ABSTRACT UNCANNY PROCESSING: MISMATCHES BETWEEN PROCESSING STYLE AND FEATURAL CUES TO HUMANITY CONTRIBUTE TO UNCANNY VALLEY EFFECTS

Transcription

1 ABSTRACT UNCANNY PROCESSING: MISMATCHES BETWEEN PROCESSING STYLE AND FEATURAL CUES TO HUMANITY CONTRIBUTE TO UNCANNY VALLEY EFFECTS by Steven Michael Almaraz The uncanny valley is the tendency for highly humanlike, but non-human agents (e.g., robots, animated characters, dolls) to be perceived as creepy or unsettling, relative to their less humanlike counterparts. Recent research has pointed to mismatching signals of humanity as a possible explanation for the uncanny valley. The current work aimed to extend this hypothesis by investigating whether conflicting signals of humanity from face processing styles and featural cues can trigger negative affect. To this end, participants viewed faces that were morphed on a continuum from full dolls to full humans and indicated the extent to which these faces are unsettling. Critically, on half of the trials, faces were inverted to disrupt configural face processing, a processing style that involves viewing faces as a single Gestalt and is a cue for humanity. When faces were highly humanlike, they were experienced as less creepy than less humanlike faces, but when such targets were inverted, processing and featural signals did not disagree with one another, and some of the feelings of unease were alleviated.

2 UNCANNY PROCESSING: MISMATCHES BETWEEN PROCESSING STYLE AND FEATURAL CUES TO HUMANITY CONTRIBUTE TO UNCANNY VALLEY EFFECTS A Thesis Submitted to the Faculty of Miami University in partial fulfillment of the requirements for the degree of Master of Arts by Steven Michael Almaraz Miami University Oxford, Ohio 2017 Advisor: Kurt Hugenberg Reader: Heather Claypool Reader: Jonathan Kunstman 2017 Steven Michael Almaraz

3 This Thesis titled UNCANNY PROCESSING: MISMATCHES BETWEEN PROCESSING STYLE AND FEATURAL CUES TO HUMANITY CONTRIBUTE TO UNCANNY VALLEY EFFECTS by Steven Michael Almaraz has been approved for publication by The College of Arts and Science and Department of Psychology Kurt Hugenberg Heather Claypool Jonathan Kunstman

4 Table of Contents Introduction... 1 The Uncanny Valley 1 Configural Face Processing as a Perceptual Cue to Humanness. 4 Current Work... 7 k Experiment 8 Method. 8.Participants.. 8.Materials.. 8.Procedure. 9 Results.. 9.Human Likeness Uncannyness k Discussion References k iii

5 List of Figures Figure Figure Figure Figure Figure k iv

6 The uncanny valley is a phenomenon whereby people react to somewhat humanlike objects, usually dolls or robots with humanlike faces, with increasing positivity as the objects become more similar to humans, but only up to a point. As objects become very highly humanlike, they are often perceived negatively, and typically as creepy or uncanny. This sharp evaluative drop in appeal occurs just before a stimulus is perceived as truly human, and is known as the uncanny valley (see Figure 1). In his original observation of the uncanny valley, Mori (1970) proposed that understanding the uncanny valley is necessary if human-machine interactions are to improve as technology becomes more sophisticated. In the current work, I propose that understanding the uncanny valley can provide important insight into both how faces are processed, and how that face processing generates social inferences. Although extensive research has demonstrated that the uncanny valley is multiply determined, in the current work, I propose that the uncanny valley may be caused, at least in part, by perceivers employing configural face processing, which is typically reserved for human faces, for stimuli that are nonhuman. To this end, I begin with a discussion of the uncanny valley and its proposed causes, before moving on to discuss configural face processing and its relationship to the experience that a face is human. Finally, I then propose one study designed to test the relationship between configural face processing, the perception of humanness, and the uncanny valley. The Uncanny Valley Although the uncanny valley has a history of ad hoc demonstrations in literatures on robotics and human-machine interactions (Matsui, Minato, MacDorman, & Ishigoru, 2005; Minato, Shimada, Ishigoru, & Itakura, 2004; Mori, 1970), only recently have behavioral scientists worked to reliably replicate and explain the mechanisms behind the uncanny valley. The original concept devised to measure the uncanny valley (Mori, 1970), bukimi, translated loosely to eeriness, and a lack of shinwakan, which translates to comfort. These issues of eeriness and comfort remain central measurement foci in this research literature. More recently, some research has focused more on comfort or likeability (e.g., Bartneck, Kulić, Croft, & Zogbhi, 2009), other work has focused on discomfort or eeriness, associated with high arousal natural affect such as disgust, fear, and anxiety (e.g., Burleigh, Schoenherr, & Lacroix, 2013; Ho & MacDorman, 2010; Ho, MacDorman, & Pramono, 2008). In this latter vein, the most recent psychometric evidence (Ho & MacDorman, 2010) indicates that eeriness is a more optimal fit to the data in the uncanny valley literature. In the current paper, I will use the terms 1

