It is widely assumed that 'predictable' words are often phonetically reduced. It is currently unclear what brings about this reduction: Do speakers strengthen and lengthen certain words for the benefit of their listeners, and reduce other words that they deem likely to be intelligible even when reduced? Or do speakers lengthen or shorten words depending on how much time they themselves need for planning and encoding their utterances? This paper summarizes and partly extends some recent work investigating what happens when intelligibility and ease of production lead to different expectations about whether words are likely to undergo reduction or its opposite, phonetic strengthening. The work capitalizes on the fact that words in 'dense' phonological neighborhoods tend to be challenging targets for recognition, other things being equal, but easy targets for production. The main finding is that regression models of word duration and vowel centralization in conversational speech suggest that words in dense neighborhoods are phonetically reduced, other things being equal.