This paper responds to MacDonald and Christiansen’s (2002) experience-based account of subject vs. object relative clause processing based on Simple Recurrent Network simulations. They found that object-extracted relative clauses exhibit performance penalties that are absent in subject relative clauses, and more so in less trained networks. Whereas MC argue that their finding reflects a differential amount of word order regularity in subject- vs. object-extractions, a detailed analysis of the word-by-word output-activation pattern suggests that it is caused by the network failing to distinguish verbs from the relative pronoun that during early training epochs. This interpretation is supported by other aspects of the activation pattern that indicate incomplete grammar acquisition. Nevertheless, the results point at a viable source of complexity in sentence processing.