Anthropomorphizing Machines: Reality or Popular Myth?

Abstract

According to a widespread view, people often anthropomorphize machines such as certain robots and computer and AI systems by erroneously attributing mental states to them. On this view, people almost irresistibly believe, even if only subconsciously, that machines with certain human-like features really have phenomenal or subjective experiences like sadness, happiness, desire, pain, joy, and distress, even though they lack such feelings. This paper questions this view by critiquing common arguments used to support it and by suggesting an alternative explanation. Even if people’s behavior and language regarding human-like machines suggests they believe those machines really have mental states, it is possible that they do not believe that at all. The paper also briefly discusses potential implications of regarding such anthropomorphism as a popular myth. The exercise illuminates the difficult concept of anthropomorphism, helping to clarify possible human relations with or toward machines that increasingly resemble humans and animals.

1 Introduction

The idea that people anthropomorphize machines continues to be a subject of much interest both within and beyond robot and computing sciences (Airenti, 2018; Damholdt et al., 2023; Darling, 2021; Duffy, 2003; Li & Suh, 2022; Salles et al., 2020).¹ A particularly intriguing claim about anthropomorphism in many discussions is that humans commonly and instinctively attribute or ascribe mental states such as feelings of fear, desire, happiness, love, pain, joy, anger, and sadness to certain machines such as robots.

To anthropomorphize, this idea goes, is to erroneously believe that certain machines, usually machines that are human-like or animal-like in some way, really have such feelings or phenomenally experienced mental states, which of course they lack. For example, Matthias Scheutz tells of a colonel who condemned as “inhumane” a military test that blew the legs off a mine-detection robot. Scheutz (2012, p.210) claims that

the only explanation for not wanting to watch a mindless, lifeless machine…destroy itself, is that the human projected some agency onto the robot, ascribing to it some inner life, and possibly even feelings.

Amanda and Noel Sharkey (2012, p.311) discuss at length the ostensibly widespread “illusion that a robot is sentient, emotional, and caring or that it…loves you.” And Kate Darling (2016), who has written extensively about anthropomorphism of both machines and animals, claims that

not only are we prone to ascribing more agency [to robots] than is actually present, we also project intent and sentiments (such as joy, pain, or confusion) onto [robots].

Such claims are now common. Many writers discuss how readily, if unintentionally, certain robots can trick, fool, deceive, mislead, or delude human beings (Coeckelbergh, 2012; Nyholm & Frank, 2019; Sharkey & Sharkey, 2021; Sparrow & Sparrow, 2006). The idea here is that we are tempted and gulled into thinking that these machines do not just appear to have emotional states but that they actually have emotional states. For if it was held that people do not believe that machines actually have emotional states, it would be quite misleading to speak in this context of people being tricked or deceived, or of people being under an illusion about robot sentiments or love, or of people ascribing to robots inner qualities that are not ‘actually present’.

To be sure, ‘anthropomorphism’ need not always refer to this sort of false belief in machine feelings or mental states. For example, the term can refer to the belief or assumption that a machine has abilities not involving phenomenal experience, such as forms of machine autonomy, agency, or intelligence, that are not manifest in the immediate interaction. Alternatively, it might refer to positive attitudes such as affection towards ‘friendly’ robots. Or again, ‘anthropomorphism’ may merely signify machine features that resemble eyes, mouths, or hands or that involve human-like voice, text, or movement (Coeckelbergh, 2021).

But there is a more philosophically interesting sense of anthropomorphism,² pointed to in the above quotes. It relates to the claim that human beings commonly but mistakenly believe that some machines really or truly have phenomenally experienced feelings.³ This is what some scholars mean by ‘anthropomorphism’. I will call the claim that people tend to form the above false belief the ‘A-claim’. The A-claim often centers on robots with human or animal-like bodies and physical behaviors. However, it could also refer to ‘disembodied’ AI, such as recent chatbots like ChatGPT, Bard, and Gemini that competently ‘converse’ with us.

Though widespread, the remarkable claim that humans are prone to forming false beliefs about non-existent feelings in some machines has not been sufficiently critically examined. However plausible the claim initially appears, there are reasons for suspecting it might be a popular myth. While I will not argue for that position definitively, I will suggest the claim needs stronger justification than it has received.

The paper runs as follows. Section 2 provides background to type of anthropomorphism under consideration, Sect. 3 clarifies the A-claim, and Sect. 4 outlines alleged mechanisms of anthropomorphizing beliefs. Section 5 then raises problems for the A-claim. Section 6 considers an objection to the argument and Sect. 7 some implications. Section 8 concludes.

2 Background to the A-Claim

Some commentators make the A-claim explicitly. Others imply it. Still others use the term ‘anthropomorphism’ without saying whether it involves a cognitive or descriptive error about machine nature and feelings. Used without qualification, the term often conveys an impression that our stance towards machines is in error, for ‘anthropomorphism’ is historically linked to purportedly specious attributions of mind-features to religious entities, inanimate natural features, and nonhuman animals (Kennedy, 1992). Consequently, invoking anthropomorphism often tacitly raises the A-claim.

Sharkey and Sharkey (2021, p. 313) note that humans are anthropomorphic about “any inanimate objects with certain features.” Nonetheless, machines like robots are often significantly more human-like than other things (Sica & Sætra, 2023). Moreover, the intriguing human responses they elicit can seem to support the A-claim.

The responses that supposedly reveal a belief that machines really do have mental states have two forms. One is explicit assertions involving mentalistic language that people might make (e.g., ‘the robot is happy/sad/suffering’). The other (more discussed) indication relates to certain human behaviors towards machines like robots and chatbots. Like mentalistic assertions about machines, certain behaviors also allegedly suggest or reveal the presence of the erroneous belief. Here are just a few of many possible examples.

