Artificial Intelligence Systems That Understand

From its beginnings, artificial intelligence has borrowed freely from the vocabulary of psychology. The use of the word "intelligence" to label our area of research is a case in point. Other terms referring originally to human mental processes that have considerable currency in Al are "thinking," "comprehending," and, with increasing frequency in the past five years, "understanding." In fact, these terms are probably used more freely in Al than in experimental psychology, where a deep suspicion of "mentalistic" terminology still lingers as a heritage of behaviorism. It is not my intent to engage in a barren lexi-cographic exercise, nor to bait those among us who are aroused to indignant emotion whenever terms from human psychology are used in reference to computers. We employ these anthropomorphic terms because we find them useful in defining our research goals, and therefore it is important that we attach clear operational meanings to them. Before discussing computer programs that understand , we need to consider how we can judge whether such programs are successful. When we attribute intelligence to a computer, we mean that it is doing something which, if done by a human, would be called " intel 1 igent" (Felgenbaun and Feldman, 1963, p.3)-The criteria we use to judge whether a man or woman is acting intelligently are generally external criteria; we do not often attach electrodes to our subject's head, much less open it up to peer inside. Instead, we ask questions like, "Is he behaving appropriately to the circumstances?" or "Is she solving the problem?" 1 We do not get much help from either experimental psychology or ordinary speech in defining "under-standing" in a precise way. It is doubtful that there is a single, unambiguous concept lying behind the diversity of usage. This vagueness of everyday language should prepare us to encounter a multi-plicity of different kinds of understanding, and a corresponding multiplicity of criteria that we must apply to behavior (whether of a computer or a human being) to test whether understanding has been achieved. 1 The discussion here is limited to the goals of Al, as usually defined by its practitioners. On the other hand, when we are using the computer as an instrument of psychological research, to simulate human thinking, the criteria for defining intelligence are stricter. In the latter case we must insist on similarity of process as well as similarity of product. Moore and Newell (1974), …