7 uncanny, eerie, and creepy interchangeably (as is typical in the literature) to describe an affectively negative feeling incorporating nervousness, unattractiveness, disgust, and anxiety (e.g., MacDorman & Ishiguro, 2006; MacDorman et al., 2009; Mathur & Reichling, 2016; Pollick, 2009). In one of the first methodologically rigorous replications of the uncanny valley, MacDorman and Ishiguro (2006) morphed two images of robot faces with two images of human faces, creating images that varied in their objective humanness in 10% increments, ranging from fully robot to fully human. Participants rated each face morph, one at a time, along three dimensions: very machinelike to very humanlike, very strange to very familiar, and slightly eerie to extremely eerie. As predicted, they found that although there was a monotonic increase from machinelike to humanlike, ratings of familiarity (i.e., valence) showed an upward trend as the stimuli became more human, but with a dip near the middle, indicative of the uncanny valley. This drop in familiarity ratings approximately matched a peak on the eeriness scale, serving as one of the first rigorous demonstrations of the uncanny valley using highly controlled stimuli. More recently, Looser and Wheatley (2010) demonstrated similar findings using a closely related technique, but targeting perceptions of targets animacy. Similar to MacDorman and Ishiguro, Looser and Wheatley had participants judge a series of morphed doll-human faces, which ranged monotonically from 100% doll through 100% human, in 10% increments. Looser and Wheatley found that the perception that a stimulus is animate (i.e., is human; has a mind) is a non-linear function, relying heavily on high perceptual similarity to human faces. Highly humanlike, but inanimate faces are easily parsed from very similar but truly human faces. Further, and in line with the findings of the uncanny valley, these judgments are made categorically that is, it is only when a face has a very strong perceptual signal of humanness is it judged as human. Somewhat human faces are still seen as inanimate. Although Looser and Wheatley did not collect data on evaluations, these humanlike yet inanimate faces would presumably be evaluated negatively (i.e., fall within the uncanny valley). Beyond establishing the uncanny valley effect, multiple researchers have recently investigated the causes of the uncanny valley. A number of theoretical mechanisms for the uncanny valley have been proposed, many of which argue that the uncanny valley is caused by perceivers experiencing conflicting signals of humanness and non-humanness from nearly human stimuli. 2

8 As previously noted, Looser and Wheatley (2010) established that perceptions of faces as possessing human traits is relatively categorical we typically experience faces as either human or not. Some researchers have suggested that stimuli which fall right at the boundary between these two categories (human and non-human) generate ambiguity, and this disfluency in categorizing the stimuli may be experienced as unease or eeriness (i.e., the uncanny valley; MacDorman & Ishigoru, 2006; Pollick, 2009; Ramey, 2005). A closely related, but separable model suggests that inconsistencies between how people are expected to look or act, and how a humanlike entity actually looks or acts leads to the feelings of eeriness that constitute the uncanny valley (Brenton, Gillies, Ballin, & Chatting, 2005; Kätsyri, Förger, Mäkäräinen, & Takala, 2015; Pollick 2009). Though the basic form of this argument simply states that there will be unease in response to inconsistencies between expectations and perceptions, different scholars have argued for the importance of different inconsistencies. For example, both Brenton and colleagues (2005) and Pollick (2009) have proposed that inconsistencies between static and dynamic cues may lead to feelings of unease. In a separate argument regarding inconsistencies triggering eeriness, Gray and Wegner (2012) have proposed that a mismatch between the perception that a target has a mind, yet is inanimate, may be a possible trigger for the uncanny valley. In three studies, Gray and Wegner investigated whether the perception that a non-human had a mind could lead to uncanny valley effects. In their first study, participants were shown videos of either a highly humanoid robot or a highly mechanical (i.e., low humanoid) robot and rated how unnerved, uneasy, and creeped out they felt. In addition, they rated the robots on dimensions of mind, including agency (the ability to act on the world) and experience (the ability to feel; Gray, Gray, & Wegner, 2007). Gray and Wegner found that robots (i.e., inanimate agents) believed to have sophisticated human mental faculties were more unsettling than mechanical (i.e., unsophisticated) robots. Further, this was predicted by differences in perceived experience, but not agency robots that can feel are creepier than robots than can think. They also wanted to test whether the ability for inanimate objects to experience the world, even in the absence of humanlike appearance, was sufficient to elicit uneasiness. To this end, Gray and Wegner had participants read about a hypothetical computer that was either high in experience (i.e., could feel hunger and pain), high in agency (i.e., could make decisions), or high in neither, with the prediction that a high experience computer would be creepy even without humanlike appearance. Indeed, they found that the high 3