Several decades back, Joseph Weizenbaum’s (1966) simple psychotherapy chatbot ELIZA prompted ordinary people to disclose their mental health problems to it. Weizenbaum became deeply concerned that even very basic machines can fool people about their true (mindless and non-feeling) nature. Today the ELIZA-effect refers to “the tendency to project human traits—such as experience, semantic comprehension or empathy—into computer programs.”⁴

More recently, YouTube viewers have felt sorry for robots that stagger when forcefully kicked.⁵ Some have refused to ‘torture’ a robot that yowls when held upside down (Darling, 2016). Studies show people being reflexively polite to computers and affectionately naming their robot vacuums (Sung et al., 2007). Research also shows people can be inclined to be abusive or violent towards smart speakers and robots (Bartneck & Hu, 2008).

Some people remove their robot “from the room while changing” to avoid being ‘watched’ or have “feelings of guilt when putting the device back in its box” (Darling, 2021, p. 8). There are reports of individuals ‘falling in love’ with virtual human personas such as Replika⁶ and becoming distraught when the personas are downgraded.

Now, such examples certainly show that people tend to respond—i.e., speak about and behave—towards certain machines in ways somewhat like their responses to fellow humans. These emotionally engaged responses are then presented as evidence for the A-claim. Indeed, as future machines get more human-like in capability and appearance (Orlando 2023), some think these startling human behaviors and their associated (alleged) false beliefs may intensify.

One reason such anthropomorphism is considered important concerns its possible practical implications. In fact, the above kinds of responses to machines have been viewed both positively and negatively (Boden et al., 2017). For example, a worry exists that lonely or vulnerable individuals will be fooled into treating human-like machines as romantic partners or friends (Scheutz, 2012) and that human-like robots should perhaps be prohibited in sensitive care settings. Bowles (2018, p. 185) argues that “humanoid robots…should follow a principle of ‘honest anthropomorphism’ and break through the fourth wall, reminding us of their non-human nature.”

Others (e.g., Duffy, 2003), however, treat anthropomorphism as a relatively harmless illusion or even as a productive fantasy we might voluntarily engage in. For example, the illusion and false belief of feelings, some suggest, might actually facilitate desirable co-operation with machine assistants and benefit people in the process (Złotowski et al., 2015).

Fully assessing the A-claim is both an empirical and philosophical task. It is partly empirical because identifying the alleged false beliefs requires attention to evidence about people’s language and behavior toward machines. But the assessment is also partly philosophical: it requires trying to understand and interpret that behavior and language and any associated beliefs. The philosophical task is my focus.

Undoubtedly, studies in Human Computer and Robot Interaction and psychology have revealed intriguing and strange responses to machines. Furthermore, one can appreciate how they have tempted the A-claim. Nonetheless, that particular interpretation of those responses is not beyond criticism. Before raising some problems for it, we should clarify the type of anthropomorphism under consideration.

3 Clarifying the A-Claim

The A-claim runs:

People commonly but mistakenly believe that some machines really do have mental states like feelings.

‘Feelings’ is linked to concepts like phenomenal states, inner life, subjective experience, and sentience. Humans and many animals have phenomenal mental states, whereas rocks, chairs, and dolls totally lack any inner life or sentience. Most people believe this.

The term ‘really’ refers to the belief that a machine actually, literally, or truly has mental states like feelings. As we noted, that is why A-claim proponents speak of false belief, deception, trickery, and illusion. People can, of course, use mentalistic language to describe things without believing they are really sentient. One example of this is speaking wholly metaphorically or purely comically about an artifact’s ‘feelings’ (‘This damn dishwasher hates me’).⁷ Such modes of speech clearly do not involve the false belief or mistaken ascription.

However, on the A-claim, people (allegedly) do indeed believe that certain machines have feelings just as, say, dolphins or humans have feelings and as rocks and dolls do not. Of course, this view does not imply that people believe that the mental states of certain machines, dolphins, and humans are identical in nature. People may well believe there are important differences between the mental states in these different entities or beings (e.g., that robot joy is somehow different from dolphin joy which is somehow different from human joy). But that qualification is compatible with believing that they are all equally mental states or phenomenal experiences in these different entities or beings.

A-claim proponents do not necessarily hold that the false attribution or belief about machines need be particularly self-aware, reflective, or persistent. Proponents may well acknowledge the relative rarity of cases like Blake Lemoine, the Google engineer who passionately argued that Google’s chatbot LaMDA was sentient.⁸ But while most people do not believe it “intellectually,” as Darling (2021, p. 111) puts it, they might nonetheless believe “on a more intuitive and subconscious level…that certain robots have much more of an inner life…than they really have”. For Darling (2016) and others, “the line between lifelike and alive” gets “muddled in our subconscious.”

This notion of subconscious belief regarding robots is common enough, although it seems underexplained. Perhaps the notion simply refers to false beliefs that occur transiently and with minimal reflective awareness during human-machine interactions. Still, the idea of genuinely subconscious or unconscious belief makes sense in various contexts. For example, we might readily imagine a man who is terse with his mother-in-law because of an unconscious belief that she resents him, even though he sincerely denies having that belief.

Likewise, someone might deny ever believing that the robot they treated politely had actual feelings, but that polite treatment might be interpreted as evidence that they did believe it at some subtle or lower level after all. Even when the belief is ‘subconscious’, it is still, on this view, a belief that the relevant machine really does or did have feelings. Thus, according to A-claim proponents, it is not the case that people with the subconscious belief only appear to believe; rather, they do indeed believe (subconsciously or unconsciously) that those machines actually have feelings.

