Successful sentence comprehension routinely requires us to establish dependencies between non-adjacent words and phrases, which are subject to strict grammatical constraints on their structural, morphological, and semantic properties. In this talk, I will discuss how investigating the successes and failures of real-time grammatical constraint application provides a window into the core cognitive processes for sentence comprehension. In particular, I will demonstrate that even though language interpretation seems immediate, effortless, and accurate, we often fail to implement even the simplest of grammatical relations like subject-verb agreement, giving rise to so-called “linguistic illusions.” Drawing on experimental and computational evidence from agreement, anaphora, and ellipsis, I will show how to systematically switch linguistic illusions “on/off”. These findings motivate a new conception of how we mentally encode, access, and update syntactic representations during sentence processing. I will argue that even when things go wrong on the surface, as in the case of illusions, the underlying cognitive processes are more grammatically sophisticated than expected under current models of sentence processing, and that the selective errors are confined to limitations of the domain-general memory retrieval mechanisms that implement the grammar. To conclude, I will describe several avenues for future research and discuss how we can use linguistic illusions to engage the general public in the language sciences.