In past decades, researchers have conducted a myriad of masked priming lexical decision experiments aimed at unveiling the early processes underlying lexical access. A relatively overlooked question is whether a masked unrelated wordlike/unwordlike prime influences the processing of the target stimuli. If participants apply to the primes the same instructions as to the targets, one would predict a response congruency effect (e.g., "book-TRUE" faster than "fiok-TRUE"). Critically, the Bayesian Reader model predicts that there should be no effects of response congruency in masked priming lexical decision, whereas interactive-activation models offer more flexible predictions. We conducted 3 masked priming lexical decision experiments with 4 unrelated priming conditions differing in lexical status and wordlikeness (high-frequency word, low-frequency word, orthographically legal pseudoword, consonant string). Experiment 1 used wordlike nonwords as foils, Experiment 2 used illegal nonwords as foils, and Experiment 3 used orthographically legal hermit nonwords as foils. When the foils were orthographically legal (Experiments 1 and 3; i.e., a standard lexical decision scenario), lexical decision responses were not affected by the lexical status or wordlikeness of the unrelated primes, as predicted by the Bayesian Reader model and the selective inhibition hypothesis in interactive-activation models. When the foils were illegal (Experiment 2), consonant-string primes produced the slowest responses for word targets and the fastest responses for nonword targets. The Bayesian Reader model can capture this pattern, assuming that participants in Experiment 2 were making an orthographic legality decision (i.e., anything legal must be a word) rather than a lexical decision.