9 experience computer was rated as more uncanny than either the control or the high agency computer. Finally, Gray and Wegner predicted that a human without experience should be as creepy as a robot with experience. This time, participants read about a human who was either low in agency, low in experience, or normal (i.e., high in both agency and experience). As predicted, when the human was described as lacking in experience, participants perceived that person as creepy and unsettling. Thus, a mismatch between expectations for an agent s ability to experience the world and the agent s actual capacities can lead to uncannyness. In a third example of inconsistencies triggering unease, MacDorman, Green, Ho, and Koch (2009) showed that featural mismatches could also lead to uncanny valley effects. They showed participants humanlike faces that were distorted by changing the texture of the skin, changing the facial proportions, or by changing the size and shape of the eyes. These faces, which contained otherwise humanlike features, were perceived as creepier than their more realistic counterparts. MacDorman and colleagues also found that a single atypical feature is particularly unnerving, especially when all other features are highly humanlike. This suggests that, as predicted by the uncanny valley, stimuli that are highly humanlike, but are not quite human, will be evaluated most negatively. Work by Seyama and Nagayama (2007) has found similar results, showing that when the eyes were either too large compared to the rest of the face or when the realism of the eyes was mismatched with the realism of the rest of the face, the uncanny valley effects were most strongly observed. Taken together, it appears that a variety of inconsistencies, both between expectations and perceptions, and between different features and our conception of what it means to be human, can generate uncanny valley style effects. Drawing from this cue inconsistency hypothesis, I aim to investigate whether the specific processing style with which a target is encoded can influence these feelings of unease. In other words, can the uncanny valley be caused, in part, by a mismatch in processing a stimulus as a human face is typically processed, while that stimulus possesses physical characteristics that are somewhat unlike those of humans? Configural Face Processing as a Perceptual Cue to Humanness Humans possess the remarkable ability to differentiate and recall thousands upon thousands of human faces. This ability is helped in part by the use of configural face processing, a feature integration process that allows us to view a face as a single Gestalt, rather than as a series of separate features (see Maurer, Le Grand, & Mondloch, 2002, for a review). This 4

10 integration is performed almost immediately (typically within 200 ms of processing a face), and occurs for the faces of fellow humans, but for few other stimuli (Tanaka & Farah, 1993). Although several techniques have been used to manipulate or measure configural face processing, the longest standing and most widely used is the face inversion paradigm (Rossion & Gauthier, 2002; Valentine, 1988; Yin, 1969). By inverting the face, one disrupts the eyes-overnose-over-mouth configuration of faces, while retaining their features. Indeed, at the heart of this inversion effect is the reliable observation that inversion dramatically reduces humans ability to process faces, but not other non-face stimuli such as automobiles and houses. Importantly for the current work, configural face processing has been implicated in a variety of phenomena surrounding social judgment. Generally speaking, configural processing is both cause and consequence of a variety of social motives. Perhaps best established is that configural processing can be the result of top-down motives. For example, configural processing is hampered for cross-race targets (Michel, Rossion, Han, Chung, & Caldara, 2006), for mere outgroup members (Hugenberg & Corneille, 2009), and for people who violate social norms (Fincher & Tetlock, 2016). Further, individuals who are deemed worthy of punishment are also not processed configurally (Fincher & Tetlock, 2016). Taken together, it seems that simply believing that a face is a valid target for social affiliation triggers the motive to configurally process. Importantly, configural face processing also has important bottom-up effects as well. Recent research by Hugenberg and colleagues (2016) has demonstrated that processing a face configurally can act as a signal that a target is human, and conversely disrupting this configural processing (such as by inverting a face) can disrupt this bottom-up signal of humanness, causing apparent dehumanization in judgments. Specifically, Hugenberg and colleagues (2016) tested whether configural processing can trigger perceptions of humanness. First, they tested whether configural processing led to the activation of human-related words. Participants were primed with an image of an upright or an inverted human face and then indicated whether a string of letters following the face was either a word or a non-word. Because inverted faces disrupt configural processing, Hugenberg and colleagues hypothesized that only upright but not inverted faces should activate human-related concepts. As expected, human-related words were identified more quickly than machine-related words, but only for upright faces, indicating that human related concepts are activated by upright, but not inverted, faces. In their second study, 5

11 Hugenberg and colleagues investigated whether configural processing made categorizing faces as human easier. They found that inverted, compared to upright, human faces were more slowly categorized as human, but inverted chimpanzee faces were categorized as quickly as their upright counterparts, suggesting that configural processing is used in the categorization of humans, but not other animals. Finally, Hugenberg and colleagues provided evidence that, beyond activation of human concepts and aiding categorization, configural processing actually affects judgments of others humanness. Images of upright faces were rated as more humanlike than were inverted faces. Taken together, it seems that configural face processing can serve as a cue to a target s humanity. Recently, we (Deska, Almaraz, & Hugenberg, in press) replicated and extended these findings using a technique quite similar to that employed in MacDorman and Ishiguro s (2006) pioneering work on the uncanny valley and Looser and Wheatley s (2010) work on mind inferences. Specifically, we employed Looser and Wheatley s (2010) doll-to-human morph continua, in which highly human doll faces were morphed in 10% increments with fully human faces. In our work, however, we also manipulated participants ability to configurally process the faces, via a within-subjects manipulation of face inversion. Thus, participants saw each stimulus both upright and inverted, and rated each face on the extent to which it was animate (i.e., had a mind). Interestingly, inversion only affected face morphs that were rated as more animate than inanimate (i.e., faces beyond the point of subjective equality [PSE]; see Figure 2). Disrupting configural processing for a face that was obviously a doll had no effect one can t dehumanize an obviously non-human stimulus. However, inverting highly human faces, faces that were typically rated as animate, did have the dehumanizing effect previously observed by Hugenberg and colleagues (2016). It is important to note, however, that inversion did not only affect fully human faces (i.e., 100% morphs). In fact, inversion reduced the perception that even 70% morphs had a mind. Although morphs at this level are very humanlike and are processed configurally, neither upright 70% nor even 80% morphs reach the peak levels of humanlike mind afforded to 90% or 100% morphs (see Figure 2). Given that we see inversion effects for highly human, but not 100% human stimuli, this indicates these stimuli are being processed configurally. This combination of configural processing and apparent inhumanity occurs at exactly the location hypothesized in the uncanny valley: just before actual humans. 6

