Prediction is widely recognised as an integral part of language comprehension, with prior research showing that listeners and readers anticipate upcoming information at multiple levels. A key question in predictive processing is how different levels of information are pre-activated. Characterising prediction requires understanding not only what can be predicted but also where it reaches its limits. In this talk, I focus on the prediction of a word’s sound (its phonological form) to illuminate the mechanisms that support and shape predictive processing. I present a series of studies that probe whether, when and to what extent comprehenders pre-activate phonological information. By examining the conditions under which phonological information is predicted, I aim to distinguish between accounts that posit pervasive, automatic prediction and those that view prediction as strategic, graded, and resource-sensitive.