R1.1 Real versus realistic

The bare nouns “tree,” “gun,” and “dog” ordinarily denote trees, guns, and dogs that are real or genuine. The phrases “artificial tree,” “fake gun,” and “toy dog,” on the other hand, denote depictions of trees, guns, and dogs as used, for example, in the theater or make-believe play. Artificial trees, fake guns, and toy dogs can be described as “realistic,” but not as “real” (Oxford English Dictionary). In these cases, “real” (or “genuine”) contrasts with “artificial.” For something to be construed as real or genuine, it needs to pass two tests:

The humanoid robot Nao, in our account, is therefore an artificial social agent that depicts a genuine social agent. And for us, a genuine social agent is a living being that is able to interact socially with humans. So, although Nao_prop might be described as “realistic” for a social being, it would not be described as a “real” social being. Hortensius and Cross (Reference Hortensius and Cross2018, p. 93) agree: “We use the term artificial agents to refer to robots (including those that are machine-like, pet-like, or human-like).”

Many commentators, however, describe social robots such as Nao_prop as genuine social agents. Orgs & Cross, contrary to Hortensius and Cross, insist “The robot [Smooth] does not pose as a social agent, it is a genuine social agent” even though it doesn't pass either test for being a real social agent.

Eng et al. say “Robots are not human beings, but neither are they mere depictions of social agents. Instead, they are seen as real social agents.” But for an object to be “seen as” a real X, the object cannot itself be a real X. That is, to say that a social robot is seen as a real social agent is to imply that it is not a real social agent – that it only looks like one, that it depicts a real social agent. So, despite their objections, Eng et al. appear to agree that social robots are depictions of social agents.

R1.2 Imagination

Still, people interacting with social robots imagine that they are interacting with real social agents. Owners of Sony's robot dog Aibo, for example, offer spontaneous reports such as these (Friedman, Kahn, & Hagman, Reference Friedman, Kahn and Hagman2003): “I feel I care about him as a pal.” “He always makes me feel better when things aren't so great.” “My emotional attachment to him … is strong enough that I consider him to be part of my family, that he's not just a ‘toy.’” Many of these “feelings,” we argue, are based on people's engagement, engrossment, or immersion in the scenes depicted, a phenomenon that has long been recognized in novels, films, and plays.

Novels and films, according to Chatman (Reference Chatman1980), divide into a discourse (the medium people process) and a story (the content they are to imagine), and people get engrossed in the content (Clark, Reference Clark1996, p. 366; see Clark & Van Der Wege, Reference Clark, Van Der Wege, Tannen, Hamilton and Schiffrin2015). As Gardner (Reference Gardner1985, p. 132) puts it about novels, “The writer's intent is that the reader fall through the printed page into the scene represented.”

Literary theorists, Gerrig notes, call this experience an aesthetic illusion. He cites Wolf (Reference Wolf, Hühn, Pier, Schmid and Schönert2009, p. 144), who said that the illusion “consists predominantly of a feeling, with variable intensity, of being imaginatively and emotionally immersed in a represented world and of experiencing this world in a way similar (but not identical) to real life.” In Wolf's view, being immersed can range from “the disinterested observation of an artifact” to “the complete immersion (‘psychological participation’) in the represented world” (p. 144). Much the same feelings arise, we suggest, with ventriloquist dummies, hand puppets, and social robots.

R1.3 Supporting imagination

Orgs & Cross assert that “Clark and Fischer link the quality of a social robot to its resemblance to a human agent.” This is a serious misreading. “All social robots,” we wrote, “represent nonstandard characters,” beings that one may never have met, seen, or thought about before. And the “quality of a social robot” is tied not to its literal resemblance to a character, but to the depictive devices by which the character is represented (see Clark & Van Der Wege, Reference Clark, Van Der Wege, Tannen, Hamilton and Schiffrin2015). Here are a few such devices that film makers, play directors, and puppeteers have used to immerse people in the scenes depicted.

R2.1 Real experiences

Reeves argues that engaging with social robots includes not only imagined experiences “guided by pretense,” but “natural experiences that are direct, automatic and independent of any thoughtful mapping between what is real and depicted.” At an IMAX film, we are surrounded by an 18 × 24 meter screen, and when the camera goes over a mountain ridge, we feel our stomachs rise into our throats. Reeves calls experiences like this “natural responses.” (See also Förster, Broz & Neerincx [Förster et al.], Seibt, and Vogeley.)

