"The Uncanny Valley" is a translation of the Japanese name for a phenomenon noted by Japanese roboticist Masahiro Mori
in the 1970s. Mori was concerned with human reaction to the presence and appearance of artificial creatures, and studied the reactions of many test subjects to creatures of varying degrees of anthropomorphicity
(izzat a word? Heh). He graphed their responses to both the creature's or device's movement and to their appearance on a two-dimensional graph, with the degree of positive or negative reaction on the part of the human on the Y-axis (the dependent variable
) and the similarity of each quality to that of a human along the X-axis (independent variable
The results surprised him. Instead of a roughly linear or even smoothly curved line indicating the human reaction becoming increasingly positive as the test object approached entirely human-like movement or appearance, there was instead a slope leading up the expected path for the first two-thirds of the graph. However, in the upper values of similarity, there was a sharp drop in response, all the way down into the negative (repulsion) in some cases, before an equally sharp rise upwards as the similarity grew past some particular point.
I'm lousy at ASCII art, but I'll give it a try and illustrate a notional graph:
E | A o
C | o o
T | o o
I | o o
O | o o
N | o o
| o o o
| o <--- The 'Uncanny Valley'
| o o o
| o o
Inhuman (Robotic) <------------> Human
On the above graph, the point labelled 'B' is the depths of the Uncanny Valley, where peoples' response to a simulacrum or robot is the least favorable. Point 'C' indicates a fully human-appearing object; responses to this are the most favorable.
Why the sudden dropoff? Masahiro postulated that initially, as the object of our attention is recognizably not only not human but not really even close to human in appearance, we tend to disregard 'nonhuman' appearance traits, and to anthropomorphize by attributing humanoid motives or reasoning to the object's actions. Pets, for example, look nothing like a person, so any recognizably non-human behaviors they exhibit - or, in the case of Hiro's two variables, non-human movement and non-human appearance - is ignored or interpreted as 'cute' or 'interesting' or other non-person to person cues.
However, at a certain point, as the object grows more human, we tend to stop cataloging the similarities and begin concentrating on the dissimilarities. As a result, highly-humaniform (but not perfect) renderings of people suffer because their motion, for example, looks wrong to people. Similarly, eyes lack some 'glitter', or facial muscles aren't present - these things become central to our perceptions of the object, and as a result, we think of them as 'passing for' human. The corpse, for example, looks almost human - but the color is wrong, there is no animation of the skin or muscles - and people can instantly recognize that there is 'something wrong with this picture.'
Hiro's findings were, of course, intended to aid in producing robots that could be easily accepted by people. His counterintuitive finding was that designers should avoid making their creations too lifelike, for they risked falling into the Uncanny Valley. Rather, they should strive for point 'A' on the graph - the last peak before the trough. This may be one reason the 'pseudohuman' characters of Japanese videogames and anime are more highly regarded than the fetishistically rendered ones of first person shooters - the pseudohuman characters fall on the graph to the left of the valley, whereas the almost-but-not-quite renderings of the FPS genre manage only to improve on the appearance of the model enough to push it to, perhaps, point 'B'.