12 If configural processing occurs in the absence of fully human features, as in Deska and colleagues (in press), the perceiver may be receiving mixed signals about the humanity of the target. Whereas a signal of humanness may be activated by processing the stimulus configurally, a signal of inhumanity may be present from features that are highly human, yet not fully human. I propose that when this occurs configural processing in the absence of fully human features it may create a disagreement amongst the signals of humanness, causing trouble interpreting the stimuli and generating an uneasy feeling as a result (i.e., the uncanny valley). Thus, I propose that one reason that highly humanlike, but nonhuman stimuli may be experienced as creepy is the conflicting signals of humanness between the configuration of a face and the features of the face. Current Work The current work aims to investigate the possibility that the uncanny valley is caused in part by a mismatch between facial featural cues of humanness and the cues of humanness from processing style. Put differently, my primary hypothesis is that faces that are processed configurally, but are not fully human (i.e., upright and highly human, yet inhuman faces), will elicit the uncannyness typical of the uncanny valley. A secondary hypothesis, which is also testable with my current design, also falls from the logic of the primary hypothesis: fully human faces that are not processed configurally (i.e., inverted and fully human faces) also send conflicting signals of humanness, and should therefore elicit uncanniness. If observed, this would not be, definitionally speaking, the uncanny valley, but would be evidence that the uncanny valley may actually be a subset of a broader phenomenon produced when signals of humanness in faces are in conflict. By using face inversion as a vehicle for manipulating configural processing, I will test whether the processing style used by perceivers interacts with the features of the face to determine when an image will be seen as creepy. To this end, I will manipulate both the humanness in the features of facial stimuli (morphs ranging from non-human dolls to fully human faces) as well as the orientation of the images (upright versus inverted face presentation). Face inversion is used in this study because it is the most widely used method of manipulating configural processing and because, unlike some other manipulations, it maintains the facial features themselves, and their relationship to one another, changing only the configuration of the features vis-à-vis the viewer. Further, the stimuli remain ecologically valid. Unlike other 7

13 manipulations of configurations, such as scrambling the features locations within the head, upside down human faces are still, logically speaking, human faces. I hypothesize that faces that are processed configurally but that are not fully human will be perceived as creepy, and that inverting these faces will reduce the creepiness. I further hypothesize that inverted fully human faces will be more creepy than their upright counterparts. Experiment The current experiment was designed to determine whether conflicting signals of humanity lead to feelings of unease. To accomplish this, I borrowed from the experimental designs employed by Looser and Wheatley (2010) and Deska and colleagues (in press). Specifically, participants were asked to view faces that lie on a morph continuum between fully doll and fully human in 10 percent increments. They rated each face on several factors related to creepiness (allowing me to index the uncanny valley) and humanness. Critically, half of the faces were presented inverted, which disrupts configural processing (Yin, 1969), and half of the faces were presented upright, which allows for configural processing. Based on my primary hypothesis, I predicted that there would be a range of faces in the close-to-human range of the continuum, which are sufficiently human that configural processing occurs, but whose features are not fully human, that would be more creepy upright than inverted. Based on my secondary hypothesis, I also predicted that fully human faces would seem creepier when presented inverted than when presented upright. Method Participants. A power analysis using G*Power was conducted, assuming a small effect size and 95% power, and suggested 36 as a minimum sample size. As such, I aimed to collect at least 36 participants, and ran participants until the end of the week during which the target N was achieved. Because research has demonstrated that ingroup faces are processed more configurally than outgroup faces (Hugenberg and Corneille, 2009; Michel et al., 2006), and the available morphed stimuli were White, I employed only White participants in this study, to match the race of the target images. Ninety two participants (M age =18.54, SD=0.72) were recruited from the participant pool at Miami University. Most of the participants were female (77.2%) and no participants were excluded from analyses. Materials. The stimuli used in this experiment were the same as those used in previous research (Deska et al., in press; Hackel, Looser, & VanBavel, 2014; Looser & Wheatley, 2010). 8

