This Creepy Face Perfectly Explains the Uncanny Valley
A viral image captures that creepy feeling we get when something is humanlike but not quite right.
Something weird happens if you put “average” faces together—they get more attractive. Some research suggests that this may be because we find symmetrical faces more attractive, and a crooked nose on one face and a droopy eye on another get washed out by other noses and eyes making an overall face more free from flaws.
Other research, however, suggests a different possibility. Broadly speaking, the easier something is to process—think a really legible font or a saying that rhymes—the more we tend to like it, and average-looking faces may be more attractive because they’re easier to process. With their edges smoothed out, these faces fit more snugly into the mental template we use to recognize faces.
In an image that went viral last month, New York-based artist Robby Kraft gave us a small glimpse into what that template looks like, and part of what’s so striking about it is how really, really creepy it is. It’s ghostly, looking vaguely like Jeff Daniels, with a clearly defined set of lips, eyes, a nose, and cheekbones. But what is it, exactly, that makes the face look so unsettling? Recent research suggests it may be the same reason we find dolls and cheap CGI so unnerving: the uncanny valley.
Coined by Japanese roboticist Masahiro Mori in the 1970s, the uncanny valley describes the cold and eerie feeling we experience when something seems not quite human. Short Circuit’s crude robot Johnny Five is charming and endearing, but the plastic characters from The Polar Express are weirdly creepy in a way that more modern CGI isn’t. If something is obviously not human or convincingly human, then there’s no problem. There’s a gap between the two, however, where something’s just off, and that makes us strangely uncomfortable—which is certainly the case in Kraft’s image.
Compiling 2,500 photos from Instagram under the #FacesInThings tag, Kraft used code to scale and align each object by their “eyes” and “mouths,” laying them one over the other with equal transparency. The resulting image is a ghostly and surprisingly realistic composite of a face. This isn’t a perfect reconstruction of how our minds represent faces, however, but an approximation. Kraft explained to me in an email that the final image does show signs of some of the underlying algorithm and the choices he made.
“After I face-detected all the images, I averaged the vertical distance between the eyes and the mouth and scaled all the images in the vertical to match this average,” Kraft told The Daily Beast. “So that value was entirely human-decided and had nothing to do with the algorithm.”
Nonetheless, the brunt of the work is done by the sorts of things we see faces in—an imprint of our minds at work, shown in found objects—and there’s an undeniable pattern that looks startlingly lifelike.
Strangely, however, the face may be so startling and unsettling because it’s in a sweet spot of lifelikeness. It wouldn’t be nearly so creepy if it looked more crude, nor if it looked indistinguishable from a regular face. Which is to say that it’s a great illustration of the uncanny valley.
Accounts diverge in explaining why exactly the uncanny valley exists, and the debates quickly grow technical and theoretical. Mori, who coined the term, along with other more contemporary researchers, argue that the uncanny valley is a survival instinct, protecting us from things like dead bodies or from unfit mates who aren’t quite human, like Neanderthals.
Others suggest that it’s the fact that we are unsure whether or not these faces are human, and it’s that uncertainty that we find unnerving. We like to put things into boxes, the reasoning goes, and that inability, particularly when it comes to people, seems threatening.
Other researchers, still, suggest that it may be the opposite effect of why we like averaged faces—when faces don’t quite fit into the mental template we have, that unexpected difference can be jarring. Particularly so, since some parts of these faces are realistic (the defined cheekbones, lips, and nose of Kraft’s composite) while others are more obviously fake (the hollow eyes, missing ears, and long, dissolving forehead).
A forthcoming paper in the journal Cognition sets out to test two of these competing explanations—do we experience the uncanny valley because we’re unsure of what we’re looking at? Or is it because the incongruity between the real and the fake makes us uncomfortable?
The paper’s authors, Karl F. MacDorman and Debaleena Chattopadhyay from the Indiana University, presented more than 500 subjects with images of people, animals, and objects, all of which varied in how computer-generated they looked. Participants rated how eerie each image felt and whether it was real or fake, and the results were interesting. First, there was no uncanny valley effect for objects, which suggests that it’s limited to things that are alive, like people or animals. And second, the images that were most ambiguous—that is, the images where people were more or less at chance guessing which was real—weren’t rated to be particularly eerie at all. They definitely weren’t the eeriest.
The faker-looking images, however, were consistently rated to be eerier and less warm than the real and ambiguous-looking ones. This effect was especially pronounced when some parts of the face, like the eyes or the mouth, looked less real than other parts. This isn’t particularly surprising since those regions of the face convey the most social information and are critical to understanding the intentions and goals of others—particularly for figuring out whether someone means us harm.
While discussions about the uncanny valley may seem cute and abstract—so some faces look creepier sometimes, so what?—the stakes get higher as our world becomes increasingly automated and machines take on a greater role in our lives. Another forthcoming study, also published in Cognition, suggests that the creepier robot faces seemed, the less trust we were willing to put in them. In fact, the trust showed a similar “valley” effect: more-crude robots were trusted just fine, and so were hyper-realistic ones. The problem lies in between.
So while the uncanny valley may seem like a fun novelty when we’re talking about old CGI movies, it takes on a new level of importance once robots start doing things like prescribing us medicine or taking care of our kids. And that’s an important thing for researchers and designers to keep in mind—perhaps our future robot helpers will seem less like human-replications and more like something distinct and friendly.