For many commentators, the A-claim applies to technological experts and adults as much as to the technologically naïve (e.g., Sharkey & Sharkey, 2021) and to very young humans (Jipson & Gelman, 2007). A-claim proponents do not have to claim that people form false beliefs about all robots and AI. For example, they may think that some vacuuming and industrial robots do not necessarily trigger the erroneous anthropomorphic beliefs, whereas robots with faces, certain patterns of autonomous behavior, or language ability do often trigger them.⁹

Nonetheless, a common theme in these discussions is that such anthropomorphizing occurs very easily and naturally, and for some even automatically, for a variety of machines; and for many A-claim proponents, this can apply to starkly mechanical-looking machines with only rudimentary human-likeness. In any case, I will be questioning the general idea that people often believe that any current (and perhaps future) robot or AI system really has phenomenal mental states.¹⁰

4 Supporting Arguments for the A-Claim

Often the A-claim is merely assumed or asserted with minimal argument. Yet the A-claim does need support: we cannot just accept it at face value. For not only is it a surprising claim, but, as its proponents themselves agree, people to whom the false belief is supposed to apply may deny that those machines really have feelings or mental states. Indeed, some people to whom the belief is attributed may deny that they ever believed such a thing, even transiently. While such denials are not, as we noted, definitive, they do demand some argument showing that any initial interpretation of people’s behavior as supporting the A-claim is not mistaken or has not been accepted too quickly and uncritically.

Furthermore, while observations and reports of people’s behavior and language regarding machines is arguably suggestive, it is compatible with those people not believing the machines really have mental experiences. It is possible that people may sometimes appear to act and speak as though they believed it, or in ways that prompt the A-claim, without actually believing it. Even if people use mentalistic language, such as when they say ‘the robot was suffering’, they might not believe the robot really was suffering. We will examine this possibility before too long.

Some proponents have recognized the need to support the A-claim. Proponents often start by observing that a general tendency to anthropomorphize is ancient and runs deep in human nature. Humans strongly tend to perceive or describe by means of mentalistic terms many different things, even things that are only vaguely human-like. A-claim proponents may say that these things, whether they are substantially or only distantly life-like, push our Darwinian buttons, naturally eliciting the attribution of human qualities to nonhuman objects.

Consider the humanizing and personification of some dwellings or boats, or the way that tree, cloud, and rock formations can appear to have angry or happy forms. In addition, proponents may observe, even comparatively primitive machines may be anthropomorphized. For example, the 1990’s social robot Kismet prompted observers’ recognition of emotions despite its mechanical facial features and its very crude expressions of surprise, fear, sadness, and happiness (Breazeal, 2003).

Some proponents highlight cultural influences from sci-fi robots and AI as drivers of the anthropomorphism phenomenon. Epley et al. (2007) cite human desires for social connection. Others point to psychological causes like higher trait empathy which, for instance, makes people more hesitant to strike robots (Darling et al., 2015).

However, while these factors may help explain various intriguing human responses to machines, they are not in themselves sufficient to show that people believe that machines really have mental experiences or feelings. Consequently, some proponents have presented additional arguments about the nature of that ostensible anthropomorphic belief and its underlying mechanisms.

A prominent argument of this kind is that people automatically infer that machines have phenomenal mental states. For example, Epley et al. (2007) claim that we infer the existence of mental experiences in nonhuman entities by drawing on knowledge of human minds. Nonhuman things with human-like features provoke a (perhaps instantaneous) inductive move to a belief that those entities have mental experiences.

Perhaps even more common is the view that we naturally project human mental qualities onto nonhuman entities (Damiano & Dumouchel, 2018; Darling, 2021; Zawieska et al., 2012). Some think this involves employing heuristic human mental models (Bruni et al., 2018) to make sense of certain nonhuman phenomena (Kiesler, 2005). Howsoever it occurs, the idea is that human-resembling features can trigger projection.

To further support claims about inference and projection and how these belief processes work with machines, A-claim proponents have offered an analogy based on animals. According to this analogy, people make inferences or projections regarding machine feelings in the same way they make inferences or projections regarding animal feelings. Indeed, nonhuman animals supposedly provide a perfect model of how the processes of inference and projection operate in our responses to AI and robots (Darling, 2021).

While the alleged mechanisms of inference and projection involve an automatic or instinctive transfer or ascription of mental states beyond human beings to nonhuman things, a different and less passive belief mechanism has also been also proposed: willing suspension of disbelief (Duffy & Zawieska, 2012; Prescott, 2017). This alternative argument for the A-claim presents humans as voluntary self-deceivers. In deliberately suspending our disbelief, this idea goes, we actively misperceive machine mental states.

While some have doubted the A-claim or similar ideas (Coghlan et al., 2019; Damiano & Dumouchel, 2018; Rodogno, 2016; Watson, 2019), the A-claim remains compelling for many scholars and has even attracted media attention.¹¹ Questioning this alleged kind of anthropomorphism may be attempted by more closely examining proponents’ supporting arguments about the nature or mechanisms of the belief.

5 Questioning the A-Claim

In this section I raise some problems for the A-claim. I first address the argument of willingly suspending disbelief, and subsequently address the arguments about inference and projection. The skeptical approach taken below frequently involves comparing machines with other entities, including movie characters and puppets, and our responses to and beliefs about these sorts of things. When taken together, the following points cast some doubt on the A-claim, or at least demonstrate it needs a better defense. As we go and afterwards, I consider some objections to the critical points being made.

5.1 Willing Suspension of Disbelief?

Consider first the claim that our impressive ‘anthropomorphizing’ responses to machines can be explained by a willing suspension of disbelief (Duffy & Zawieska, 2012). Here we are talking about a kind of belief suspension in which we voluntarily fool or actively delude ourselves that machines really have mental states.

The idea of a willing suspension of disbelief seems somewhat unclear and underdeveloped in this context, but we might make some sense of it. Imagine that a child, playing with an expressionless and immobile doll, declares, “Look! Charlie is upset because I didn’t take her to the playground,” and then cuddles the doll to soothe it. In this scenario the child speaks about and behaves towards the doll in ways that somewhat resemble speech and behavior the child may have towards upset humans or animals.

