Model-free metacognition

Abstract

Extensive work has been done on the metacognitive capacities of humans, as well as to investigate metacognitive processes in nonhuman animals. What we propose here, however, is that there are two very different forms that metacognition can take: either model-based (implicating at least a simplified model of the thinker's own mind), or model-free (representing some mental state or process in oneself in the absence of any such model). The focus of all work on human metacognitive judgments has been on the model-based variety, as have been most attempts to discover metacognition in animals. We first review recent studies suggesting that there are no resources shared between human metacognitive judgments and the sorts of behavioral tests employed with animals, implying that the latter fail to provide evidence of even simplified forms of model-based metacognition. Thereafter the question of model-free metacognition in animals is pursued. Negative verdicts are rendered on a pair of possible claims of this sort. But two positive answers are defended. One is that epistemic emotions like curiosity and interest, as well as the signals involved in failed memory searches, implicate representations whose content is, unknown. The other is that decisions to deploy attentional / mental effort (which many animals besides humans can do) depend on appraisals of an analog-magnitude signal representing the extent to which executive resources are engaged.