Hoeks, Vonk, & Schriefers (2002) showed that pragmatic information determining the topic of the next sentence can eliminate processing difficulty in a locally ambiguous "gardenpath" sentence. To further investigate the use of pragmatic information, an ERP experiment was conducted using temporarily ambiguous sentences which were eventually disambiguated in favor of the PREFERRED syntactic structure. These locally ambiguous sentences were preceded either by a context biasing towards the non-preferred reading or by a neutral context. In addition, to investigate the effect of local ambiguity per se, an unambiguous control condition was added (preceded by a neutral context). First, the pragmatic context biasing toward the non-preferred reading elicited an early left negativity followed by a late centro-parietal positivity (P600) at the point of disambiguation, suggesting that pragmatic information is used extremely fast in on-line sentence processing, inducing effortful syntactic (re-)processing even in sentences with the "preferred" syntactic structure. Secondly, the comparison of the ambiguous and the unambiguous sentences following a neutral context revealed significant differences at predominantly left frontal electrodes, strongly suggesting a penalty for ambiguous sentences, even when the ambiguity is resolved towards the most preferred reading. This result provides support for the race model of ambiguity resolution (Van Gompel et al., 1998).