How do bilinguals detect the language of the words they read? Recent electrophysiological research using the masked priming paradigm combining primes and targets from different languages has shown that bilingual readers identify the language of the words within approximately 200 ms. Recent evidence shows that language-detection mechanisms vary as a function of the orthographic markedness of the words (i.e., whether or not a given word contains graphemic combinations that are not legal in the other language). The present study examined how the sub-lexical orthographic regularities of words are used as predictive cues. Spanish–Basque bilinguals and Spanish monolinguals (control group) were tested in an Event-Related Potential (ERP) experiment, using the masked priming paradigm. During the experiment, Spanish targets were briefly preceded by unrelated Spanish or Basque words. Unrelated Basque words could contain bigram combinations that are either plausible or implausible in the target language (Spanish). Results show a language switch effect in the N250 and N400 components for marked Basque primes in both groups, whereas, in the case of unmarked Basque primes, language switch effects were found in bilinguals but not monolinguals. These data demonstrate that statistical orthographic regularities of words play an important role in bilingual language detection, and provide new evidence supporting the assumptions of the BIA+ extended model.