14 These stimuli are ideal for this work because they allow me to concentrate on the human end of the uncanny valley spectrum the part of the spectrum in which the uncanny valley is likely to occur while systematically varying the objective humanity of the targets. Eight base faces have been morphed with a doll or statue face with similar structure. Each human-doll pair has 11 total faces in the set, varying from 0% to 100% human (100% to 0% doll) in increments of 10%. Finally, each face will also be inverted, making 176 total target stimuli. Procedure. Participants were informed that they would be viewing faces of either humans or humanlike dolls. For each face, they rated how uncanny and how humanlike it appeared. Uncanniness was measured using a four-item composite containing measures of how uncanny creepy, unsettling, and eerie the face appeared. These measurements have been used in past uncanny valley research as measurements of bukimi or eeriness (Carter, Mahler, & Hodgkins, 2013; Gray & Wegner, 2012; MacDorman and Ishigoru 2006; MacDorman et al., 2009), and are aimed at aligning with the disgust, fear, and anxiety that encompass eerieness. Measuring how humanlike each face appears will allow me to replicate Deska et al. (in press), and also find the point of subjective equality (PSE) the point at which stimuli begin being processed in a humanlike manner which will help isolate where I predicted the uncanny valley effects would begin to be observed. Each participant viewed all 11 upright versions of four human-doll pairs and all 11 inverted versions of the remaining four pairs. Following Deska et al. (in press), targets were presented in two blocks, one upright and one inverted. The order of presentation of the blocks as well as the human-doll pairs in each block was counterbalanced between-subjects. Within each block, stimuli were presented in a random order. After rating all stimuli, participants completed a demographics survey and were debriefed before being excused. Results As a preliminary analysis, I first aimed to replicate past work by Looser and Wheatley (2010) and Deska and colleagues (in press) by submitting the single-item humanity rating for each morph to a 2 (target orientation: upright v. inverted) x 11 (morph percentage) repeatedmeasures ANOVA. Replicating Looser and Wheatley, the ANOVA revealed the predicted main effect of morph percentage, F(10,910)= , p<.001, η!! =.926, showing that morphs with objectively 9

15 stronger signals of humanity were perceived as more humanlike. A main effect of target orientation was also revealed, F(10,910)=29.30, p<.001, η!! =.244, such that upright targets were rated as more humanlike than inverted targets. Replicating Deska and colleagues, the predicted interaction between target orientation and morph percentage was also significant, F(10,910)=9.98, p<.001, η! =.10 (See Figure 4). To interpret this interaction, pairwise comparisons were performed, comparing the upright and inverted targets at each morph level. I found that only at and beyond the Point of Subjective Equality (PSE; M=.52) were the upright targets perceived as more humanlike. For morphs from 0% to 40% human, inversion had no effect on perceived human likeness, but for morphs from 50% to 100% human, inversion were rated as less humanlike (50%, t(91)=3.29, p=.001; 60%, t(91)=4.69, p<.001; 70%, t(91)=5.67, p<.001; 80%, t(91)=4.98, p<.001; 90%, t(91)=4.78, p<.001; 100%, t(91)=5.30, p<.001). Of primary interest in the proposed experiment is whether the uncanny valley is observed differently for upright or inverted faces. To investigate this, I first formed an uncannyness composite of the four uncanny valley traits (creepy, uncanny, eerie, unsettling; Cronbach s αs.900), separately at each morph level, split across the upright and inverted conditions. For the primary analysis, I tested whether inversion affects ratings of uncannyness differently depending on how much objective signal of humanity is present (i.e., how far along the human-doll continuum the face is). Again, I hypothesized that uncannyness should follow from configural processing of upright humanlike but non-human faces, and from the failure to configurally process fully human inverted faces. Thus, I hypothesized that, for faces in the uncanny valley, inversion would partially reduce the perception of uncannyness, especially after the PSE (i.e., after which configural processing has been shown to affect processing; see Figure 2), but that uncannyness would increase for inverted, fully human faces (relative to upright, fully human faces). As such, the uncannyness composite data from both inverted and upright targets was submitted to a 2 (target orientation: upright v. inverted) x 11 (morph percentage) repeatedmeasures ANOVA. A main effect of morph level was revealed, F(10,910)=282.60, p<.001, η!! =.756, such that, overall, more dolllike morphs were perceived as more uncanny, and a main effect of target orientation, F(10,910)=29.19, p<.001, η!! =.243, such that inverted targets were more uncanny. Finally, an interaction between morph level and target orientation obtained, F(10,910)=18.48, p<.001, η! =

16 As seen in Figure 5, the most dolllike morphs are all relatively high in creepiness and there are no differences between the upright and inverted versions. As the morphs become more humanlike, they decrease in creepiness overall, but this downward trend is steeper for upright morphs, leading to inverted humanlike morphs being creepier than their upright counterparts. Specifically, pairwise comparisons showed that only for morph levels between 40% and 100% were upright targets rated as less uncanny than inverted targets (notably before the PSE; 40%, t(91)=-2.92, p=.004; 50%, t(91)=-3.68, p<.001; 60%, t(91)=-4.74, p<.001; 70%, t(91)=-6.31, p<.001; 80%, t(91)=-7.74, p<.001; 90%, t(91)=-7.38, p<.001; 100%, t(91)=-7.52, p<.001). Essentially, moving from left to right, targets exit the uncanny valley faster when upright than when inverted. That is, faces become less creepy at an increased rate when upright, compared to when inverted. Though this pattern for highly dolllike stimuli was not hypothesized, the results do reflect the predictions for highly humanlike stimuli (Hypothesis 2). That is, I predicted that the most humanlike morphs would be more uncanny when inverted than when upright. I did find this effect; fully human morphs were more uncanny when inverted than when upright; but it was true for more of the morphs than I anticipated. Toward the dolllike end of the scale, I predicted that the means for upright and inverted would switch from upright being more uncanny to inverted being more uncanny (Hypothesis 1). Although this crossover was not observed, the upright and inverted morphs did meet for the highly dolllike morphs. This pattern of results and the implications are discussed at greater length below. Discussion In the current work, I replicated previous work, showing that both sufficient levels of objective humanity and configural processing were necessary in order for targets to be perceived as fully human. I also found evidence that these factors were also important in perceptions of uncanniness, but not fully in the way that I predicted. I demonstrated that more humanlike morphs were perceived as less uncanny than less humanlike morphs and that morphs with more objective humanity were perceived as more uncanny when inverted than when upright, whereas morphs with less objective humanity did not differ depending on their orientation. First, these findings inform research on the uncanny valley, both in terms of where it lies and what causes it, or more accurately, what may help lead to the escape from the valley. Second, these findings also 11

