“I’ll be back.” This now-famous movie line from Arnold Schwarzenegger placed the Terminator film series firmly in the minds of filmgoers around the world. In the films, Schwarzenegger plays a robot sent back in time to protect a family from other robots bent on eliminating it. But the Terminator films, focused on the frightening plot of people pursued endlessly by single-minded human-looking assassin androids, reveals something interesting about the human psyche. More than just the story line of the explicit threats posed by killer robots, the films touched on a more subtle, but far more profound paradox of the human mind – our preference for agents that appear more human, but fear and revulsion at non-human agents that look “too human.”
Masahiro Mori (1970), a Japanese robotics engineer, noticed that as his robotic creations looked more and more like humans, they seemed more aesthetically pleasing and popular. However, once these robots reached higher levels of human likeness, they elicited strong feelings of discomfort from those who interacted with it. Mori (1970) used these experiences to theorize a pattern whereby, as human likeness of an object increased people were more “familiar” with it, until some point around 70 percent, when “familiarity” would dip into negative values and people would express a strong revulsion to the object until it reached levels of 100 percent human likeness. This curve, reprinted as Figure 2, formed the basis for what Mori (1970) described as the Uncanny Valley Theory (UVT), where the drop in “familiarity” represents the valley created by the uncanniness of the object. While other researchers have worked to translate “familiarity” into more understandable terms such as “eeriness” (MacDorman, 2008), the existence of the valley remains the same.
While the theory remains under-researched, a few scientists have attempted to explain this phenomenon. The leading researcher in the area, Karl F. MacDorman, introduced his theory of Mortality Salience (2005), which explains that very human-like stimuli that do not share enough characteristics with a living human evoke thoughts of our own mortality, triggering defense mechanisms such as revulsion. Victoria Groom and her team (Groom et al, 2009) used previous social research by Fiske and Taylor (1991) to annunciate a Consistency Theory where humans prefer to interact with agents whose behavior is consistent and can be predicted. Those agents, such as human-like robots, with which humans are unfamiliar and whose behavior is not readily predictable, create a nonpreferential reaction in people.
MacDorman’s studies have been key to the advancement of the UVT, but additional researchers are taking an interest in the subject. The most interesting of these came from the marketing field, when Oyedele, Hong and Minor (2007) looked into cross-cultural and contextual effects of the UVT. Oyedele et al studied South Korean and American adults who had self-reported as either low or high for a phobia of technology. They took this a step further by asking respondents about their comfort levels with interacting with robots of different human likenesses in varying contexts, ranging from simply touching the robot, to watching a movie with it, or even living with it. They’re surprising findings showed that Korean respondents, whose society encourages advanced technology, were more wary androids than their U.S. counterparts.
We decided to add to the existing research by looking, not at cultural differences, but gender differences. The work of Grossman and Wood (1993) on the differences in emotional intensities of men and women, along with findings that the uncanny valley effect is more closely related to emotions than the preferential design of the Consistency Theory (MacDorman, 2008), led us to consider this study. We hypothesize that, given images of robots varying along the human-likeness continuum, men and women will feel about the same overall comfort level. However, looking at comfort levels for each image, we hypothesize that women’s stronger emotional intensity as compared to men will make them more sensitive to the robot images, both in the positive direction leading up to the uncanny valley, and the negative direction in the uncanny valley. Our experiment employed a 2×6 design to test the interaction effect of gender and human likeness on comfort levels and presented male and female participants with ten images of robots of varying human likeness along with a scale recording their ratings.
Fifty-eight men and women (29 men and 29 women) participated in the study. Participants were in the age range of 20 – 53 years old and varied in their occupation and geographic location along the United States eastern coast states of South Carolina to Connecticut.
Materials and Design
Researchers set up an experimental website that housed instructions, the ten robot images, response forms, and an open-ended question about participants’ reasons for low comfort ratings. The images used were accessed from previous research (MacDorman, 2008) and the website www.androidworld.com, following the Oyedele et al study (2007). We selected ten robot images along the human likeness scale, using the ratings either provided from the MacDorman study (2008) or along the lines of Mori’s work (1970), and their human likeness levels were ranked as follows: 20, 40, 60, 70, 70, 70, 80, 90, 90 and 90. These robot images, duplicated in Appendix A, were standardized in black in white and for size to limit any confounding effects of color reproduction or size.