However, it is quite a stretch to claim that the child believes the doll is really upset. A more plausible interpretation is that the child is not deluded or fooled by the doll but is instead employing a vivid imagination towards it, playfully ignoring the fact that the doll does not look at all upset. Furthermore, even if we allow that some very young children might have false beliefs about doll feelings in such circumstances, it is less plausible that older children and adults have them. So, if willingly suspending disbelief is the model for understanding responses to machines, it would appear not to explain the putative false anthropomorphic belief.

Yet this model is surely not a good general analogy for the idea of self-deceiving behavior with machines in the first place. For the kind of behavior we should perhaps have front of mind here is not efforts of vivid imagination and active “pretense” (Melson et al., 2009) towards expressionless dolls or teddies (cf. Sweeney, 2021, pp. 468–469), but rather the more automatic responses people have to nonhuman things, such as their responses to robots that make life-like crying and writhing motions when they are beaten.

In such cases, people may, for instance, effortlessly feel sorry for the robot, refuse to beat it, and describe it as upset. The case here thus looks quite different to the case of the expressionless doll. No more imaginative effort is needed for (say) the crying and writhing robot than is needed to for people to recognize the emotional features of, for instance, some expressive cartoons, puppets, or statues.¹²

To be sure, some machines are less ‘expressive’.¹³ For example, some people may declare that their Roomba is upset merely when it gets stuck under furniture (Sung et al., 2007) or that it is sad when it runs out of battery charge and can’t complete the vacuuming. In that sort of case, the notion of actively suspending disbelief seems to get more grip. But for many cases, including many of those advanced by A-claim proponents, it seems less apt to describe human responses to machines as involving a ‘willing suspension of disbelief’.

To be clear, the explanation of willingly suspending disbelief perhaps has some marginal application for machines and our responses to them. But it does seem that the alternative arguments for the A-claim, namely arguments about some mechanism of inference or projection of mental states, might be more promising justifications of that claim. I address this more pertinent line of argument in following subsections.

5.2 The Problem of Simple Examples

A-claim proponents often argue that the process of inference or projection involves the same basic mechanism in the respective cases of machines and other nonhuman things. As we have seen, proponents often emphasize how little an object needs to be human-like to provoke the false belief (Scheutz, 2012). Thus, the mechanism of belief is taken to be the same whether the object is only vaguely human-like, such as with a crude metallic robot, or vividly human-like, such as with an intelligent animal or a soft-skinned talking robot. But this pointing to a common belief mechanism across such cases creates a challenge for the A-claim.

An extremely basic example of the apparent anthropomorphic belief comes from a famous 1940s psychological study of perception. Fritz Heider and Marianne Simmel (1944) produced a short film¹⁴ of a circle and triangles that moved on a screen and interacted highly schematically. The experimenters deliberately chose plain geometric figures to avoid as far as possible the human appearances used in previous psychological studies of perception. Nonetheless, most participants instinctively and naturally described the simple shapes as angry, jealous, afraid, and excited when the shapes ‘behaved’ and interacted in certain ways.

Compared to many robots and AI, Heider and Simmel’s animated shapes are not very human-like. Indeed, they are composed merely of a few black pixels, their expressive characteristics confined to the ways they move and interact with each other and with lines on the screen. But now recall that many A-claim proponents invoke the same epistemic processes of inference or projection in all cases in which people naturally exhibit certain suggestive behavior or automatically employ the relevant mentalistic language or descriptions.¹⁵

The problem is that claiming that people believe Heider-Simmel shapes really have mental experiences is highly implausible. Indeed, many of us would surely find it more plausible to say that any ‘attribution’ of (say) fear or happiness to those animated geometric figures is done in a way that does not involve believing that they really have mental states. That is, we may think that people might behave and talk in ways that resemble behavior and talk that goes with believing it, without believing it. At most, we may feel, such people only appear to believe it. Indeed, Heider and Simmel’s (1944) article itself makes no claim that the participants, despite automatically and naturally employing psychological terms to describe the “figural units,” had any belief whatever that those figures really felt fear or happiness.

Examples like the Heider-Simmel shapes are therefore instructive. Certainly, people may be more strongly and variously engaged by machines with more pronounced human resemblances. For example, they may bond with and refuse to harm them, act politely toward them, trust them, and so on (Sica & Sætra, 2023). Clearly people do not do such things with animated triangles. But if the nature or processes of belief are identical in both cases, and if it is true that people generally do not believe things like Heider-Simmel shapes really feel fear or happiness even when they apply these mentalistic terms to them, then the A-claim looks less certain.

A-claim proponents may give one of two replies. First, they may insist that, regardless of what the authors of the study might say or think, the participants did in fact believe, perhaps by inferring or projecting feelings, that the Heider-Simmel shapes really had feelings, even if those participants deny that they previously had or currently have the belief that those geometric shapes really had/have feelings (both of which denials, we may think, are likely).

Second, proponents may argue that while people do not have mistaken beliefs about very basic things like Heider-Simmel shapes, people do have mistaken beliefs about certain human-like machines (Levillain & Zibetti, 2017). Proponents may thus confine the A-claim to much more human-appearing and human-behaving robots and AI. This brings us to a key supporting argument for the alleged belief mechanisms of inference and projection: the analogy with animals and our responses to them.

5.3 Questioning the Animal Analogy

A-claim proponents have used animals as a model par excellence for the alleged anthropomorphic belief process. Darling (2021, p. 112) defends this view at length, writing that “our relationships with animals are driven by the same impulse that forges human-robot relationships – our tendency to anthropomorphize.” On this view, understanding how the belief process about mental states works with animals reveals how it works with machines. Because the animal analogy plays a significant role in some justifications of the A-claim, it deserves our attention. If that analogy is problematic, perhaps alternative supporting arguments should be sought.

