Abstract

This essay attempts to distill the relationship of the frame problem to the Turing test in order to support two claims. First, it is argued that passage of the test presupposes a solution to the frame problem. Second, it is argued that the prospects for solving the frame problem are sufficiently poor that it is unlikely that any digital computer can be programmed so that it can pass the Turing test. It is argued that these results count as new evidence that artificial intelligence's understanding of intelligence, namely, that it is the manipulation of symbols under the control of a computer program, is fundamentally mistaken. ;The essay attempts to reach a considered perspective on what it means for a digital computer to compute some function. The conclusion reached is that basic computer operations are both surprisingly simple and epistemologically different in kind from the operations that evidently occur in human brains. The basic distinction developed in the essay is that between epistemologically simple and epistemologically complex learning. The claim is that computers are capable of ES learning but not EC learning. ;The frame problem, it is claimed, has to do with how an axiom system can revise its representation of the world in light of new experiences. Since there appears to be no "principled" way to do this, the result is that computers suffer from a "relevance space" problem which blocks natural conversational ability needed to pass the Turing test. The essay also argues that the inaccessibility of brains and mind-like programs is such that the prospects for hoping to understand them are poor. The result is that we will have to rely on I/O test such as Turing's in assessing the significance of AI. ;The stance of the essay is the implications of the frame problem for the Turing test count in favor of the "poor substitute" strategy for opposing AI and against the "hollow shell" strategy for opposing it favored by Searle and Gunderson. The essay argues that, since EC learning is presupposed by the Turing test, and computers are capable of only ES learning, they will not be capable of passing the Turing test. This result, it is concluded, counts against the "physical symbol system hypothesis" of classical AI