reciprocal interactions between the expressivity of one’s own face and those of others (see Powell et al., 2018). This stands to reason given that this ‘primary intersubjectivity’ is sustained by ‘affective proximity’ (Varga and Gallagher, 2012: 252), and the sharing of affect is most importantly sustained by the sharing of facial expressions. Before she has ever looked in a mirror or come to understand that she has a face in any conscious way, an infant enacts her faciality in a reciprocal relationship with the faces of others, and one only needs to see the wince of someone watching video footage of someone else having a painful accident to know that the sympathetic dimension of faciality continues beyond childhood. Experiments even suggest that having one’s facial mobility reduced makes it more difficult to interpret the expressions of others (Wood et al., 2016a, 2016b).

It is clear that human faciality arises through the interaction of multiple factors distributed across more than one body. Consequently, it stands to reason that there could exist a state of ‘partial faciality’, where some but not all of these factors are present. At best, current efforts to technologically reproduce the face might attain such a partial faciality, but for a robot to produce a genuine simulation of human faciality, it would need to go further than simply presenting a simulation of human facial features – even mobile and expressive ones. It would need to go beyond any previous technology of representation by creating a facial expressivity that exists in a reciprocal relationship with the faces of others. No existing representation of the face possesses a capacity to convincingly reproduce the dynamic, reciprocal character of a real human face, and this is precisely why philosophers such as Deleuze and Sloterdijk, whose accounts of faces are largely based on a consideration of film and portraiture respectively, are unable to provide fully elaborated accounts of human faciality in their work (Deleuze, 1986: ch. 6; Sloterdijk, 2011: ch. 2).

It is easy enough for something non-human to have a partial face. So easy, in fact, that partial faces can be thrown together randomly by natural forces, like the Man in the Moon or the face on Mars. ⁴ The role of human perception in the production of faces means that face-hungry human vision can misattribute faciality to anything even remotely approximating facial anatomical features. However, attribution alone is no more sufficient for faciality than physical features alone; just as something can’t be a face without an observer to perceive it as such, perception in the absence of real physical expressivity is also not enough. Faciality requires reciprocity: to see another’s expressions while that other sees your own so that their affective charge can be shared. It is common for young children to ‘hide’ by covering their eyes, and in the past this behaviour has generally been put down to toddlers’ solipsistic lack of understanding of object permanence. If the child can’t see anything, this explanation went, then the child believes that there is nothing there. This explanation was always problematised by babies’ love of peek-a-boo, in which they are endlessly surprised and delighted by the disappearance and reappearance of an adult who covers his or her own face, rather than the face of the child, but in any event research has established that this experience of invisibility is caused by the absence of eyes rather than the absence of sight (Russell et al., 2012). Young children understand perfectly well that their bodies remain visible while their eyes are covered, but nonetheless believe that their selves are inaccessible without mutual looking, and the selves of others are inaccessible under the same circumstances. Looking into the face is so crucial that when their faces are not visible, it seems that other people simply aren’t there.

It is also not enough for one body to possess the physiological capacity to express itself in certain ways alone, and another to possess the cognitive capacity to read certain kinds of meaning in that expression alone, the resulting combination of expression and perception supplying the two halves of the whole of faciality. If this were all that was required, we could imagine some hypothetical inter- or intraspecies relationship between two kinds of bodies, one with only the anatomical components of the face and one with only the cognitive components, a relationship in which facial communication only moves in one direction. In fact, such an inter-species relationship does exist: watching a dog chasing a ball in the park, the dog’s owner might look at the dog’s open mouth and widely pulled-back lips and perceive the dog as made happy by the game, but the dog is only maximising the flow of air over its tongue to cool its overheating body; dogs can’t smile because dogs don’t have full faces, even if the human cognitive equipment of facial perception sees a face on a dog’s head. We can similarly produce partial faces in other animals, but the dog is unique in that tens of thousands of years of co-evolution seem to have equipped it to use the limited mobility of the musculature of its head specifically to trigger human facial attribution (Waller et al., 2013). Nonetheless, neither this nor the more reciprocal facial interactions seen in apes are able to sustain the kind of sympathetic intercorporeal relationships that exist between bodies that are fully faced. ⁵

