In the research field of human-agent interaction, it is a crucial issue to clarify the effect of agent appearances on human impressions. The uncanny valley is one crucial topic. We hypothesize that people can perceive a humanlike agent as human at an earlier stage in interaction even if they finally notice it as non-human and such contradictory perceptions are related to the uncanny valley. We conducted an experiment where participants were asked to judge whether faces presented on a PC monitor were human or not. The faces were a doll, a CG-modeled human image fairly similar to real human, an android robot, another image highly similar and a person. Eyes of the participants were recorded during watching the faces and changes in observing the faces were studied. The results indicate that eye data did not initially differ between the person and CG fairly similar, whereas differences emerged after several seconds. We then proposed a model of the uncanny valley based on dual pathway of emotion.