17 speak to the original question presented in this paper, that the mismatch between signals of humanity may lead to contribute to uncanny valley effects. When considering why the predicted effects for Hypothesis 1 were not observed, but the effects for Hypothesis 2 were, perhaps the most central issue is that this study did not observe a full uncanny valley pattern (see Figure 1). Instead, the results do resemble the latter half of the uncanny valley (see Figure 5). Rather than seeing an increase in creepiness, a peak in creepiness, and a decrease in creepiness (i.e., the pattern typical of the uncanny valley), I observed only a decrease in creepiness across the morph continuum. Thus, it is possible that Hypothesis 1 is still valid for a subset of stimuli that exist outside of the current morph continuum. Indeed, in other research on the uncanny valley (e.g., Ishiguro & MacDorman, 2006), the morph continua extended into much less humanlike stimuli than the current doll stimuli. However, it is clear that manipulations of configural processing for unambiguously dolllike stimuli is not sufficient to modulate the creepiness of those dolllike stimuli. Although it is possible my hypotheses may pan out over a wider range of stimuli, the current data show a pattern that is not consistent with the proposed mismatching hypothesis. Instead of configural processing in the absence of fully humanlike features leading to uncannyness, both objective humanlike features and configural processing contributed to reducing uncannyness. This pattern is more similar to that found by Deska and colleagues (in press) for mind ascription. As such, one reasonable conclusion from the current study is that each signal of humanity is more directly related to creepiness than I had originally proposed. Further investigation, such as that suggested above, would test this possibility. Despite the failure to demonstrate the full uncanny valley, these data do give us information about how configural processing may be useful for exiting the uncanny valley. As can be seen in Figure 5, the apparent creepiness of upright stimuli appears to wane more quickly than does the creepiness of inverted stimuli, as the stimuli morph from dolllike to humanlike. Specifically, the finding that the reduction in creepiness as stimuli became more humanlike was steeper for upright than inverted stimuli shows that configural processing is important to this process. Further, this effect is clearly observed only just before the Point of Subjective Equality of humanness (40% morph). Thus, in terms of a mismatching signals of humanity hypothesis, there appears to be preliminary evidence that this hypothesis may be true, but perhaps not fully as I had initially hypothesized. On the human half of the stimuli, the place where humanlike 12

18 facial configurations are shown to affect processing, we find that the upright stimuli are less creepy than the inverted counterparts. This means with a sufficiently high signal of humanity from the features of the face, configural processing leads to reductions in the perception of uncannyness. In other words, for these stimuli at least, when featural signals of humanity and processing styles typical of human processing co-occur, stimuli are perceived as less creepy, just as predicted. Unfortunately, without the other half of the uncanny valley, it is unclear whether or how the presence of a humanlike configuration in the absence of humanlike facial features will affect uncannyness. The endpoints of the morph continuum were simply not far enough apart. Future Directions and Implications In order to fully understand how these signals of humanity affect the perception of creepiness, it will be important to gain a more full representation of the uncanny valley (i.e., that includes both the upslope and the downslope of the valley) and to see how inversion affects judgments at all points. However, beyond this issue, I believe that the current data are encouraging and suggest several avenues for further research. The current work also adds to existing knowledge of the effects of configural processing. Whereas past research has shown that configural processing is required to be in tandem with sufficient humanlike features if it is to positively affect ratings of stimuli (Deska et al., in press), the current work shows that even before the point of subjective equality, inversion increases the creepiness provided by a stimulus. Thus, sometimes configural processing does not require objective humanity to affect subjective perceptions. This work also helps us understand social phenomena. Specifically, because perceivers only selectively use configural processing (Fincher & Tetlock, 2016; Hugenberg & Corneille, 2009; Michel et al., 2006) these results inform our understanding of how processing style affects interactions even with other humans. Considering configural processing is used to a lesser degree for outgroup, compared to ingroup, faces (Hugenberg & Corneille, 2009) even though they are objectively human, whole groups of people whose features are fully human, but who may not be processed configurally, may be perceived as off-putting or creepy. That is, if mismatches between objective featural humanity and a lack of configural processing lead to feelings of unease, as shown here, this naturalistic difference in configural processing for objectively human faces may inadvertently trigger dislike for outgroup members. Importantly, outgroup members are not the only group of people that are not processed configurally. Past work has shown that 13

