Two types of Spanish-English bilinguals were tested in an event-related potential (ERP) experiment on a contrast in the two languages exemplified in (1) and (2) in order to investigate linguistic permeability during processing of Spanish (1a and 2a). In Spanish, but not English, absence of the complementizer que is ungrammatical. (1) a. Qué hermana confesó Inés que había comido la tarta? b. *What sister did Inés confess that had eaten the cake? (2) a. *Qué hermana confesó Inés Ø había comido la tarta? b. What sister did Inés confess Ø had eaten the cake? In a first analysis, we grouped subjects by generation and compared ERP responses to que-less vs. que-full sentences. A significant N400 effect was found for first-, but not second-generation, suggesting reduced sensitivity to missing que for the latter. However, a second analysis, using linear mixed modeling to test predictiveness of individual speaker variables revealed generation to be non-predictive of N400 amplitude. Instead, current language use, cumulative exposure to English, and socioeconomic status (SES) were significant predictors for all subjects: increased English use, exposure, and SES resulted in smaller N400 amplitude to the anomaly in Spanish shown in (2a). Our results show that a priori classification of bilinguals masks gradient cross-linguistic effects, and processing is permeable in all bilinguals depending on amount of language use. Results from an acceptability judgment task administered to the same subjects using a subset of the same stimuli show that both subject groups judge que-less and que-full to be equally natural. These results suggest that behavioral measures that rely on metalinguistic judgments may not be good indicators of processing, and that having to appeal to metalinguistic knowledge may mask intrinsic knowledge.