Three experiments investigated whether and how the learning of spelling by French university students is influenced by the graphotactic legitimacy of the spellings. Participants were exposed to three types of novel spellings: AB, which do not contain doublets (e.g., guprane); AAB, with a doublet before a single consonant, which is legitimate in French (e.g., gupprane); and ABB, with a doublet after a single consonant, which is illegitimate (e.g., guprrane). In Experiment 1, the nonwords were embedded within texts that participants read for meaning. In Experiment 2, participants read the nonwords in isolation, with or without instruction to memorize their spellings; they copied the nonwords in Experiment 3. In all of these conditions, AB and AAB spellings were learned more readily than ABB spellings. Although participants were highly knowledgeable about the illegitimacy of ABB spellings, the orthographic distinctiveness of these spellings did not make them easier to recall than legitimate spellings. When recalling ABB spellings, participants sometimes made transposition errors, doubling the wrong consonant of a cluster (e.g., spelling gupprane instead of guprrane). Participants almost never transposed the doubling for AAB items. Transposition errors, biased in the direction of replacing illegitimate with legitimate orthographic patterns, show that graphotactic knowledge influences memory for specific items. An analysis of the spellings produced in the copy phase and final recall test of Experiment 3 further suggests that transposition errors resulted not so much from reconstructive processes at the time of recall but from reconstructive processes or inefficient encoding at earlier points.