19 people with facial stigmas, such as scars, port-wine stains, and birthmarks, afford featural processing, rather than configural processing (Madera & Hebl, 2012). These findings could imply that these groups are perceived as creepy or unsettling, which may help serve as an explanation for the existence of facial stigmas. Future work can aim to investigate the question of whether configural processing is involved in the production of facial stigmas. Conversely, on the perceivers end, some groups of people have difficulty engaging in configural processing for any target. People with prosopagnosia (or face-blindness), for example, are characterized by their inability to use configural face processing, in favor of more featural processing (Farah, Wilson, Drain, & Tanaka, 1995). Importantly, the face-blind also tend to have some difficulty in social situations (Kress & Daum, 2003). Additionally, people on the Autism spectrum, who are relatively unable to understand the thoughts and intentions of others and struggle interacting with others, also do not use configural processing for human faces (Behrmann et al., 2006). If these people perceive others as unsettling or creepy because of the mismatch between processing style and featural signals of humanity, it naturally follows that these disorders would be accompanied by social difficulty. Examining the role of uncanny valley effects in causing social difficulty for people who have difficulty using configural processing has the potential to greatly improve understanding of these disorders, if not help find possible solutions. Another interesting question uncovered by the current work is whether the uncanny valley acts similarly if the target and the perceiver are the same race or social group. Because configural processing is reduced for outgroup members (Hugenberg & Corneille, 2009), it is possible that the uncanny valley could be either reduced or even exacerbated for outgroup members. Recent work has shown that targets that were either believed to be in or actually in an outgroup had a stricter threshold for perceiving humanness (Hackel et al., 2014; Krumhuber, Swiderska, Tsankova, Kamble, & Kappas, 2015). That is, in general more humanlike features were required before targets were perceived as having mind if they were members of an outgroup. Importantly, the tipping point for these outgroup members also proved to be less extreme than for ingroup members. Because the perceptual humanity curves are different for ingroups and outgroups and they provide different levels of configural processing, the uncanny valley may also be different. 14

20 Ultimately, finding that a mismatch between processing style and objective features of a face leads to the uncanny valley would have far-reaching implications, extending beyond robotics to interpersonal interactions. 15

21 References Bar-Cohen, Y., & Breazeal, C. (2003). Biologically inspired intelligent robots. Smart Structures and Materials (pp ). International Society for Optics and Photonics. Bartneck, C., Kulić, D., Croft, E., & Zoghbi, S. (2009). Measurement instruments for the anthropomorphism, animacy, likeability, perceived intelligence, and perceived safety of robots. International Journal of Social Robotics, 1(1), Behrmann, M., Avidan, G., Leonard, G. L., Kimchi, R., Luna, B., Humphreys, K., & Minshew, N. (2006). Configural processing in autism and its relationship to face processing. Neuropsychologia, 44(1), Brenton, H., Gillies, M., Ballin, D., & Chatting, D. (2005). The uncanny valley: Does it exist? In Proceedings of Conference of Human Computer Interaction, Workshop on Human Animated Character Interaction, Napier University, Edinburgh. Burleigh, T. J., Schoenherr, J. R., & Lacroix, G. L. (2013). Does the uncanny valley exist? An empirical test of the relationship between eeriness and the human likeness of digitally created faces. Computers in Human Behavior, 29(3), Carter, E. J., Mahler, M., & Hodgins, J. K. (2013). Unpleasantness of animated characters corresponds to increased viewer attention to faces. In Proceedings of the ACM Symposium on Applied Perception (pp ). Deska, J. C., Almaraz, S. M., & Hugenberg, K. J. (in press). Of mannequins and men: Ascriptions of mind in faces are bounded by perceptual and processing similarities to human faces. Social Psychology and Personality Science. Farah, M. J., Wilson, K. D., Drain, H. M., & Tanaka, J. R. (1995). The inverted face inversion effect in prosopagnosia: Evidence for mandatory, face-specific perceptual mechanisms. Vision Research, 35(14), Fincher, K. M., & Tetlock, P. E. (2016). Perceptual dehumanization of faces is activated by norm violations and facilitates norm enforcement. Journal of Experimental Psychology: General, 145(2), 131. Gray, H. M., Gray, K., & Wegner, D. M. (2007). Dimensions of mind perception. Science, 315(5812), Gray, K., & Wegner, D. M. (2012). Feeling robots and human zombies: Mind perception and the uncanny valley. Cognition, 125(1),