Partial faces abound in the world of robotics. Social robotics seeks to leverage the human propensity to produce partial faces in order to have us misrecognise machines as face-bearing – and thus human-like – entities. The attempt to leverage this human facial orientation in robotics is most associated with the influential concept of the ‘uncanny valley’ (Mori, 1970), according to which pursuing realistic simulations of human appearance can backfire. Add some simple features evocative of a face to a simulacrum and the human perceptual apparatus will fill in the gaps and recognise it as human-like; but our capacity to interpret faces is so nuanced and sophisticated that, the more detail you add to your simulated face, the more likely it is that our incredibly fine perceptual discrimination of facial attributes will be discomfited by a mismatch between that detail and the real thing. As a result, a simpler, more basic representation of a face can be more convincing than a more sophisticated and complex one – at least one that falls short of perfect verisimilitude. ⁶

But what about a robot visage that lies on the other side of the uncanny valley? The concept of the uncanny valley suggests that this point – when designed facial features have moved through the valley and precipitously risen on the other side – represents a final success in the creation of an artificial face; but in reality it only represents arrival at a partial face whose capacity to generate misrecognition is the most compelling possible. Current attempts to create realistic robot faces, such as the robots of Hiroshi Ishiguro, ⁷ appear nightmarish in their contorted efforts to mimic human expression, but even if their reproduction of human facial features was flawless, this would only achieve the level of the dog’s partial face, not the human being’s full face.

A non-sentient machine can never be said to possess a face in the fullest sense of the word because a face requires an exchange of signification, a sharing of affect, and a sense of shared humanity. But a sufficiently sophisticated machine, one equipped with a realistic expressive mobility and ability to register and appropriately respond to human facial expression that far exceeds current technological capabilities, might produce a flawless illusion of faciality. The kind of human engagement invited by such a machine would be fundamentally dysfunctional in its manipulation of human empathy. It could be argued that this misrecognition would be no different from that produced by a dog, which has itself evolved a capacity to play upon our perception of faces; however, even if it lacks a full face, the dog does possess a genuine capacity to experience positive and negative affect, to thrive or suffer and to develop meaningful bonds with human beings. As a result, feelings of empathy, and a desire to treat the dog with dignity and respect, do not seem misdirected whether they are founded on a kind of facial misattribution or not. In the case of a machine with a face, on the other hand, not only is there no genuine affect in the illusory face, but, more importantly, the viewer knows this to be the case, potentially causing the face to come unstuck from its moorings in feelings of empathy and shared humanity. Such a face would have no reality outside the perception of an external human observer; just as the dog can ‘smile’ as it gasps its last breaths, lying, agonised, in the gutter after being hit by a car, such a robot would continue to ‘smile’ cheerily as its head is bashed in with a hammer.

Such a partial face lacks precisely what Emmanuel Levinas presents as definitional to the face as representative of living human exteriority: its quality of always ‘exceeding the idea of the other in me’, as being something that ‘at each moment destroys and overflows the plastic image it leaves me’ (Levinas, 1988: 50–1). With such a perfect physical simulation of the face, what you see is what you get; it is never anything more or other than the impression it makes on an observer. For Levinas, the face represents the fact that others always exceed their summation in our perception, which produces a moral responsibility towards them. To create a face that does not do this is to create a face that exists only as a product of our solipsistic viewpoint on those around us, lacking any existence beyond our bestowal of faciality, a face that exists only to service our requirements and whose appeal to empathy is a facile illusion that can be rationally dismissed. In other words, such a robot would allow every one of us to see faces in the way that psychopaths already see faces now.

The Face and Violence

In fact, an influential account of psychopathy ties it to the face’s role in the development of a posited ‘violence inhibition mechanism’:

[E]thologists noted that the display of submission cues to a conspecific aggressor resulted in the termination of the attack. For example, dogs when attacked by a stronger opponent bare their throats. This results in the cessation of the fight…. [H]umans might possess a functionally analogous mechanism: a violence inhibition mechanism (VIM)…. [The] VIM [is] a cognitive mechanism which, when activated by non-verbal communications of distress (i.e., sad facial expression, the sight and sound of tears), initiates a withdrawal response; a schema will be activated predisposing the individual to withdraw from the attack. In line with this suggestion, Camras (1977) has observed that the display of distress cues (a sad facial expression) does result in the termination of aggression in 4- to 7-year-olds. She studied the use and the effect of facial expressions in children defending possessions. When a child displayed a sad facial expression when resisting another child’s attempt to take a possession, the aggressor child usually terminated his/her demands and allowed the original possessor to continue playing ‘for a relatively long time’. (Blair, 1995: 2–3)