Darling and some others believe we project human mental states onto animals. Such ‘projecting’ is typically not explained in detail, but the idea seems to be that this ‘projecting’ onto other entities is something people do without sufficient epistemic warrant. After all, these A-claim proponents do not say that we project or superimpose mental states onto other humans and their behavior. Rather, it seems to be assumed that, at least where humans are concerned, we justifiably believe or know they have mental states, perhaps rendering any need for projection redundant.

The idea, then, seems to be that when we see fellow humans groan, cry, or smile in certain circumstances, we have no need to ‘project’ mental qualities of pain, grief, and contentment onto them—at least, not normally. We simply recognize those mental experiences without superimposing them on such behaviors.¹⁶ It is this contrast between our beliefs about humans and our beliefs about nonhuman entities like animals that apparently underlies this mechanism of ‘projection’. This general story would also seemingly apply to the proposed mechanism of ‘inference’.

A-claim proponents (or some of them) apparently believe that because other animals are different from us, we have or experience significantly less certainty about their minds. They may think the projecting of such psychological characteristics onto animals needs further backing, such as a scientific justification, to be trusted. In any case, we can say that the contention is that the process of projecting or superimposing mental states onto animal behaviors illustrates and explains the process of anthropomorphizing machines.

However, the contention (which, to be fair, is sometimes more of an assumption) that we do such ‘projecting’ onto animals is eminently contestable. It is perhaps plausible to think that people ‘project’ feelings onto animals occasionally, as when a person declares that their dog is ‘guilt-ridden’ after being caught red-handed eating the birthday cake (Horowitz, 2009). But it is far from clear that people ordinarily or systematically project mental states like pain or pleasure onto animals, or that people infer these experiences from perceived similarities with human behaviour (Morris, 2017). This suggestion needs explanation.

In providing one possible but important explanation, I will use a helpful analogy from Rob Sparrow in an essay on robot dogs (2002, p. 311). It concerns our understanding of virtues. Consider loyalty. One might assume that we recognize loyalty in animals like dogs by inferring or projecting human behaviors or states onto them using our knowledge of human behavior. Here we would be transferring to canines concepts and an understanding that derive from human not animal life.

However, Sparrow questions the often little-scrutinized supposition that we transfer concepts or understandings of virtues from humans to animals. He argues that we do not first have an understanding of (say) loyalty informed by human behavior, and only later apply that originally human-exclusive understanding to dogs. Rather, we have a concept or an understanding of loyalty that is already partly informed by the behavior and character traits of animals like dogs.

In fact, animals give us some “paradigm cases” of loyalty (Sparrow, 2002, p. 311). Think of dogs devoted to specific humans or popular stories of dogs who starve waiting for their absent humans to return. Hence, Sparrow (2002, p. 311) says, it “makes as much sense to compare the character of humans to that of animals as to judge the animal’s behavior against the human model.” Such character traits are by no means strictly human-exclusive at their origins: animal behavior is already a part of our concepts and our understanding with respect to those traits.

We can apply this argument to the idea of inferring or projecting phenomenal mental states vis-à-vis animals as follows. As with virtues, our concepts and understanding of mental states are not informed originally and exclusively by our life with other humans. Rather, they are also partly informed by human experience with animals. It is therefore wrong to think that our concepts of pain, pleasure, fear, and joy are first created in respect only of humans and later applied to animals. They are rather created in respect of humans and animals together (Cockburn, 1990; Gaita, 2016; Winch, 1980).¹⁷

If this is right, then the idea that we transfer by inference or projection mental states to nonhuman animals based on their similarity with human beings is misguided. We no more routinely transfer various mental states belonging to humans onto animals than we transfer mental states belonging to animals onto humans. The point here is not that it is impossible to make projections or inferences about animal mental states, but that this is quite possibly not what people ordinarily do. The upshot is that the idea we naturally transfer human mental states to certain machines may be based on a fallacious analogy with animals.

These considerations should at least affect our confidence in the animal analogy as a supporting argument for the A-claim. While philosophically controversial, they cannot be blithely dismissed. For even these considerations do not convince, they show that the animal analogy argument and its notions of projection and inference needs a better defense than it has received. To date, little attempt has been made to justify it.

At this point, however, A-claim proponents could abandon the animal analogy, while retaining the contentions about projection or inference with machines. Indeed, taking their cue from our questioning of common assumptions about the nature of belief regarding animal minds, proponents might instead embrace the idea of a relevant difference in our responses to animals and machines.

A-claim proponents could now argue, for example, that while our understanding of animals is indeed well established from millennia of human experience and language use, our language and behavior toward human-like machines are still being formed. Accordingly, even if people do not generally project or infer mental states vis-à-vis animals, that is exactly what they do with human-like and animal-like machines.

It is time, however, to consider an alternative understanding of our language and behavior regarding such machines. If this alternative understanding is plausible, the claim that people often believe such machines really have mental experiences will seem still more uncertain.

5.4 An Alternative Analogy and Explanation

Our behavior and language regarding machines might seem to indicate that people believe at some level and in particular circumstances that certain machines really have mental states like feelings. But there may be a better analogy than the animal model for understanding and explaining these responses to robots and AI.

This alternative model may suggest that people use mentalistic language to describe machines without believing those machines really have mental states. Similarly, a more apt analogy may show that people sometimes behave towards insentient machines in ways that (sometimes impressively) resemble their behavior towards sentient beings, without believing that those machines are actually sentient. In such ways, people might, at most, only appear to believe those machines have mental states.

