Proudfoot, Diane - Rethinking Turning's Test

11/15/2013

The Journal of Philosophy, Vol 110 No 7, July 2013

Author starts by reviewing the history of Alan Turing's "Computing Machinery and Intelligence" paper in Mind. The central question is whether a computer could "be said to think", but it was replaced with the question: "are there imaginable digital computers which would do well in the imitation game". For Turing, the criterion for thinking morphed into what could be taken for an imitation of human activity, and author sets to give a fresh interpretation of the test as a "response-dependence" approach (pg393).

First author reviews the "canonical" view, which employs behaviorism as the criterion: if the machine behaves as a thinking creature would in a given context, it is thinking. Author argues that Turing did not mean to give even sufficient (let alone necessary) conditions for thinking using behaviorism. The interrogator's (the human's) response to the behavior is a necessary part to success in the imitation game. Turing used a different game, one of a man trying to fool an interrogator into believing he was a woman, as the benchmark for how well a computer could fool an interrogator into believing it was thinking. Author discusses how puzzling this emphasis on the interrogator's response is for behaviorists. Instead of writing off this crucial element as misguided, author suggests that the behavoristic interpretation is what is misguided.

The new interpretation author puts forward starts with Turing considering the "idea of 'intelligence'" (pg396) as an "emotional" one, but meaning specifically that we apply the label "intelligent" partly due to our own mental state and perspective; it is response-dependent (pg397). The trick now is to indicate what kinds of subjects and what kinds of conditions are "normal" to elicit the prototypical response, which was Turing's point with the imitation game (pg398). Author argues that common objections to Turing's thesis is that it fails to capture response-independent notions of intelligence miss the point (pg399).

Author reviews an objection to Turing's test, first from Ned Block's "Aunt Bubbles" thought experiment (pg400-2). Block's thought experiment points out the logical possibility of a (very) large index of conversations, which would emulate intelligence if there was an (impossibly) fast search and probability mechanism. Author responds by putting the operator "actually" into the Turing schema.  

Next discussed are rival views to the behavioristic one; the first is Moor's response-independent view (pg402-3). Moor wants the test to be evidence for thinking, which would be internal to the mechanism. Stuart Shieber argues that the test is an "interactive proof" of intelligence. Author also believes this view misses the response-dependence of the test (pg404).  

Author talks about how response-dependence doesn't need to undermine objectivity, or that it is compatible with "qualified" realism (pg405). Author then talks about more the more recent understanding that "intelligence is in the eye of the observer", a form of (perhaps) illusionism, with comments on the Chinese Room from Jordan Pollack. Further discussion is about a problem with response-dependent concepts: that humans anthropomorphize (just about anything) (pg407-8) "the forensic problem of anthropomorphism". Author believes the Turing test has a two-part solution to this: (1) create a situation where there is a disincentive to anthropomorphize (the judge's accuracy would decline), and (2) making anthropomorphism into a controlled variable (pg409).