The face is the part of the body most powerfully involved in the maintenance of empathetic connections between human beings. The face manifests, and often reflects, expressions that maintain shared affective relations, and these expressions are an evolved underpinning of human sociality and mutual care. But to what degree could interaction with a face that was not attached to a person really weaken this function? It might be argued that such a face could only produce one of two outcomes. If we lean towards a prioritising of rational agency and conscious control in our interpersonal dealings, we might believe that we would clearly differentiate simulated bodies from real, and therefore our interactions with the former would be unrelated to our dealings with the latter. On the other hand, if we accept the powerful role of non-rational, pre-conscious influences such as our perception of faces on our interpersonal dealings, we might believe that these influences would make us averse to treating realistic robots like objects or doing them harm, whatever rational understanding of their lack of personhood we might have, in which case they could never habituate us to behaviours that would be cruel when directed at real people. So we either won’t treat robots like people because we understand that they aren’t people, or we won’t treat robots like robots because we feel that they are people; either way, the boundary between human and non-human, person and non-person, remains uncompromised and our behaviour towards real people is unaffected. The problem with this argument is that the boundaries of human personhood have historically been unstable even in the absence of convincing technological simulations.

In 1511, when Fray Antonio Montesinos asked regarding the natives of Hispaniola, ‘Are these not also men?’ (Seed, 1993), many Spanish colonists obviously didn’t agree with the implied answer, their belief in the non-humanity of the indigenous people enabling them to perpetrate acts of murder, rape, mutilation, enslavement, and other forms of cruelty. The history of colonialism and slavery, war and genocide, is filled with instances of people committing acts of savage barbarism towards other human beings despite their faces’ appeal to empathy. Some of the perpetrators of these acts could be categorised as murderous psychopaths indifferent to such empathetic appeals, but most, presumably, cannot. Most must simply have been people whose victims had been dehumanised to such an extent that they were perceived as much like humanoid robots themselves.

The very existence of this history of intraspecies cruelty is likely to be cited as evidence against my argument for the significance of the face. If the face is such a powerful force in the creation of empathetic relationships between human beings, then why is it that there are seemingly endless examples of one human being looking at another and seeing only a sub-human object whose suffering evokes no sympathy? However, in the history of mass, organised, intraspecies violence it is possible to see, rather than an absence of any countervailing influence generated by faces, a need to actively neutralise the power of faces in order to facilitate the prosecution of violent acts.

The sheer weight of history makes it impossible to deny that human beings are capable of great violence and cruelty towards one another. At the same time, however, there is good evidence that there are powerful, innate forces that work to push human beings away from doing harm to their own kind. While there are clearly circumstances in which rage or panic – whether justified or not – can make a person murderously violent, the history of organised warfare is one of attempts to cause large numbers of human beings to kill one another in the absence of any direct, individualised sense of animosity or threat. This attempt to produce organised mass killing conducted in a methodical and rational way can be understood as working against a set of basic, pre-rational, evolved empathetic connections between human beings that are importantly maintained by the face.

Perhaps the best-known assertion of a basic human aversion to killing in war was made by the journalist and military historian S.L.A. Marshall when in the 1940s he claimed, on the basis of after-action interviews, that most US Second World War soldiers simply were not prepared to shoot at the enemy during military engagements (Marshall, 2000). Marshall’s claims would later be disputed, perhaps to some degree because they were believed to impugn the fighting prowess of US soldiers, but certainly for being based on data that was too limited and lacking in methodological rigour (Smoler, 1989). At the same time, however, there have been no similar studies whose results might contradict his claims, and they remain broadly supported by other kinds of scholarship, as well as anecdotally by veterans (Bourke, 1999: 73–6; Jordan, 2002). In the 1990s, psychologist and US Army lieutenant colonel Dave Grossman’s book On Killing (1995) drew on Marshall’s claims and combined them with other evidence ⁹ to argue that military training hinges on the problem of how to overcome a deep-seated human resistance to killing. Certainly, the evolution of training techniques for infantry and other groups of people expected to kill has, since the mid-20th century, pursued the aim of installing a habituated, pre-conscious capacity to kill, for example through a movement from shooting at bullseyes to shooting at responsive human-shaped targets, and then on to the use of computer simulations and realistic training environments with human actors (Bourke, 1999: 84; Protevi, 2008). If Marshall’s narrative is to be believed, the combined effect of such techniques aimed at dehumanising enemies and installing the shooting of human targets as a conditioned reflex was to lift the ‘ratio of fire’ from 25% or less in the Second World War to around 55% in the Korean War, and then to over 90% in the Vietnam War (Grossman, 1995: 35). ¹⁰