The analogy I shall examine involves human creations like puppets and fictional characters. Rodogno (2016) suggests that we are not deceived by robots because we respond to them more as we do to fictional characters than as we do to living humans. For example, when we feel sorrow for Anna Karenina we do not thereby believe (even subconsciously or unconsciously) that she was someone who actually suffered.

We “do not attempt to console her,” for instance (Rodogno, 2016, p. 260). Rodogno argues that being engrossed in robots resembles our emotional relations towards fictional humans. He says that “my joy at the robot…involves my imagining, accepting, mentally representing or entertaining the thought (without believing) that it is happy to see me” (Rodogno, 2016, p. 262).¹⁸

Rodogno (2016) reviews philosophical arguments that we are irrational or mistaken when we respond in relevant emotional and engaged ways to fictional characters (Radford, 1975). This is too large a debate to enter here. Instead, I will assume that the idea that people believe fictional characters actually have mental states is not widely held, even if a few philosophers hold it.

Of course, there may be some A-claim proponents who think that we do form these false beliefs about artistic creations, perhaps by projecting mental states onto them. My guess, for what it is worth, is that many do not. Moreover, the idea that people in general believe in relevant circumstances that puppets or fictional characters really have feelings seems to be highly implausible. Let us examine the situation more closely.

Clearly, we can feel things like joy for a fictional character’s successes, fear for their possible ruin, and sadness¹⁹ when they come to grief. Furthermore, people can become emotionally attached to fictional characters and care about their fortunes. Those characters may even serve as role models for some people, somewhat as real people can.

Nonetheless, there seem to be some crucial differences between our usual responses to characters from art and living humans. These differences appear to show a radical divergence in the nature of our beliefs about each. To explore this difference, consider some of our responses to movie characters.

In Nora Ephron’s romantic comedy Sleepless in Seattle, the character played by Rita Wilson cries (very dramatically) while recounting the misfortunes of Terry, the female hero of her favorite movie An Affair to Remember. She begins to laugh at herself, however, when her male companions, who are not moved by the film or by Terry’s plight, bring to her attention the strangeness²⁰ of crying for a person in a movie.²¹ She is well aware that Terry (played in the movie by actor Deborah Kerr) did not actually suffer when she was struck by a car and left disabled. It would obviously be crazy of her to think, for example, about sending Terry a sympathy card. Here we apparently see a striking difference between our beliefs about fictional characters and our beliefs about living humans.

Of course, the A-claim proponent may acknowledge that people do not have the false belief that the machines they ‘anthropomorphize’ really have feelings after their interaction with them. Proponents may similarly acknowledge that such people do not have the false belief most of the time, or when they are quizzed about it, or when they reflect on the relevant interactions with machines. Nonetheless, the proponent claims that forms of behavior such as some affective engagement with machines and certain uses of mentalistic language show that people did believe it at the time of that behavior and language use.

What can we glean from our movie example? We can first observe that a key difference between the response of the Rita Wilson character while she was tearfully recounting what befell Terry on the one hand, and the responses of her male companions and also of herself after her companions had affectionately mocked her tears on the other hand, is one of engrossment or emotional response to the suffering character in the movie. Crucially, this is not a difference of two conflicting beliefs about mental states.

It is not as if Rita Wilson’s character first believed that Terry actually suffered pain and humiliation, and then later recanted that belief. It is rather that she was first emotionally engaged with that character and her misfortune, and then later she was not.²² Nor did her male companions think for a moment that she believed Terry actually suffered. What they mocked was not a delusion or a false belief (for they did not assign any such belief to her), but their friend’s dramatic emotional response to the fate of a movie character.

We can apply this lesson to machines. Our movie example suggests that there need be no false belief when people speak or act in various emotionally laden ways towards nonhuman or nonanimal objects—precisely the sorts of ways that have tempted scholars to make the A-claim. People may act in ways that resemble behavior associated with a belief that fictional characters or machines really have feelings, and they may become engrossed and engaged with the appearance of mental states in those machines, without believing at any level that the machines really had or have mental states.

The same apparently applies to other artistic creations that move us in various ways. For example, people may respond to an expressive statue or a human-like cartoon character with sorrow, joy, pride, empathy, and compassion. It is conceivable that people might respond with automatic protectiveness and politeness towards a puppet in a dramatic puppet show. But these responses, as interesting as they are, may occur without the belief that these human-resembling entities really feel anything or that there is ‘anyone home’.

There seem to be lessons here too about our use of language. In the above sorts of cases, we may well say that ‘Person P believes the puppet/cartoon/statue/robot is sad’. Those objects, after all, are not like, say, a motionless volleyball. We do not normally say that a motionless volleyball is sad (or happy etc.)—unless, for instance, someone imaginatively suspends their disbelief, or names it ‘Wilson’ and draws a sad face on it.

But it is, we might suggest, natural and unobjectionable to describe a wailing, frowning, sulking humanoid robot as sad in certain contexts, and natural and unobjectionable to sometimes say we then believe that it is sad.²³ Calling such a robot sad in that way may be like calling a wailing, frowning, sulking puppet or cartoon character sad. Here the word ‘believe’ need have no implication of a belief in the actual presence of mental states in those puppets or cartoon characters. Nor must it have that implication for robots and other machines.

Indeed, someone who cannot see or believe that a robot or puppet or cartoon character with a mournful appearance, vocalization, and behavior is sad probably has a perceptual or cognitive problem. And someone who never feels even slightly sorry for a sad-looking puppet/movie character/robot/etc. might have an emotional problem.²⁴ But the point is that such locutions—‘the robot is sad’; ‘P believes the robot is sad’; ‘P feels sorry for the robot’—do not necessarily imply any belief, even of a transient or subconscious kind, that the robot is really or actually sad.²⁵

Someone might say it would be less objectionable to speak less loosely and more carefully here, saying instead something like ‘the robot appears sad’. In some contexts, that may be fair advice. Yet in general we are quite capable of understanding that the former, more direct expressions can be perfectly good and non-misleading descriptions of many robots and our beliefs about them.

