According to the standard signal-detection model of recognition memory, confidence judgments for recognition responses are reached in much the same way that old-new decisions are reached (i.e., on the basis of criteria situated along the strength-of-evidence axis). The question investigated here is how the confidence criteria shift when recognition accuracy is manipulated across conditions. Although several theories assume that the old-new decision criterion shifts when recognition accuracy changes, less is known about how the confidence criteria move. An analysis of data previously reported by R. Ratcliff, G. McKoon, and M. