Illusory Negative Polarity Item (NPI) licensing in (1) is ungrammatical, yet temporarily treated as acceptable in online processing.

(1) *[The bill [that no senators voted for]] will ever become a law.

This phenomenon has been presented as evidence for a cue-based retrieval mechanism. A competing account argues that the effect reflects overapplication of pragmatic inference mechanisms involved in regular NPI licensing. Existing evidence shows that illusory NPI licensing is robust across tasks and languages, but almost all evidence comes from a single syntactic context. The two accounts make divergent predictions about how the effect should extend to other environments.

In this talk, we present the results from several experiments testing a new syntactic context that should have the ability to distinguish the competing accounts of illusory NPI licensing. Results from the studies show that illusory NPI licensing is limited to environments that support negative pragmatic inferences, favoring the semantic/pragmatic account. Throughout, we consider the relationship between illusory NPI licensing and another grammatical illusion, agreement attraction, asking whether these grammatical illusions reflect different cognitive domains (memory retrieval vs. inference). We find evidence that the profiles do diverge, suggesting that there is not a homogeneous cause for the problems.