It is well-known that individuals can identify novel words in a stream of an artificial language using statistical dependencies. While underlying computations are thought to be similar from one stream to another (e.g. transitional probabilities between syllables), performance are not similar. According to the “linguistic entrenchment” hypothesis, this would be due to the fact that individuals have some prior knowledge regarding co-occurrences of elements in speech which intervene during verbal statistical learning. The focus of previous studies was on task performance. The goal of the current study is to examine the extent to which prior knowledge impacts metacognition (i.e. ability to evaluate one’s own cognitive processes). Participants were exposed to two different artificial languages. Using a fully Bayesian approach, we estimated an unbiased measure of metacognitive efficiency and compared the two languages in terms of task performance and metacognition. While task performance was higher in one of the languages, the metacognitive efficiency was similar in both languages. In addition, a model assuming no correlation between the two languages better accounted for our results compared to a model where correlations were introduced. We discuss the implications of our findings regarding the computations which underlie the interaction between input and prior knowledge during verbal statistical learning.