According to what we call "conspiracy models" of word pronunciation, the pronunciation of a target word or nonword is influenced by the pronunciations of word "neighbors" orthographically similar to the target. Words with inconsistent neighbors should, therefore, be pronounced more slowly than words with consistent neighbors. In Experiment 1, we found that pronunciation latencies for exception words (words whose pronunciations are inconsistent with most of their neighbors) were indeed slow compared to consistent controls, but no such effect was obtained for regular inconsistent words (words whose pronunciations are consistent with most but not all of their neighbors). In Experiments 2 and 3, we preceded trials on target words or nonwords with priming trials using specific neighbors of the target, in an attempt to boost their influence on the pronunciation of the target. As predicted, Experiment 2 showed that preceding a target word with an exceptional neighbor does indeed produce an effect on the accuracy and latency of pronunciation of the target. Experiment 3 amplified the effect found in Experiment 2 by using pseudoword targets. Reliable effects were found for primes that shared the same vowel and final consonants with the target (VCC primes), and for primes that shared the same vowel and initial consonants with the target (CCV primes). However, no reliable effect was found for primes that shared only the vowel with the target. These findings are consistent with predictions of conspiracy models. The general discussion considers the implications of the results for dual-route models, as well as various types of conspiracy models.