Knowledge considered harmful [AI]

One of the major achievements of artificial intelligence research has been the recognition of the central role of knowledge in intelligent performance. This is the `knowledge is power hypothesis´. This has led to a great emphasis on research into the problems of providing systems with knowledge: knowledge representation and knowledge acquisition. An underlying assumption of most of this work appears to be a belief that additional knowledge of the problem domain is always beneficial (or at worst of neutral value) to a problem solver. The authors´ work on machine learning has led them to question this optimistic view of the effects of adding knowledge to a system. They believe that the knowledge is power hypothesis is correct if interpreted to mean that substantial quantities of domain knowledge are necessary for effect problem solving. However, they reject the view that additional knowledge is always either beneficial or, at worst, neutral in its effects. The authors show that additional knowledge may often have a detrimental effect on problem solving. Consequently, the relationship between knowledge and performance is not monotonic; a conclusion that has far-reaching implications for the development of knowledge based systems