The idea is that when you face a thing that resembles a person, it becomes more and more likable as it becomes more human-alike, except that between the “almost human” and the “perfectly human” lies a point where you actually dislike the thing for it’s resembled humanity. The interpretation is that computer games still need to become more realistic if 3d-CG characters are ever to become as likable as Mario.
But there is a fatal flaw in the hypothesis: it sprung from the experience of building robots progressively more human. That means that the data in the “small resemblance” side of the graph is much more reliable than the data in the “resembles a lot” side. In other words, the first peak and following decrease in likability is reliable, but the further ascension is not. The first side of the graph comes from the robot builder’s experience, the later from extrapolations, analogies (and i suspect a bit of science-fiction).
My guess is that we stop liking anthropomorphic things when we begin to think they are trying to fool us. If things are clearly just things, having anthropomorphic qualities is a plus. But if things start to try too hard to look like people it becomes a person but with flaws. The two ways of “dealing with”: dealing with things and dealing with people cannot be forced into the same graph, and that is the reason for a “valley” there.
The answer to creating visually indistinguishable from human CG characters was found years ago when Lucas found out that the only way to make digital film believable was to ADD NOISE so that it wouldn’t mismatch with the rest of the footage. Simply put, diminish the level of detail. Humans can recognize other humans even with very bad resolution, the thorn of 3d CG is exactly that it has too much resolution, human beings do not have this much.
But if i am right, making each CG character indistinguishable from human will not solve the problem. They will still be “fake humans”, and if the difference becomes harder to spot, the prejudice against “fake humans” will become stronger, just as the fear of being fooled is also bigger.
I think what is needed to enhance the likability of CG characters is to create clear limits to their simulating attempt. To find some specific, distinguishable, obvious and rigidly enforced ways in which they will not be and not attempt to be human AT ALL.
But as i am not involved in any CG, my theories will likely remain unproved.