One objection to the present attempt to problematize the A-claim is that machines like robots are very different to, say, fictional characters (Sweeney, 2021, p. 468). For robots are embodied²⁶ and may have autonomous agency of various sorts (see also O’Hara 2023). Unlike book characters, we can hug robots, refuse to smash them, trade insults with them, apologize to them, order them about, and converse with them.

Our relationships with such machines, it may further be suggested, are not static as they are with book characters such as Anna Karenina. Instead, those relationships can be ongoing and evolve over time. Therefore, it may be argued, we may be more inclined to believe some robots really have feelings.

I have two replies. First, this contrast may be slightly overdrawn. Millions of fans had an ongoing relationship of sorts with Harry Potter as each installment in the series came out. People can get very attached to made-up characters and even seek to influence their future adventures and fate.²⁷ Furthermore, unlike fictional characters, some other creations, like puppets, have physical bodies. Although puppets are operated by humans, puppets can take on a life and personality of their own and interact with members of the audience in ongoing and unpredictable ways.

Second, it is unclear that the various differences between certain robots, and other entities to which people respond and engage in many similar ways, point to a false belief regarding those robots but not the other entities. After all, the differences between (say) physical puppets and fictional book characters (not least physical embodiment) do not appear to be associated with a difference in belief regarding mental states.

That is, it seems improbable that we have the false beliefs about puppets and not fictional characters—or vice versa—even though there are various differences between these things and some of our respective responses to them. When, say, we describe puppets and fictional characters as happy or sad we do not seem to have false beliefs about the mental states of either. In this way at least, puppets and fictional characters are more similar than different.

Likewise, if people do not believe that Anna Karenina, Mary in Michelangelo’s Pieta, Terry from An Affair to Remember, or marionettes in a dramatically stirring puppet show really have feelings—notwithstanding the emotionally complex responses and the different interactions we may have with such various things—why should we think that people are given to believe that of human-like robots?

5.5 Something Fishy about the A-Claim?

I have not sought to definitively demonstrate that the A-claim is wrong but rather to show it can be challenged and needs a better defense. It certainly cannot be merely assumed. We have raised the problem that people who act in ways that seem to suggest they believe the machines had feelings, may, when asked, emphatically deny those machines did or do really have feelings. But that is not all: It is possible that such people, or that many people, believe that machines are not the type of thing that could intelligibly have mental experiences. If that is the case, further doubt is cast on the A-claim.

It does not seem impossible to suppose that most people do not believe that it even makes sense to think that robots can really be joyful or despairing. It seems quite possible that they no more think that of robots, even intelligent autonomous ones, than they think it of rocks, cars, puppets, statues, and fictional movie characters. On this view, it is not that many people think it unlikely that PARO, Kismet, or Pleo the robot could really be joyful or despairing, but that they think it is sheer nonsense to believe that they might be and misguided to seriously speculate about whether they are.²⁸

It seems possible, for example, that many people (and not just AI experts) would feel that Blake Lemoine, the Google engineer who claimed LaMDA the chatbot actually had feelings, was not so much mistaken as sadly divorced from reality. If this suggestion is on the right track, then the notion that people sometimes believe, even transiently or unconsciously, that a machine really has mental states is rendered still more curious.

This is not to say that we cannot have various false beliefs about machines and mental states. For example, people may (in a sense) believe a machine has phenomenal mental states in Turing imitation game situations. Consider the time an AI system (Google’s Duplex) fooled restaurant staff on a phone call that it was a real person making a dinner reservation.²⁹

However, the A-claim does not apply to such cases; it applies to beliefs that occur in full view of the robot or AI system. If it is true that few people would think that it even makes sense to think those machines really have mental states—and that anyone who thinks they might actually be sentient is profoundly deluded and perhaps in need of help—it is not exactly clear that they would so readily be tricked or fooled into believing, even momentarily or subconsciously, that those machines really have feelings and experiences. While this position is contestable, it does serve as another reminder that the onus of justification is now with the A-claim proponent.

6 Objections

A-claim proponents may reply that our doubts can be empirically refuted. Some of them may insist that it is just obvious or evident that a person who feels bad about smashing a struggling robot that is pleading for mercy believes (at least believes in the moment) that the robot actually fears for its life. However, we have already discussed how such human behavior need not imply that sort of belief. Still, proponents may reply that closer attention to how people behave and speak about machines will demonstrate the relevant anthropomorphic belief is there after all.

Once again, the onus of justification—empirical demonstration in this case—falls on the A-claim proponent. Should they produce additional evidence for the A-claim, they cannot merely give more examples of the types of responses we have already considered. The new evidence must show not just a difference in (say) the strength of human responses, but a difference in the kind of response (O’Hara 2023)—specifically, a difference that shows not just that people superficially appear to believe in machine mental states, but a difference that shows they do indeed believe in them. That might be achievable, but we would still need to see it done.

Some proponents might argue that empirical findings about certain bodily reactions or processes could support the A-claim. A proponent might place hope, for example, in discovering physiological and neurological processes (Wykowska et al., 2016) in people responding to machines that resemble the processes which occur when people respond to human mental states.

Yet such discoveries would be insufficient to substantiate the A-claim. Brain images and the like (Złotowski et al., 2015) may well be similar when we react emotionally to human distress and to robot ‘distress’. But what really matters is the nature of the beliefs that accompany those bodily processes. By themselves, empirical findings about bodily processes are inconclusive.

Another objection is that even if our skeptical questions are convincing for some types of machines, they may nevertheless be unsuccessful for other types of machines. For example, someone might agree that it is unlikely that people falsely attribute real feelings to existing robots, while claiming that the case is different for, say, new kinds of advanced conversational chatbots.

