One question of enduring interest for psycholinguists is the question of how closely real-time sentence processing routines align with grammatical knowledge: does the competence grammar directly constrain sentence comprehension, or does it play a secondary role, 'cleaning up' the results of a comprehension process driven by heuristics (Lewis & Phillips, 2015; Patson & Ferreira, 2007; Townsend & Bever, 2001)? Much experimental work has provided evidence for the view that the human sentence processor is directly constrained by grammatical knowledge. A challenge for this view is the observation that there are a number of apparently simple grammatical constraints that comprehenders fail to respect in comprehension, such as subject-verb agreement (e.g. Wagers, Lau & Phillips, 2009). Providing an explanation of why we see these 'grammaticality illusions,' and why only certain dependencies seem to be be susceptible to illusory grammaticality, has been a productive research project that has led to new and diverse models of linguistic dependency formation in real-time comprehension (Phillips, Wagers & Lau, 2011). In this talk, I will review this work, and present two case studies from the research group at UMass that provide some new perspectives, and new puzzles, for this project. In the first part of the talk, I will present evidence from English reflexive processing (joint work with Shayne Sloggett) that suggests that grammaticality illusion associated with English reflexives are created by assigning a logophoric interpretation to the reflexive form. In the second part, I present work with Jérémy Pasquereau on a novel grammatical illusion, French quantifier/de-phrase dependencies. This work suggests that not all intrusive licensors are created equal, and that only some quantifiers have the potential to create grammaticality illusion effects in French. Taken together, these studies suggest that grammatical illusions are conditioned by the availability of alternative structures and interpretations made available by the grammar. In a slogan, it appears that what could've been said, but wasn't, seems to play a role in creating grammaticality illusion effects in comprehension.