Proceedings of the Annual Meeting of the Cognitive Science Society

Previous work on concept learning has focused on how con-cepts are acquired without addressing metacognitive aspectsof this process. An important part of concept learning froma learner’s perspective is subjectively knowing when a newconcept has been effectively learned. Here, we investigatelearners’ certainty in a classic Boolean concept-learning task.We collected certainty judgements during the concept-learningtask from 552 participants on Amazon Mechanical Turk. Wecompare different models of certainty in order to determineexactly what learners’ subjective certainty judgments encode.Our results suggest that learners’ certainty is best explained bylocal accuracy rather than plausible alternatives such as totalentropy or the maximum a posteriori hypothesis of an idealizedBayesian learner. This result suggests that certainty predomi-nately reflects learners’ performance and feedback, rather thanany metacognition about the inferential task they are solving.