I do not want to rule that out. For example, many sophisticated robots still look pretty mechanical, while some large language models can partake in startling human-like conversations without being so patently weighed down by various nonhuman cues. Perhaps, then, many people do or will subtly attribute feelings (or other mental states) to machines like ChatGPT and its descendants.

Given that robots are often the stars players in the A-claim, the concession that we may not after all have illusions or false beliefs about robot feelings is a significant concession. But we might also have reservations about any attempt to re-focus the A-claim on new-fangled AI. For the arguments we adduced earlier also seem to apply to such AI as much as they do to robots and to earlier computing systems.

Current and future studies may show that people behave as if new conversational chatbots have feelings and that people even report believing that to be so. Yet, as we have stressed, we cannot necessarily accept such behavior, and even such self-reports, at face value—we also need to interpret their meaning. For example, people may say they believed the chatbot had feelings, without meaning at all to say the chatbot really had feelings as humans and animals have feelings. Such studies will therefore need to be conducted with care and a degree of philosophical sophistication.

7 Implications of Rejecting the A-Claim

An accurate understanding of our responses to human-like and animal-like machines is, I think, worthwhile seeking in itself. But we might also wonder whether our reassessment of anthropomorphism has further practical implications. We have noted that the kind of anthropomorphism described by the A-claim generates both concern and enthusiasm. I only have space to briefly discuss a subset of examples, which we can now do in the light of the possibility that such anthropomorphism is a popular myth.

A commonly voiced concern is that illusion or deception about the true (mindless or insentient) nature of machines could lead people to form unhealthy or problematic relations with them. For example, vulnerable people might get addicted to their robotic or virtual ‘partners’ and socially withdraw. Sherry Turkle (2017) worries that “empathy machines” are too “seductive and offer the wrong payoff: the illusion of companionship without the demands of friendship, the illusion of connection without the reciprocity of a mutual relationship.” Another worry is that people may over-trust human-like machines and then do risky things, like surrendering their privacy to them.

If the A-claim is dubious, the worry that people might have relations with certain machines based on a false belief about subjective experience is also dubious. However, rejecting the A-claim need not mean rejecting other concerns about human-machine relations. For example, people might still develop dangerous, addictive, over-trusting, worthless, or undignified relations with autonomous machines whether or not they believe those machines really have feelings (Turkle, 2010). For some people, the compelling semblance of mental experience in machines may be enough to trigger the problematic behaviors without requiring any illusion of sentience.

We might even want charming robots to offer warnings to users—not so much warnings that the machines do not really or truly feel love or affection,³⁰ but warnings that they are not as ‘friendly’ or capable as they seem, or warnings about developing excessively strong feelings towards them.

A person who ‘fell in love’ with Anna Karenina, Harry Potter, or Misery Chastain (as Annie Wilkes does in Stephen King’s horror story) and, in their obsession, neglected their friends and family or (as in Annie’s case) did even worse things, would be doing something unhealthy or immoral or otherwise problematic irrespective of the possession of deluded beliefs about those fictional personalities. But the fact that human-like robots and emerging AI may be much more seductive companions for some people than various human creations (including earlier, less human-like machines) could be socially significant.

A parallel response can be made to enthusiasm towards the A-claim. The idea that we naturally and instinctively attribute actual mental experiences to machines and think of them as (say) truly caring and compassionate could conceivably facilitate co-operative human-machine relations (Onnasch & Roesler, 2019). This could occur, for instance, in plane cockpits, factories, and care settings where machines and humans work together.

But again, there is arguably no need for people to believe the machines have real feelings for these positive effects to occur. We already know that people have responses such as protectiveness, friendliness, affection, and politeness toward machines. Indeed, these are the very kinds of responses that led scholars to posit the existence of the false beliefs.

Machines can be good at simulating mental states and inner lives and will surely get better and better at it. As we have suggested, people can say things like, ‘the robot is upset’, without meaning or implying that the robot is really or actually upset. Humans can have many interesting and strong emotional responses to robots without needing to have illusions about robot sentience or to make erroneous attributions of mental states.

Finally, some commentators have suggested that being ‘cruel’ (or kind) to robots could make us cruel or (kind) to humans and animals (Darling, 2016, 2021; Flattery, 2023). One version of this contention relies on the notion that a belief that robots have feelings sits behind this knock-on effect. The idea here recalls the idea that cruelty to animals which we believe are sentient can promote cruelty to humans. Rejecting the A-claim would mean rejecting this particular position.

However, the general contention can perhaps be partially salvaged, in a now familiar way: If robots or other machines are human-like or animal-like enough to prompt various kinds of strong emotional engagement, perhaps some sort of worrying or welcome moral conditioning effect will still occur, without anyone attributing real feelings to these machines (Coghlan 2019).³¹

8 Conclusion

The claim that people often naturally and irresistibly believe at some level that robots and AI systems really or actually have mental states like feelings has become widespread, almost orthodox. I sought to problematize that claim by questioning common arguments about beliefs that are used to support it and suggesting alternative explanations. Contrary to a common view, it is not yet clear that humans are prone to having false beliefs in human-like machine feelings—even transiently or subconsciously and even when impressively autonomous and intelligent humanoid machines are involved. To insist that such machine-related illusions are in fact a widespread phenomenon would require further argument.

If anthropomorphizing of the type we discussed turns out to be not prevalent after all, it seems we do not have to be so concerned about machines engendering illusion and deception about their non-existent mental lives. Nonetheless, even if the A-claim is a popular myth, some concerns and enthusiasms about various emotional responses to machines might still apply, and some of these concerns may need addressing. Finally, simply getting clearer about the difficult concept of anthropomorphism is useful for improving understanding of human behavior, belief, and relations with machines that increasingly resemble humans and animals.