While the accuracy of these figures is open to dispute, the idea that modern military training conditions soldiers to shoot to kill despite a powerful countervailing aversion is given support by the problem of post-traumatic stress disorder (PTSD) among US veterans of the Vietnam War and later deployments. It has been known for some time that PTSD is correlated most strongly, not with a fear of being killed in war, but rather the stress and guilt that comes from killing others (Grossman, 1995: 51–66). Surveys of US veterans of the Vietnam War and more recent conflicts have established that whether a soldier has killed is the strongest predictor of PTSD (Fontana and Rosenheck, 1999; Maguen et al., 2012, 2013). This suggests that training strategies aimed at producing a near-universal willingness to shoot to kill is responsible for the scale of PTSD among US veterans of recent conflicts, a scale illustrated by the fact that, by the late phase of the occupation of Iraq, US soldiers were more likely to die from suicide than enemy attack (Maguen et al., 2012: 918).

All of this suggests that there is, indeed, a powerful innate human aversion to the killing of other human beings, at least outside extreme states of rage or panic. Armies have developed behaviour modification techniques intended to circumvent this, whose effectiveness is attested to by the fact that they can overcome this aversion even where this causes extreme and lasting psychological damage to individual soldiers. And the face plays the most important role in initiating and maintaining this aversion to violence.

This is apparent in the broadest and most obvious way from the fact that the less visible another person’s face is, the easier it seemingly becomes to do that person harm. Those who kill at a distance in war, for example by firing artillery shells or dropping bombs from aeroplanes, seem to do so more readily and with less psychological harm to themselves (Grossman, 1995: 51–66). Furthermore, when killers have the opportunity to arrange the scene of their killing, they have a tendency to ensure that the faces of those being killed are not visible, for example by shooting them in the back of the head, or, in the case of the firing squad or hanging, having them wear a blindfold or hood (Grossman, 1995: 128). The meeting of selves is effected by looking at one another, a reciprocal act of mutual perception, and a line can be drawn between infant play and the firing squad, in which the logic of peek-a-boo is employed in the blindfold that renders the condemned man’s personhood invisible. Modern military imaging technologies such as night vision and infrared scopes, or camera feeds from drones, which reduce bodies to dark silhouettes, all analogously function to erase the faces of human beings in contexts where they are about to be killed.

Dave Grossman argues that, in ancient battles, most of the killing that took place occurred, not during the clash of massed soldiers, but rather after one side’s formation had broken, when fleeing enemies could be chased down and killed from behind (1995: 12–13). Going back further still, biologist Paul M. Bingham has argued that the evolution of a human specialisation for throwing missiles was driven not simply by its usefulness in repelling predators, but rather primarily as a means of doing harm to other human beings at a distance, presumably without having to look them in the eye while doing so (Bingham, 1999). According to anthropologist Paul Roscoe, all of these factors result from a process of:

devising psychological and cultural ‘technologies’ to overcome the limitations imposed by the genetic emotions and dispositions to which humans are heir…[creating] a set of techniques that side-step or short-circuit humanity’s aversion to this act with results that have been as consequential for human lethal violence as any projectile weapon or suit of armor. (Roscoe, 2007: 488–9)

Counterfeit Human Beings

In a piece on robotics research written for the Washington Post, ¹¹ journalist Joel Garreau recounts the story of a US Army colonel who – despite being a man whose duties involve making dispassionate decisions to put human beings in situations where they might be killed or maimed – cancelled a demonstration of a new mine-clearing robot because he ‘just could not stand the pathos of watching the burned, scarred and crippled machine drag itself forward on its last leg’ (Garreau, 2007). This is, of course, yet another instance of a human observer anthropomorphising a robot (in this case one that was not even modelled on a human body). But what is the half-life of such a sympathetic response to a simulated body? If that colonel watched the robot being crippled by mines over and over, wouldn’t he at some stage become hardened to the spectacle? And why wouldn’t he watch this spectacle over and over? After all, the purpose of the robot is to save human lives, and surely saving human lives is more important than the emotional discomfort of seeing a machine drag itself through a minefield. If the robot was considered effective in its life-saving work, any irrational squeamishness could be pushed aside by a rational understanding that sympathy is misplaced and unhelpful. In all likelihood, the colonel would come to look back on his initial, naïve response with embarrassment. Over time, surely he could inure himself to the spectacle of the robot, training himself to ignore the inappropriate emotions it generated, and doing so would perhaps become easier on each subsequent viewing.

