Experiments using memory paradigms have shown that general terms receive context-dependent encodings. The results of these experiments can only provide indirect evidence about representations set up at the time of presentation. Those studies which have concerned themselves with on-line processing of general terms have produced only equivocal results. The present experiment investigates the encoding of category and instance nouns as a function of (i) whether their referents were previously described by category or instance nouns; (ii) whether preceding context suggests what kind of thing is being referred to. Where context gives no specific information about a category member, the occurrence of an instance noun in a following sentence provides extra new information about its referent. It was found that this extra new information takes time to process. However, when context has already suggested an instantiation of a category noun, then the later occurrence of the appropriate instance noun does not slow readers down. The results indicate that representations set up during reading are the product of both the linguistic input and of general knowledge. There is no reason to suppose that further elaboration of these representations is required in order to achieve a full understanding of a text.