Are the peculiar design characteristics of human languages merely the result of evolutionary accidents that determined innate limitations on linguistic competence? Or do they stem from the fact that human languages are used for practical purposes, and that efficient use demands sentence structures that can readily be produced and parsed? The only way to decide this question is to look for functional explanations of language structure; the search should be informative, whether it ultimately succeeds or fails. Some grammatical constraints have the effect of ruling out one structural analysis of a potentially ambiguous word string, thus disambiguating it and reducing uncertainty for the sentence parsing routines. So far, however, there has been no convincing account in these terms of the major ‘island’ constraints on transformational operations. It is not plausible, for example, that these are motivated by the strain of holding a transformational ‘filler’ constituent in memory until its associated ‘gap’ has been found, or by the difficulty of identifying gaps in certain sentential contexts. I suggest that the parsing routines play a crucial but less specific role in determining the constraints. They call for some restrictions on filler and gap positions, but not for any particular restrictions. The actual pattern of permitted filler—gap relations stems from the expressive function of transformations. Only fillers and gaps that are central to expressive purposes are admitted; all others are excluded by the constraints. Despite the difficulty of defining expressive function, this approach offers promising explanations for why different transformations, and similar transformations in different languages, are subject to different constraints. I conclude that the form of a language may indeed be shaped by its function, but only if ‘function’ is construed broadly enough to include communicative goals as well as the mechanics of the encoding and decoding operations. Finally, there are indications that these functional pressures on language design can operate only within certain very narrow confines established by the nature of linguistic competence.