In metamemory research, it is well stablished that metamemory judgments are inferences made based on cues. Several cues have been found to underlie predictions of future memory performance – judgments of learning (JOLs) – such as concreteness, emotionality, etc. However, little is known about specific mechanisms for learning cues. To explore this gap, we tested statistical learning as a mechanism to extract regularities from the environment and use them as cues to inform JOLs. Across two experiments, participants were exposed to a continuous auditory stream of artificial words with fixed transitional probabilities between adjacent syllables. Afterwards, they studied and made JOLs for items that 1) were presented in the stream (word), 2) were not presented in the stream, but followed transitional probabilities (phantom), and 3) did not follow transitional probabilities (non-word). Results showed that JOLs were based on the transitional probabilities: JOLs were higher for word and phantom than for non-word items. In Experiment 1, using an old-new recognition memory test, discrimination was worse for word and phantom than for non-word items. In Experiment 2, using a 2-alternative-forced-choice recognition memory test with the same type of target and distractor item in each trial, performance did not vary across trial types, but there were group differences depending on whether statistical learning was assessed during learning or not. Hence, linguistic cues for JOLs can be acquired through statistical learning, but such cues are not necessarily valid and may lead to metamemory illusions.