Natural responses, Reeves argues, are a product of what Kahneman (Reference Kahneman2011, Reference Kahneman2012) called system 1 thinking. System 1 is fast, intuitive, and involuntary, whereas system 2 is slow and “performs complex computations and intentional actions, mental as well as physical” (Kahneman, Reference Kahneman2012, p. 57). The time-locked processes we described in section 7.3 belong to system 1. The audience at Hamlet must imagine Hamlet stabbing Polonius at precisely the same time as the actor playing Hamlet is “stabbing” the actor playing Polonius. The percept-based processes we discussed in section 7.3 also belong to system 1. People are usually able, without reflection, to recognize an apple as an apple. Both of these processes would be natural responses.

Reeves concludes with a significant insight: “Much of the history of media technology is about inventions that promote natural responses.” This is especially clear in depiction-based media, such as film, television, video, and telepresence technology. Reeves' point applies just as forcefully to social robots. People's experience with them seems real because it is based in part on natural responses.

R2.2 Experiencing emotions

If system 1 “generates emotions” as Kahneman (Reference Kahneman2012, p. 57) argued, then emotions should be part of people's experience with depicted scenes. In research we cited (Gross, Fredrickson, & Levenson, Reference Gross, Fredrickson and Levenson1994), students viewing a clip from the film Steel Magnolias often became so immersed in the story that they got sad and cried. Clips from other films reliably evoke emotions ranging from amusement, anger, and contentment to disgust, fear, and sadness (Gross & Levenson, Reference Gross and Levenson1995).

These emotions, Blatter & Weber-Guskar note, are “cases of what others have called fictional emotions.” They cite Gendler and Kovakovich (Reference Gendler, Kovakovich and Kieran2006), who contrast “real” emotions, which are about real situations, with “fictional” emotions, which are about fictional ones. But emotions require a finer analysis.

R2.3 Continuity of experience

Aesthetic illusions with novels and films tend to be continuous. Once people immerse themselves in a story, they stay immersed in it until they break out of it. The same should hold in people's engagement with social robots.

Rueben takes a different view. “There are reasons to suspect that meta-cognition about construing social robots as depictions would be more difficult – or absent – than Clark and Fischer discuss.” He goes on: “The amount of time and effort that participants give to this reflection could greatly affect their responses.” But in novels, people's immersion is continuous; they don't have to re-immerse themselves with each new sentence or paragraph. The same is true with social robots. People don't need extra “time and effort” for “reflection” at each new step of their interaction with a robot. Once engaged with a robot, people can stay engaged.

A final point is due to Wrede, Vollmer & Krach (Wrede et al.) (see also Healey, Howes, Kempson, Mills, Purver, Gregoromichelaki, Eshghi & Hough [Healey et al.]). People find it easy to stay immersed in an imagined scene as long as it goes smoothly. But once they notice an inconsistency in the evidence, the spell is broken, they experience a breakdown, and the physical prop is foregrounded. The same happens in the theater when an actor forgets a line, the scenery falls over, or a stage light burns out. Breakdowns like these remind viewers of the base and depiction proper that lie behind the scene depicted. In our example, “When a robot stops moving, people must decide ‘Did the social agent fall asleep, or did the artifact's battery die?’” And people may go for one interpretation one minute and another the next (Fischer, Reference Fischer2021).

R3.1 Beyond “mere depictions”

At its heart, a depiction is a representation of something else – a sign that signifies an object. The philosopher Peirce (Reference Peirce, Hartshorne and Weiss1932, Reference Peirce1974; Atkin, Reference Atkin, Zalta and Nodelman2010) argued that signs come in three main types. (1) A symbol signifies an object by rule. The sound /hunt/ signifies “dog” for German speakers by a rule of German. (2) An index signifies an object by a physical connection with the object. An arrow is an index that signifies the thing it points at. And (3) an icon signifies an object by its perceptual resemblance to the object. A video of a dog barking at a squirrel is an icon that signifies the scene by its visual and auditory resemblance to that scene. Many signs, Peirce noted, are mixed signs – combinations of two or three of the basic types. (Petersen & Almor, alas, overlooked our citations in criticizing us for not tying our model to Peirce, signs, and icons.)

