Current psychological models of word production (e.g., Dell, 1986; Levelt et al., 1999) only detail how we plan the phonological content of words, and not how we articulate them. A first step towards understanding how these models may be extended involves determining the nature of the flow of information from phonological encoding to articulation. For example, does activation cascade from unselected phonological representations? So far, the clearest support for cascading has come from the finding that erroneously produced phonemes exhibit characteristics of the intended phoneme (Goldrick & Blumstein, 2006). In this talk however, I will show how computational investigations demonstrate that both cascading and non-cascading models can account for this result. An extension of model behaviour analysis to other speech error phenomena additionally shows that models based on the classic spreading activation account of word production (Dell, 1986) still have problems in replicating some key aspects of human error patterns.