Three experiments (total N = 1,058) were conducted to investigate the relationship between memory accuracy and subjective confidence using thematic lists constructed on the basis of spontaneous accessibility, that is, the frequency with which items are spontaneously generated as category members. After memorizing lists of words and performing a distractor task, participants completed tests of recognition memory and rated confidence in their memory for 4 item types: studied items (second highest spontaneous accessibility) and 3 types of nonstudied items, including strong lures (highest spontaneous accessibility), weak lures (third highest spontaneous accessibility), and semantically unrelated items. In Experiments 1 and 2, the items were presented as lists, whereas in Experiment 3, they were embedded in short vignettes. In Experiments 1 and 3, amount of material to be encoded (1 vs. 10 or 8 lists of 15 items), and in Experiment 2, modality of stimulus presentation at encoding and at test (visual–visual, auditory–auditory, and auditory–visual) were varied between participants. Across all experiments, the confidence–accuracy relationship remained consistently positive for studied items. However, increasing the amount of information to be memorized, using inconsistent stimulus modalities across encoding and retrieval, and embedding the items in vignettes resulted in (a) zero or negative confidence–accuracy relationships for weak lures and (b) highly negative confidence–accuracy relationships for strong lures. These results demonstrate that subjective confidence judgments are made on the basis of inferential processes in the course of which spontaneous accessibility is mistaken for memory strength.