People communicate by producing signs for each other, and that leads to three methods of communicating: (1) describing things with symbols; (2) indicating things with indexes; and (3) depicting things with icons. Depicting is, therefore, a basic method of communication on a par with describing and indicating (Clark, Reference Clark2016). Most acts of communication are composites of these methods.

Acts of communication, in turn, are based on the recognition of a producer's intention in producing them (Grice, Reference Grice1957, Reference Grice1969). When Kate tells Lionel “I caught a fish this long (holding up two hands, palms in, 30 cm apart),” she intends him to recognize what she means by her gestural depiction – that the fish was 30 cm long – from two types of information: her perceptual display (the content, place, and timing of her gesture); and her intention, or purpose, in producing the display (as expressed in part in “I caught a fish this long”).

Kate and Lionel's actions aren't unilateral – separate and autonomous. They are bilateral – conditional on each other. Kate has to coordinate her display (its content, placement, and timing) with Lionel's interpretation of her display. In Grice's account, these two actions are conditional on each other even when they are displaced in space, time, or both. The same requirement holds for depictions such as social robots.

R3.2 On modern art

Depictions, in this view, have a purpose that recipients are intended to recognize. Orgs & Cross disagree. “Much of contemporary art,” they argue, “neither depicts nor represents.” As evidence, they cite delightful examples from performance art such as dance and theater that “dissolve the binary distinction between depicted and depictive scene, or acting and not-acting.”

Orgs & Cross's argument, however, ignores purpose. The very point of much modern art is to have no practical point. Artists have license to entertain, divert, or fascinate however they like and often leave purpose indeterminate. When Andy Warhol painted “Campbell's Soup Cans,” why did he depict Campbell's soup cans, and why 48 of them? Why did Jackson Pollack drip paint on a canvas in the patterns he did? Why are so many works entitled “Untitled”? Other artists play with trompe l'oeil, deceptive perspective, and visual illusions. Orgs & Cross's examples are of this ilk, and viewers appreciate them for what they are.

Everyday depictions, however, have a practical purpose. People base their interpretation of Michelangelo's David, Kate's depictive gesture, and social robots in part on what they believe the creators intended. Genuine depictions and artistic creations may live in the same world, but they are not all processed in the same way.

R4.1 Joint activities

The basic idea is that whenever people interact socially, they do things together: They coordinate with each other in joint activities (Clark, Reference Clark1996, Reference Clark2005). And to coordinate with each other, they have to agree on their joint actions and positions, and that requires communication. Here is an example from two people assembling a TV stand (from Clark, Reference Clark2005):

In turn 1, Ann proposes a joint position for the two of them, and in turn 2, Burton takes up her proposal and agrees to it. People can also reach agreement with gestures, which are any “visible acts of communication” (Kendon, Reference Kendon2004):

The principle is this: “It takes coordination for people to do things together, no matter how simple, and it takes communication to achieve that coordination” (Clark, Reference Clark2005, p. 507).

Social robots require the same techniques. In section 2.4 of our paper, we illustrated two exchanges between the robot Smooth and a woman named Beth:

In the first pair of turns, Smooth offers Beth water, and she accepts his offer and takes a glass. In the second pair of turns, he makes a toast, and she takes it up and reciprocates. All four turns rely on both speech and gestures. In an example from Guo, Lenchner, Connell, Dholakia, and Muta (Reference Guo, Lenchner, Connell, Dholakia and Muta2017), a woman asked a robot concierge for the location of a bathroom. She posed the question in turn 1, and he took it up and answered it in turn 2:

So, for humans and social robots to coordinate with each other, they must reach agreement on how to get things done together. A model of social robots unable to do this cannot be complete.

R4.2 Common ground

When two people communicate, they assume certain information to be part of their current common ground, and they add to that body of information with each new act of communication (Clark, Reference Clark1996, Ch. 4; Stalnaker, Reference Stalnaker and Cole1978). The same goes for social robots. As Carroll argued, “Future robots must effectively coordinate common ground with humans.”

Common ground comes in two main types, and social robots need to track both:

1. (1) Personal common ground is information people establish based on their joint experiences – what they see, do, and communicate with each other. Suppose a woman named Jane asks an actual concierge, “Where is the bathroom?” and he answers “The bathroom is in aisle 13.” With her question, the two of them would add her request to their current common ground, and with his answer, they would add the location of the bathroom.

