The Social Embedding of Intelligence Towards producing a machine that could pass the Turing Test

I claim that in order to pass the Turing Test over any period of extended time, it will necessary to embed the entity into society. This chapter discusses why this is, and how it might be brought about. I start by arguing that intelligence is better characterised by tests of social interaction, especially in open-ended and extended situations. I then argue that learning is an essential component of intelligence and hence that a universal intelligence is impossible. These two arguments support the relevance of Turing Test as a particular but appropriate test of interactive intelligence. I look to the human case to argue that individual intelligence utilises society to a considerable extent for its development. Taking a lead from the human case I outline how a socially embedded artificial intelligence might be brought about in terms of four aspects: free-will, emotion, empathy and self-modelling. In each case I try to specify what social ‘hooks’ might be required in order for the full ability to develop during a considerable period of in situ acculturation. The chapter ends by speculating what it might be like to live with the result. Robert French (French 1990) rightly points out that the Turing Test is a test of human social intelligence rather than of a putative ‘general intelligence’ – I agree. The test is an inherently social one, which will necessitate the intelligence being embedded into the society in which it is being tested (in the case of the Turing Test, human society). This chapter will discuss why this is, how this might occur for an artificial intelligence and how this might enable the development of some of the humanlike abilities and behaviour that would be necessary in order to pass the Turing Test. In particular I will look at: free-will; the sense of self; empathy and emotion. Before that I argue that there is no such thing as a general intelligence, and so the Turing Test is a very relevant test of intelligence as far as we are concerned. The reasons why a general intelligence is not possible will help motivate the following, more constructive discussion. I end by briefly speculating on how it might be like to live with the result. 1. SOCIAL TESTS FOR INTELLIGENCE In this section I suggest that intelligence is usefully characterised by the ability to succeed with respect to one’s goals in situations of extended interaction with other entities. This will lead back to the Turing Test as a special case. This is in contrast to the classic “constrained problem solving” paradigm, where one has to find the right sequence of discrete moves to obtain a known goal. Classic examples of constrained problems used in AI include the “blocks-world” puzzle, rectilinear mazes, the iterated prisoner’s dilemma, and the travelling “salesman problem”. Most of these puzzles have the following features in common: a sequence of discrete choices have to be made; there are a (usually small) finite number of pre-defined possibilities for the choices each turn; they are played off-line with respect to the world; and the goals are known to all. Such ‘toy’ problems test for only a very restricted aspect of intelligence and almost all researchers accept that such problems are an inadequate test of general intelligence. However such tests continue to dominate the literature – presumably there is an