Most Human Human and Machine Intelligence

I've had a chance to read a galley copy of "The Most Human Human", by Brian Christian. This relates to his quest in the 2009 Loebner Prize Turing Test competition. Besides identifying the "most human computer", the annual event also designates a person who has been most confidently identified by the judges as human. In a 5 minute interactive text exchange each judge has to both select "human" or "computer" and their confidence in their selection, which leads to the two potential 'winners' (humans get no substantial awards for being human, and are discouraged from trying to simulate a machine.)
Christian raises the interesting question in the process of "how do we know what is human?" This is the focus of the book -- pursuing historical concepts from Greek Philosophers, to modern instantiations such as Garry Kasparov's chess competitions with IBM's Deep Blue. (Christian is aware of the then upcoming Jeopardy--Watson match, but went to print prior to that.) This approach ends up focusing on the diversions and "rat holes" more than the question of what distinguishes humans, intelligence, or consciousness. Since this specific instantiation of the Turing Test has a time limit (expanded from 2009 at 5 minutes, to a 25 minute head-to-chip comparison for 2011) the AI's created for the contest are "purpose built". Much of Christian's discussion focuses on differentiation from the 'single purpose' programs of the past to be convincingly distinguished as human.
Many of the points he raises provide insight on the nature of being human:
  • Persons have a consistent, unique identity (not changing point of residence, gender, relationship status or such from one input line to the next.) Exemplar AI's do not have this same sense of personal history/identify.
  • Persons have a sense of context -- except, interestingly enough, Christian points out when they are arguing ... then responses often degenerate to reply to the last comment made, not the initial topic triggering the dispute. Some (at times convincing) AI's simply respond to the most recent input with no continuity.
  • Persons 'add value' (hopefully) in interactions, ideally surfacing new concepts which were not implicit from strict analysis. (Christian touches on left brain/right brain distinctions here.)
And his listing goes on -- returning regularly to the point of "how can I use this to emerge as the most human human?"
From the technologists perspective, and getting to Turing's initial concept, "how would we know if an AI can think?" Here I find the Lobener approach to be simplistic. Fooling 30% of the judges in 2009, and 50% in 2011 does not satisfy my criteria for "thinking" (of course I'm not sure that some persons, perhaps politicians for example would clear my hurdle here either.) Consider a few alternative situations:
  • An AI which is not purpose built but consistently is considered to be a human respondent in general discourse
  • An entity known to be an AI which is generally agreed is thinking, conscious, intelligent...
Perhaps a more challenging concept is an AI that is thinking, but doesn't pass the Turing test ... perhaps because it does not care to be judged against human standards. It is a point of some arrogance in the part of humans to presume that the only instantiations of 'thinking', 'consciousness' or 'intelligence' must be evident as paralleling similar human characteristics.

No comments:

Post a Comment