Uncanny but not confusing: multisite study of perceptual category confusion in the Uncanny Valley

Android robots that resemble humans closely, but not perfectly, can provoke negative feelings of dislike and eeriness in humans (the “Uncanny Valley” effect). We investigated whether category confusion between the perceptual categories of “robot” and “human” contributes to Uncanny Valley aversion. Using a novel, validated corpus of 182 images of real robot and human faces, we precisely estimated the shape of the Uncanny Valley and the location of the perceived robot/human boundary. To implicitly measure confusion, we tracked 358 participants’ mouse trajectories as they categorized the faces. We observed a clear Uncanny Valley, though with some interesting differences from standard theoretical predictions; the initial apex of likability for highly mechanical robots indicated that these robots were still moderately dislikable, and the Uncanny Valley itself was positioned closer to the mechanical than to the human-like end of the spectrum. We also observed a pattern of categorization suggesting that humans do perceive a categorical robot/human boundary. Yet in contrast to predictions of the category confusion mechanism hypothesis, the locations of the Uncanny Valley and of the category boundary did not coincide, and mediation analyses further failed to support a mechanistic role of category confusion. These results suggest category confusion does not explain the Uncanny Valley effect.