2. (2) Communal common ground is information people share as members of the same cultural communities, such as their nationality, occupation, language, gender, age cohort, or residence. Although Beth, for example, spoke to her friends in Danish, she took for granted that she and Smooth both knew English and spoke to him in English.

For face-to-face coordination to go smoothly, communication must also be reliable. That requires a process called grounding: People in joint activities try to establish, as they go along, the mutual belief that they have understood each other well enough for current purposes (Clark, Reference Clark1996; Clark & Brennan, Reference Clark, Brennan, Resnick, Levine and Teasley1991; Clark & Schaefer, Reference Clark and Schaefer1989; Clark & Wilkes-Gibbs, Reference Clark and Wilkes-Gibbs1986). Speakers monitor their conversations both for evidence of success (e.g., “uh huh,” “good God,” “oh,” and nods from addressees) and for evidence of failure (e.g., misunderstandings that need repairing). When one woman asked Guo et al.'s robot concierge something he couldn't understand, he cleared it up before going on:

The side sequence in turns 2 and 3 is one of many strategies people use for repairs (Dingemanse et al., Reference Dingemanse, Roberts, Baranova, Blythe, Drew, Floyd and Enfield2015; Schegloff, Jefferson, & Sacks, Reference Schegloff, Jefferson and Sacks1977). Social robots, then, also need strategies for tracking success and failure in their social interactions (see Healey et al. and Wrede et al.).

The problem, as we will show, is that alternative models of social robots have no means for accumulating common ground or for grounding what they do and say.

R4.3 Social agents are individuals

In communication, common ground is accumulated by individuals and not by types of individuals (Clark, Reference Clark1996, Reference Clark2021). When the actual concierge answered Jane's question, he tried to add new information to the common ground he shared with Jane the individual and not with some generalized person. He tried to anchor his references (“the bathroom” and “aisle 13”) to an actual bathroom and an actual aisle he assumed was in his and her current common ground. It is no different for people communicating with a robot concierge.

Individual entities are fundamentally different from types of entities. Thoughts about “Jane” and “the concierge” are about the individuals they index. Thoughts about “woman” and “concierge,” in contrast, are about the types of individuals they describe. Crucially, an indexical thought cannot be reduced to a set of descriptive thoughts (Perry, Reference Perry1979, Reference Perry1993; Recanati, Reference Recanati2012, Reference Recanati2013). An individual like Jane cannot be represented as merely a bundle of attributes. Yet that is the assumption behind many models of social robots based on anthropomorphism, mind perception, and trait attributions (e.g., Girouard-Hallam & Danovitch, Orgs & Cross, Ziemke & Thellman). The point may seem technical, but it is a significant strike against those models.

R4.4 Interim summary

In short, for a robot to be a social robot, it must represent a real social agent, an individual, able to engage humans in joint activities. The agent must be able to:

1. (1) Coordinate with humans in joint activities, however restricted the activities,

2. (2) Communicate with humans well enough to advance these activities,

3. (3) Accumulate common ground with individual humans as these activities advance (as Carroll suggests),

4. (4) Ground what gets said and done well enough for current purposes.

A number of proposals in the commentaries are incompatible with these features. In the next two sections, we take up four of these proposals.

R5.1 Problems with anthropomorphism

These two examples illustrate serious problems for anthropomorphizing as an account of social robots.

R5.2 Embodiment

Ravikumar et al. argue for treating social robots as embodiments of social robots: “[E]ven if social robots are interactive depictions, people need not mentally represent them as such. Rather, people can directly engage with the opportunities for action or affordances that such robots/depictions offer to them.” This position, however, also has problems.

Suppose Brigitte saw her old friend Alain and wanted to talk to him. She knew him well enough to assume they shared a great deal of common ground, such as how to approach, hug, kiss, and gossip with each other. She wouldn't have known how to approach, hug, kiss, or gossip with an anonymous body. She needed to know it was Alain. Brigitte would have the same problem with the robot Asimo. Identifying an entity as a body is not enough to engage with it even in what Ravikumar et al. called “sociocultural settings.”

