Evaluative conditioning (EC) refers to the acquisition of emotional valence by an initially-neutral stimulus (conditioned stimulus; CS), after being paired with an emotional stimulus (unconditioned stimulus; US). An important issue regards whether, when participants are unaware of the CS-US contingency, the affective valence can generalize to new stimuli that share similarities with the CS. Previous studies have shown that generalization of EC effects appears only when participants are aware of the contingencies, but we suggest that this is because (a) the contingencies typically used in these studies are salient and easy to detect consciously, and (b) the performance-based measures of awareness (so-called "objective measures"), typically used in these studies, tend to overestimate the amount of available conscious knowledge. We report a preregistered study in which participants (N = 217) were exposed to letter strings generated from two complex artificial grammars that are difficult to decipher consciously. Stimuli from one grammar were paired with positive USs, while those from the other grammar were paired with negative USs. Subsequently, participants evaluated new, previously-unseen, stimuli from the positively-conditioned grammar more positively than new stimuli from the negatively-conditioned grammar. Importantly, this effect appeared even when trial-by-trial subjective measures indicated lack of relevant conscious knowledge. We provide evidence for the generalization of EC effects even without subjective awareness of the structures that enable those generalizations.