As already noted, there has been a development towards ever-greater realism in military training with the aim of increasing soldiers’ willingness to shoot at human beings in real combat situations. Grossman even mentions a sniper trainer making human figures with heads made from bags filled with chopped cabbage and ketchup so as to simulate the bursting of a human skull when hit with a bullet (1995: 254–5). An obvious future step would be the addition of extremely life-like robots, ones that were as close to indistinguishable from a real living person as possible. The first time a trainee soldier shot such a robot, it would presumably be a confronting, upsetting experience: the simulation of life would be effective enough that, at a visceral level, it would seem like a real death. But the trainee could tell him- or herself that these feelings of distress or revulsion were completely unwarranted: this was not a real person; the blood pouring from the wound was fake and the screams were just the recorded voice of an actor; no-one had suffered and no-one had died. Presumably, with each exercise, it would become easier to shoot the robot, knowing that the gut reaction was an error caused by an unjustified association of the robot with a living, sentient being. Later, if that soldier was ever required to shoot a real person, perhaps the basic empathetic connection stirred by the sight of another human body would have been dulled; the soldier would have been habituated to neutralising this response on the grounds that it was irrelevant or misplaced, like a phobic whose fear of spiders or heights has been dulled by a programme of controlled exposure to arachnids or steep drops.

A realistic humanoid robot would be a logical next step in military training after realistic computer simulations. But wouldn’t it therefore be only an incremental development in the realistic simulation of violence, not just in military training but popular entertainment? Even Dave Grossman’s On Killing concludes with an indictment, not of military training that uses behaviour modification to enable killing, but rather of entertainment media, whose depictions of violence he argues have the same effect. However, despite his conclusions, Grossman’s account of military training highlights its important difference from media representations of violence. At least in cases where killing is associated with PTSD, military training obviously hasn’t been able to desensitise soldiers to violence in the manner attributed to media representations at all; on the contrary, psychological damage is caused by intense feelings of guilt and horror. Such training does not aim to modify soldiers’ rational, conscious evaluation or understanding of the act of killing; rather it seeks to override a non-rational, non-conscious, instinctive resistance through the cultivation of deeply ingrained habit. The soldier sees a human figure and pulls the trigger, just like in training; it is only afterwards that the soldier reflects on the larger significance of this action and feels revulsion and remorse. Such training is not about changing what the soldier thinks, but rather how the soldier acts.

No existing representational technology is able to convincingly simulate full, reciprocal faciality of the kind that underpins human intersubjective relationships and the empathetic connections that sustain them. While a sophisticated robot would not achieve true, full faciality either, a convincing simulation would create a fundamental conflict between how bodies make us feel and how we act towards them. And the dangers of machines with faces would extend beyond their capacity to escalate the realism of military training (which in any event is complicated by the fact that an increased capacity to cause violent behaviour is the aim of this training rather than an unwanted side effect). The larger danger would come from the introduction of simulated human bodies into everyday contexts of entertainment or service as a new kind of consumer object. In this context it seems likely that they would be subjected less to an acting out of murderous assault than ongoing routines of casual, everyday mistreatment. While it seems inevitable that some people would act out elaborate cruelty on machines with faces in the service of fantasies of sadistic violence and sexual assault, most people would not want to act in such a way. They would not treat such artificial bodies as victims – instead they would treat them as objects or appliances: shoving them aside; or cursing or kicking them when they got in the way or did not perform as required; looking into their faces with indifference as they were damaged or destroyed.