Embodiment also doesn't distinguish features that allow affordances from features that do not. As we noted, “Observers of a depiction implicitly realize that only some of its features are depictive,” and only the depictive features afford the right inferences. Asimo_prop's hand depicts a real hand, which affords handshakes, but Asimo_prop's ears happen not to contain senses of hearing, so they do not afford headphones. Asimo_prop's lack of a mouth doesn't afford speaking, yet Asimo_char is able to speak. Discrepancies like these differ from robot to robot. If so, how can people “directly engage with the opportunities for action or affordances that such robots/depictions offer to them”?

Communicating with a robot is even more of a challenge. With Asimo, should Brigitte speak French, use French gestures, and kiss him on both cheeks, as she would with Alain? The affordances availed by Asimo_prop's body offer no answers.

R5.3 Intentional stance

As the commentaries by Veit and Browning and Ziemke and Thellman noted, Dennett (Reference Dennett1987, Reference Dennett1988) proposed two strategies, or stances, that attribute intentions to systems such as social robots:

There is merit in both stances. For Asimo, the intentional stance applies to the social agent it represents (Asimo_char), and the design stance applies to Asimo's physical design (Asimo_prop).

These two stances, however, are unilateral interpretations and not the bilateral ones needed for social robots. It isn't enough to attribute certain mental states to Asimo_char. People must attribute the mental states they believe they were intended to attribute to Asimo_char. More than that, their interpretation of Asimo_char must be custom-built for the nonstandard individual that Asimo depicts.

R6.1 Measuring traits

Traits are often studied by asking people to rate how much human attributes apply to nonhuman entities (see Epley, Waytz, & Cacioppo, Reference Epley, Waytz and Cacioppo2007; Gray, Gray, & Wegner, Reference Gray, Gray and Wegner2007; Reeves, Hancock, & Liu, Reference Reeves, Hancock and Liu2020; Weisman, Dweck, & Markman, Reference Weisman, Dweck and Markman2017; see Thellman, de Graaf, & Ziemke, Reference Thellman, de Graaf and Ziemke2022). In one study (Gray et al., Reference Gray, Gray and Wegner2007), participants rated 13 “characters,” which ranged from a baby, a fetus, a dead woman, and a frog to God, “you,” and a robot. All but the fetus were given proper names. People rated the entities on dimensions of “experience” (e.g., hunger, pain, fear, pride) and “agency” (e.g., self-control, morality, memory).

What people rated, however, weren't objects they had interacted with. They were static photos (e.g., Phillips, Ullman, de Graaf, & Malle, Reference Phillips, Ullman, de Graaf and Malle2017; Reeves et al., Reference Reeves, Hancock and Liu2020; Ruijten, Reference Ruijten2015), videos, labels (Lencioni, Carpinella, Rabuffetti, Marzegan, & Ferrarin, Reference Lencioni, Carpinella, Rabuffetti, Marzegan and Ferrarin2019), or descriptions of such objects. Here is Gray et al.'s description of their robot:

Clearly, this paragraph isn't about Kismet_base or Kismet_prop, but about Kismet_char. Only the character would have a proper name, be male, “engage people in natural interaction,” “observe … social signals,” “deliver his own signals,” direct his gaze, and produce a “facial expression,” “body posture,” and “vocal babbles.” It is no surprise that Kismet_char was judged to have agency.

What if people had been asked whether Kismet's machinery experienced hunger, pain, or fear, or possessed pride or morality? Their ratings would surely have changed. Even if they thought Kismet's machinery experienced “hunger” (as when its battery died), they would have based their judgment on a metaphorical, not literal, interpretation of hunger. Metaphors, indeed, are a widespread issue.

R6.2 Metaphor problems

People seem willing to attribute “mental states” and “minds” to all sorts of artifacts. These include not only cars (Petersen & Almor, Ziemke & Thellman), but gadgets. A study by Epley, Akalis, Waytz, and Cacioppo (Reference Epley, Akalis, Waytz and Cacioppo2008) examined five gadgets, including “Clocky (a wheeled alarm clock that ‘runs away’ so that you must get up to turn it off)” and “Pillow Mate (a torso-shaped pillow that can be programmed to give a ‘hug').” People were asked to rate “the extent to which the gadget had ‘a mind of its own,’ had ‘intentions,’ had ‘free will,’ had ‘consciousness,’ and ‘experienced emotions.’”