We used a 2×6 mixed factorial design with sex (male, female) as a between-subject variable and human likeness percentage (20, 40, 60, 70, 80, 90) as a within-subject variable. Higher human likeness percentages meant that the robots were rated as more human-like; lower percentages meant that the robots were rated as less human-like. Our dependent variable was level of comfort, as determined by a scaled question. The scaled question was, “Please rate how comfortable or uncomfortable the subject in the image would make you feel.” The scale ranged from “1 – Very Uncomfortable,” to “10 – Very Comfortable,” and was purposely context-free.
The researchers contacted participants via email addresses with an invitation to participate in the experiment, informed consent, and a live link to the experimental website. This email also provided an original and randomly-generated order in which participants were to access and rate the images on the website.
Upon accessing the website, participants were presented again with the informed consent information and instructed to follow the appropriate links to access first the instructions, then the images, and finally an open-ended question about the reasons for their ratings. The instructions informed participants that they were to be shown ten images and that the study would be looking at their reactions to these images. They would be shown one image at a time – in the order designated by the randomly-generated sequence in their invitation email – and that they would indicate their age, gender and response to the scaled comfort question on their reaction to the image.
Once participants completed the image ratings, they were instructed to fill out a final question about why they gave any image a rating of 1-5 if they had done so. The results of this question are presented in Appendix B. Participants were then informed that they had completed the study and that they should close their internet browsers to exit the experiment.
Table 1 shows the mean ratings for the images by human likeness for both men and women. The table shows relatively equivalent variances among the ratings for each picture on the ten-point scale – a standard deviation of around two points. This variation only shifted when the images reached the 90 percent human likeness level, with women’s deviation remaining around two, but men’s ratings deviating by two and a half points. Table 1 also shows that the overall mean ratings ranged from around 3.5 to 6.5, illustrating a common trend of ranking the images from the “slightly uncomfortable” to “neutral” to “slightly comfortable” zone. Out of the 30 subjects in our study, 28 of them responded in the predicted direction: high comfort ratings for low human likeness robots, decreasing comfort ratings for increasing human likeness up to around 70, then increasing comfort ratings as human likeness increased to 90.
Figure 1 illustrates the mean ratings for both male and female participants. This plot shows that both men and women followed a pattern similar to that proposed by the UVT – their ratings increase with the human likeness of the robot images, but takes a sharp drop at the 70 percent human likeness level, only to climb above neutral ratings again by the images qualifying as 90 percent human likeness. Figure 1 also allows us to investigate the hypothesized main effects: the main effect of gender, the main effect of image human likeness, and the interaction effect of gender and image human likeness. We can see that there is little difference between the genders’ overall ratings, an obvious difference between the levels of image human likeness, and data that trend towards an interaction effect at 60 and 90 percent human likeness.
We followed this initial plot with a repeated-measures two-way ANOVA with an alpha level of 0.05. The analysis yielded an insignificant statistic for a gender main effect, F(1,28) = 0.19, MSE = 68.83, p=0.892. Males (M=5.07) and females (M=5.24) felt the same average level of comfort for the set of images. There was a significant main effect of image human likeness, F(5,140) = 55.72, MSE = 184.11, p = 0.003. Subjects gave different ratings of comfort for each of the levels of image human likeness (20%, M=5.86; 40%, M=2.73; 60%, M=4.04; 70%, M=6.09; 80%, M=6.34; 90%, M=4.88). The interaction of gender and image human likeness approached significance, F(5,140) = 2.04, MSE = 77.90, p=0.55. Males and females experienced the uncanny valley effect similarly. As this study concerned itself with the interaction effect, we concluded our inferential analyses with the two-way ANOVA and will move on to discuss our findings in the discussion here.