22 Geller, T. (2008). Overcoming the uncanny valley. IEEE Computer Graphics and Applications, 28(4), Hackel, L. M., Looser, C. E., & Van Bavel, J. J. (2014). Group membership alters the threshold for mind perception: The role of social identity, collective identification, and intergroup threat. Journal of Experimental Social Psychology, 52, Ho, C. C., & MacDorman, K. F. (2010). Revisiting the uncanny valley theory: Developing and validating an alternative to the godspeed indices. Computers in Human Behavior, 26(6), Ho., C. C., MacDorman, K. F., & Pramono Z. A. D. (2008). Human emotion and the uncanny valley: A GLM, MDS, and ISOMAP analysis of robot video ratings. Proceedings of the third ACM/IEEE International Conference on Human-Robot Interaction Hugenberg, K., & Corneille, O. (2009). Holistic processing is tuned for in-group faces. Cognitive Science, 33(6), Hugenberg, K., Young, S., Rydell, R. J., Almaraz, S. M., Stanko, K. A., See, P. E., & Wilson, J. P. (2016). The face of humanity: Configural face processing influences ascriptions of humanness. Social Psychological and Personality Science, 7(2), Kätsyri, J., Förger, K., Mäkäräinen, M., & Takala, T. (2015). A review of empirical evidence on different uncanny valley hypotheses: Support for perceptual mismatch as one road to the valley of eeriness. Frontiers in Psychology, 6 (390), Kress, T., & Daum, I. (2003). Developmental prosopagnosia: A review. Behavioural Neurology, 14(3-4), Krumhuber, E. G., Swiderska, A., Tsankova, E., Kamble, S. V., & Kappas, A. (2015). Real or artificial? Intergroup biases in mind perception in a cross-cultural perspective. PloS One, 10(9), e Looser, C. E., & Wheatley, T. (2010). The tipping point of animacy: How, when, and where we perceive life in a face. Psychological Science, 21, MacDorman, K. F., Green, R. D., Ho, C. C., & Koch, C. T. (2009). Too real for comfort? Uncanny responses to computer generated faces. Computers in Human Behavior, 25(3), MacDorman, K. F., & Ishiguro, H. (2006). The uncanny advantage of using androids in cognitive and social science research. Interaction Studies, 7(3),

23 Madera, J. M., & Hebl, M. R. (2012). Discrimination against facially stigmatized applicants in interviews: An eye-tracking and face-to-face investigation. Journal of Applied Psychology, 97(2), 317. Mathur, M. B., & Reichling, D. B. (2016). Navigating a social world with robot partners: A quantitative cartography of the uncanny valley. Cognition, 146, Matsui, D., Minato, T., MacDorman, K. F., & Ishiguro, H. (2005). Generating natural motion in an android by mapping human motion. In 2005 IEEE/RSJ International Conference on Intelligent Robots and Systems (pp ). Maurer, D., Le Grand, R., & Mondloch, C. J. (2002). The many faces of configural processing. Trends in Cognitive Sciences, 6, Michel, C., Rossion, B., Han, J., Chung, C. S., & Caldara, R. (2006). Holistic processing is finely tuned for faces of one's own race. Psychological Science,17(7), Minato, T., Shimada, M., Ishiguro, H., & Itakura, S. (2004, May). Development of an android robot for studying human-robot interaction. In International Conference on Industrial, Engineering and Other Applications of Applied Intelligent Systems (pp ). Mori, M. (1970). Bukimi no tani [The uncanny valley]. Energy, 7: Pollick, F. E. (2009). In search of the uncanny valley. In International Conference on User Centric Media (pp ). Springer Berlin Heidelberg. Ramey, C. H. (2005). The uncanny valley of similarities concerning abortion, baldness, heaps of sand, and humanlike robots. In Proceedings of Views of the Uncanny Valley Workshop: IEEE-RAS International Conference on Humanoid Robots (pp. 8-13). Rossion, B., & Gauthier, I. (2002). How does the brain process upright and inverted faces?. Behavioral and Cognitive Neuroscience Reviews, 1(1), Seyama, J. I., & Nagayama, R. S. (2007). The uncanny valley: Effect of realism on the impression of artificial human faces. Presence, 16(4), Starkey, S. (Producer), Zemeckis, R. (Producer & Director), Goetzman, G. (Producer), & Teitler, W. (Producer). (2004). Polar Express [Motion Picture]. United States: Warner Bros. Pictures. Tanaka, J. W., & Farah, M. J. (1993). Parts and wholes in face recognition. The Quarterly Journal of Experimental Psychology, 46(2),

24 Valentine, T. (1988). Upside down faces: A review of the effect of inversion upon face recognition. British Journal of Psychology, 79(4), Yin, R. K. (1969). Looking at upside-down faces. Journal of Experimental Psychology, 81, Zemeckis, R. (Producer & Director), Bing, S. (Producer), Rapke, J. (Producer), & Starkey, S. (Producer). (2007). Beowulf [Motion Picture]. United States: Paramount Pictures. 19

25 Figure 1: The uncanny valley is the sharp decrease in familiarity or liking as an inanimate stimulus becomes perceptually similar to a human stimulus. 20

26 Figure 2: The point of subjective equality (PSE) indicates the point at which participants were equally to perceive the faces as having or not having mind. Only beyond the PSE did we find that inversion affected mind ascriptions (Deska et al., in press). Figure 3. Example stimuli of two human-to-doll stimulus morphs, ranging from 100% human (left) to 100% doll (right), one presented upright (top) and one inverted (bottom). 21

27 Figure 4. The interaction between morph percentage and target orientation on perceptions of human likeness. Replicating past work, more humanlike morphs are seen as more humanlike, and there is no difference between upright and inverted stimuli of the dolllike end of the spectrum, but for humanlike morphs, upright stimuli appeared more humanlike than inverted stimuli (PSE indicated the Point of Subjective Equality. *p<.05). 22

28 Figure 5. The interaction between morph percentage and target orientation on perceptions of uncannyness. Overall, uncanniness decreased as morphs became more human. Importantly, there is no difference between upright and inverted stimuli of the dolllike end of the spectrum, but upright humanlike stimuli appeared less uncanny than did inverted humanlike stimuli (PSE indicated the Point of Subjective Equality. *p<.05). 23