But what were these people rating? When asked about the extent to which Pillow Mate had “a mind of its own” or “free will,” they were forced to interpret “mind” and “free will” as metaphors. If they had been asked whether Pillow Mate really or actually or literally had a mind of its own or free will, they, like us, would have said no (see Thellman et al., Reference Thellman, de Graaf and Ziemke2022). Likewise, if Joe the driver had been asked if he was really or actually or literally chastising the car, he, too, would have said no. And so would the pedestrian when asked whether the driverless car could really or actually or literally “see” him or her, “understand” things, or have “intentions”? (Did Romeo think that Juliet was really or actually or literally “the sun”?)

What we have here are metaphors: “hunger,” “a mind of its own,” “intentions,” “free will,” “consciousness,” “emotions,” “chastising,” “see,” and “understand.” It is a mistake to equate metaphorical attributions like these with their literal counterparts. They are not equivalent, and treating them as equivalent leads to misleading claims about both traits and social robots.

R8.1 Theory

Is the depiction model a theory (see Bartneck)? The answer is clearly yes. A theory, according to Dennis and Kintsch (Reference Dennis, Kintsch, Halpern and Sternberg2007), should satisfy certain criteria, and the depiction model does just that. It accords with empirical data; it is precise and interpretable; it is coherent and consistent; it predicts future applications; and it provides explanations that go beyond the model itself. We have cited evidence supporting each of these criteria.

We have also investigated alternative accounts. The media equation was one of the early inspirations for our work, and while the depiction model makes many of the same predictions, it also explains phenomena not covered by the media equation (see target article, sect. 2.4). In his commentary, Reeves, one of the progenitors of the media equation, appears to agree (contra Bartneck). Anthropomorphism and trait attributions are two other alternative accounts, but these suffer from the empirical and conceptual problems we discussed earlier. The point of the depiction model, in sum, is to explain social robots in terms of a broader theory, namely, how people engage with depictions, and we believe it succeeds at that.

R8.2 Social roles

In our paper, we distinguished between self-agents, who “act on their own authority and are fully responsible for their actions,” and rep-agents, who “act on the authority of specified principals.” When Susan works as a server for Goldberg's Bakery, she is a rep-agent for the bakery, but once off work, she is on her own, a self-agent. Healey et al. (see also Carroll) worried about the roles such agents take.

Each individual person, we assume, has an individual role that is continuous and enduring. Susan remains Susan whatever else she does. But individuals also take on additional roles, social roles, that change with the social situation. They may take the social role of sister, companion, playgoer, or bus rider for people they interact with, and teacher, tutor, or concierge in working for others. Susan chose the particular role of server when she hired on at Goldberg's Bakery. Social robots could, in principle, take multiple social roles, but the robots we know of are able to take only one social role.

R8.3 Future

Predicting the future is dangerous. In about 1970, Herbert A. Simon, one of the founders of artificial intelligence (AI), suggested to colleagues that people would be able to talk with computers in 10–20 years. Fifty years later there is still no such computer. People share limited facts with virtual assistants such as Siri, Alexa, and Google Assistant, but as conversations go, these exchanges are primitive (see Marge et al., Reference Marge, Espy-Wilson, Ward, Alwan, Artzi, Bansal and Dey2022). Today's AI systems for conversation still cannot deal with such features as the timing of turns, the use of uh and um, performed depictions, pointing, anchoring, grounding, irony, sarcasm, and empathy. According to Bender and Koller (Reference Bender and Koller2020), current AI models of language cannot be complete in principle because they are based on the form of language alone.

So Malle & Zhao are brave souls. They are clear-sighted about today's robots when they say: “[C]urrent social robots are advertised to be much more capable than they really are – that is, they are largely a pretense, a fiction.” But they venture into Herbert Simon territory when they go on (see also Caruana & Cross; Franklin, Awad, Ashton, & Lagnado (Franklin et al.); and Stibel & Barrett):

As Simon's prediction shows, the future doesn't always work out the way we think – often for principled reasons. With social robots, who knows what those principles will be.

Still, no matter how humanlike social robots become, they will never be humans. They will always be artifacts intended to depict humanlike social agents. To frame them otherwise would be to engage in deception.