Two “Uncanny Canyons”

There is some controversy about how human-like robots should be, which arises especially in two new applications of robotics that are attracting a lot of attention (especially in Japan): playmates/baby-sitters for young children and companions/caretakers for the elderly.  (These address two important problems in contemporary society: daycare for children and eldercare.  Do you think it’s a good idea?)

You have already seen how programs as crude (and unintelligent) as Eliza could encourage some people to confide and discuss their problems.  Might people put too much trust in robots that look and act much more like humans?  (To see an example of the state of the art of human-looking robots, look at “Japanese Android Video” <androidvideo.com>. Note that, unlike Leonardo, this robot [named Repliee Q1expo] is not interacting with the people; “she” is just speaking prerecorded speech.)

Roboticists speak of an “uncanny canyon” in people’s responses to robot appearance and behavior.  In general, people prefer (and better understand) robots that look more like humans, but robots that are very similar, but not identical, to people in their appearance and behavior are very disturbing (“uncanny”).  The reason seems to be that we have several hundred thousand years’ of experience in sensitively judging people by their appearance and behavior; this is and has been very important in our social interactions.  (For example, people may look or act “creepy,” or suspicious, or trustworthy, or approachable, etc.)  If a robot looks sufficiently like a human, then these brain systems are engaged, and if it then falls short of comfortable human behavior, then the warning signals go off.  We have lower expectations for robots that are less humanoid (e.g., Leonardo), because these highly tuned brain systems are not engaged by them. (Similarly, for example, we are less sensitive to the mental states of chimps than to those of humans, and we are not bothered when chimps don’t act like humans.)

A second “uncanny canyon” occurs in the way people of different ages respond to very human-like androids (such as Repliee).  Children younger than 3 or 4 years do not find her disturbing, nor do adults over about 20. But between these ages is an “uncanny canyon” in which they find the android creepy.  The reason seems to be that beginning about 3 or 4, children are forming a detailed and subtle cognitive model of how humans normally look and behave, an idea of humanness we might say, and that the imperfect humanness of Repliee clashes with their developing expectations.  (By the time they are adults, their idea of humanness is more secure.)

So on the one hand, androids might be disturbing if they are not completely human-like, but if their behavior seems authentically human, it might encourage people to treat them as humans (relying on them too much, seeking empathy from them, or psychological advice, for example).  On the other hand, as we have discussed, in human-robot interaction, it is important for people to be able to “read” the internal (“mental”) states of robots, and this is facilitated by greater human-like appearance and behavior.

So here is an issue to think about and discuss: How human-like should robots be?  This is especially an issue for contemporary robots, which are far below humans in intelligence, do not have emotions, etc.


See Scientific American Mind (June/July 2006) for a discussion of this android and the two “uncanny canyons.”


Return to Minds & Machines home page

Return to MacLennan's home page

Send mail to Bruce MacLennan / MacLennan@cs.utk.edu

Valid HTML 4.01! This page is www.cs.utk.edu/~mclennan/UH267/UncannyCanyons.html
Last updated: 2006-09-04.