The exact semantic processes subserving the formation of false memories are still poorly understood. Here, we directly probed the semantic origins of false memories in a typical Deese-Roediger-McDermott (DRM) task, by predicting participants' performance in this task through data-driven distributional semantic models. Participants were required to study lists of words and then to perform a recognition task. Our findings indicate that the participants' performance is better accounted for by a local rather than a global strategy on the task at hand: the single lists composing the task activate specific semantic clusters that are responsible for the occurrence of false memories. In particular, memory performance followed a continuous semantic gradient, with higher false recognitions occurring for higher sematic similarity between the lures (i.e., the false memory items) and the words in the relative lists. Crucially, our findings also show that semantic memory is differently involved in veridical and false memories, with this pattern being consistent across two reanalyses of data from previous studies and being replicated in an independent experiment. We thus outline an empirically-driven theoretical framework to account for the semantic processes supporting veridical and false memories formation.