Uncanny Valleys

by Michael C. Dorf

In my latest Verdict column, I discuss the decision by the US and various other countries to recognize Juan Guaidó as the legitimate president of Venezuela. I frame my discussion around the question whether, and if so why, it is appropriate for outside states to deny recognition to Nicolás Maduro on the ground that his election was tainted when such states continue to recognize authoritarian leaders of other countries (I name Saudi Arabia and North Korea) who have no democratic legitimacy.

I offer two answers. First, I note that there are practical/prudential reasons why democratic regimes recognize undemocratic ones. Second, I argue that there is something particularly bad about subverting a democratic regime. To make that second point, however, I need to address the fact that some wholly undemocratic regimes have "sham constitutions" that purport to be democratic. While hardly approving of sham constitutionalism, I nonetheless acknowledge that in most cases (as in North Korea), it is easy enough to see that the constitution was never meant to be taken seriously.

To illustrate the possibility that it could be worse--at least along one dimension--for a leader to subvert his country's genuine democratic norm than for a leader to rule autocratically without the serious pretense of democratic legitimacy, I draw on the concept of an "uncanny valley," familiar from the study of robotics. A robot that looks nothing like a human is not creepy. Neither is an actual human. But a robot who looks pretty close but not quite close enough to a human is creepy. It falls within the uncanny valley. I suggest an analogy for systems of government.

Here I want to ask whether we can find other uncanny valleys. Depending on how loosely one defines the concept, my guess is that we can find a great many.

As in my jump from feeling creeped out by humanoid robots to being especially concerned about undercutting democracy, both the independent and the dependent variable can be almost anything. The idea is that the x-axis measures how close something is to some ideal, while the y-axis is some normative measure. An uncanny valley occurs when approaching the ideal causes first a dip and then a rise in the y-value. To the left is a graph for robots to human likeness, but as I'm using the uncanny valley concept, we could have other variables as well.

Here are three other uncanny valley candidates:

(1) x-axis = team quality; y-axis = fan satisfaction. I'm a New York Knicks fan. For the last several years, the Knicks have been terrible. As a result, I don't expect them to win, so I rarely watch them. If they were to get a little better, my satisfaction as a fan would improve as they would occasionally surprise me by winning. If they were to get a lot better but not quite championship-caliber, my satisfaction would then decline, because I would have false hope that would be invariably frustrated. If they were to get a whole lot better to the point where they won some championships, my satisfaction would peak. Obviously, not everyone feels this way about teams they follow, but I'm guessing that enough fans do that for some substantial number, the relationship I've just described marks an uncanny valley.

(2) x-axis = how close a cover sounds to a terrific original; y-axis = listener enjoyment. Suppose you really love some song. You listen to a totally different song. Your reaction is meh. Then you listen to a cover of the song you really like. The cover is innovative and interesting. You don't like the cover as much as you like the original, but you like the cover. For me, a good example is the Manfred Mann cover of Bruce Springsteen's Blinded by the Light. I prefer the Springsteen version, but I like the Mann version too. However, I think I would experience an uncanny valley if I were to listen to a Springsteen impersonator sing like Springsteen. I would like that version less than either the Mann or the Springsteen version, even though it would be in some sense closer to the Springsteen version that I prefer to the Mann version. Generalizing, a good cover is not simply a close but not perfect imitation of the original; it's got to be different enough to justify recording the cover. Meanwhile, a great cover is actually better than the original. (For me, Jimi Hendrix's cover of Bob Dylan's All Along the Watchtower is the gold standard here.)

(3) x-axis = how close a job candidate comes to getting hired at a particular school the first time he or she applies to be a law professor there; y-axis = likelihood of being hired by that school. I occasionally advise former students who are seeking entry-level positions as law professors. The process has three main stages: (a) submitting written material via the Association of American Law Schools (AALS); (b) attending the AALS hiring conference at which prospective employer schools send hiring teams to conduct 1/2-hour screening interviews; (c) giving a "job talk" and attending office interviews on-campus. A candidate whose written material doesn't make the cut is eliminated at stage (a). One who successfully navigates all three stages receives a job offer. That's the right-hand peak of the curve. In between are candidates who get eliminated at stages (b) and (c). A candidate who does not end up with any satisfactory job offers will often try again in the following year. If so, that candidate will likely be rejected at stage (a) from any school at which she progressed to stage (b) or (c) in the prior year, but if she was rejected by a school at stage (a) in the prior year, she might make it all the way through to an offer from that school. Why? Because the school's hiring committee will likely have no recollection of a candidate rejected at stage (a) or is using different hiring criteria in the next year (looking for a torts teacher, say, after a retirement), but a candidate who came close but didn't get the cigar will be subject to a kind of res judicata. In other words, the curve will include an uncanny valley.

To be sure, not all hiring works in this way. To give a famous example, President Clinton interviewed both then-Judge Ginsburg and then-Judge Breyer for his first Supreme Court appointment. He gave it to Ginsburg, reportedly because Breyer did not connect with Clinton during the interview. (If so, that may have been because Breyer was still recovering from a bicycling accident.) If the AALS uncanny valley curve applied to this kind of hiring, then Breyer should have been largely out of the running for Clinton's next appointment, but as we know, Clinton tapped Breyer. In many contexts, being runner up or coming close the first time around puts one in good stead for the next time around. Whether any particular graph exhibits an uncanny valley is ultimately an empirical question.

I invite readers to propose other uncanny valleys. Note, however, that not every curve that goes up, then down, then up again includes an uncanny valley. There must be some way in which the close-but-no-cigar phenomenon contributes to the effect. For example, suppose we were to plot on the x-axis the time since an addict takes a drug and on the y-axis the addict's sense of physical well-being. Well-being would go up initially as the drug travels through the body. Then, as the drug wears off, well-being would go down. It would then plummet as the addict goes into withdrawal. Well-being would then go up as the now-former addict had kicked the habit. The curve would have the right overall shape, but that would not be a consequence of the close-but-no-cigar phenomenon.