If even a faceless insect-robot could inspire sympathy in a hard-bitten US Army colonel, civilian robot-owners could be expected initially also to treat these new machines with an approximation of human consideration. However, even if this is true, presumably such feelings would quickly fade, until the affective invitation of a humanoid robot’s illusory face would ultimately seem like nothing more than a cheap trick. During an experiment that placed a robot in a shopping mall to gauge human responses to its polite requests for passage through the crowd, some Japanese children spontaneously formed groups that shouted insults at the robot and hit it with bottles (Brscić et al., 2015; see also Salvini et al., 2010), and a more general tendency of robots to provoke violently hostile responses is increasingly recognised as a problem by roboticists (Bromwich, 2019). The cause of this victimisation is perhaps the fact that we know the robot’s human-like behaviour to be a fraud; when it invokes the kind of consideration due to a real person, we aggressively act out its exclusion from the sphere of human personhood.

According to David Livingston Smith, the process of dehumanising actual human beings is dependent on a belief in the existence of ‘counterfeit human beings – creatures that look like humans, but who are not endowed with a human essence’ (2011: 101), and a convincingly realistic android would confirm such a belief. Even if we understand others’ right to dignity to be guaranteed by some inner self, we have no access to this inner self; rather we believe in some necessary connection between it and facial liveliness. To believe in the existence of counterfeit faces would thus necessarily entail a belief in the possibility of counterfeit human beings.

Of course we would still possess a rational, conscious belief that the human face was attached to a sentient subject and the robot face was not; but the intersubjectivity sustained by the face exists outside the realm of rational, conscious understanding. It begins with the innate predispositions of human infants, but then develops through the engagement with faces these predispositions initiate. How differently would this process play out if a child grew up interacting with domestic helpers or teachers whose faces she knew to be disconnected from any claim to genuine subjectivity?

Conclusion

‘[I]f I look out of the window and see men crossing the square, as I just happen to have done, I normally say that I see the men themselves’, wrote René Descartes in his Meditations on First Philosophy. ‘Yet do I see any more than hats and coats which could conceal automatons? I judge that they are men. And so something which I thought I was seeing with my eyes is in fact grasped solely by the faculty of judgement which is in my mind’ (Descartes, 1988: 85). While Descartes recognised that the attribution of personhood originates with the external observer rather than being generated directly by any essential quality of the entity being observed, he did not appreciate the degree to which our recognition of shared humanity is not the product of dispassionate conscious interpretation or conjecture, but rather a non-conscious response to dynamic physical cues. But what if even these indicators of a shared humanity were not enough to convince us that we were dealing with anything more than an automaton dressed in a hat and coat? Would this devalue or throw into doubt the deep-seated, pre-conscious bonds that tug our interpersonal interactions in the direction of mutual respect?

It is possible that machines with faces could never pose a genuine risk. It may simply be that such machines will never be built; certainly, the level of technical sophistication required to build such a machine exceeds that which is currently available. Large amounts of time and money have already been dedicated to creating believable physical simulations of human beings, and yet the results to date are far from compelling. It may also be the case that, even if a robot was created that initially seemed indistinguishable from a living human body, with familiarity the illusion would fade. The clockwork automata of the French Enlightenment were initially hailed as astounding reproductions of life, fascinating to both the public and scientists seeking new insights into the human body; but familiarity increased awareness of the limitations of the illusion, such that what was once seen as providing astonishing insights into the workings of life itself came to be seen as suitable only for novelties and children’s toys (Wise, 2007). Prolonged exposure to illusory faces is likely to make us literate in the rules of the simulation, and aware of the limitations and repetitions in their performance, such that they become progressively less convincing over time. Those who initially attributed faces to machines might later be regarded in a way analogous to the early audience for the Lumière brothers’ cinématographe who supposedly fled from a Parisian cafe in the belief that they were about to be crushed by a train. If so, the preceding discussion has value only as a thought experiment that allows us to explore the complexity and power of real human faces.

But roboticists continue in their efforts to create convincing simulations of the face, and perhaps one day one of them will succeed. If such a machine was to be created, its likely applications would include attempts to increase the human capacity to kill through military training; but even where this was not the aim it still could be expected to erode the power of the human face in such a way that human empathy and intersubjective engagement is weakened. In her book The Nazi Conscience, Claudia Koonz notes that a maxim popular with Hitler and other leading Nazis when defending genocide was, ‘Not every being with a human face is human’ (Koonz, 2003: 2), and the creation of machines with faces would establish the truth of this statement. The face is the most uniquely human feature of our bodies, which anchors a sense of shared personhood and provides external authority to our claims to empathy and mutual care. Compromising this uniqueness and its exclusive relationship with personhood threatens to undermine the humanity, not only of the other human beings with whom we interact, but also of our own actions.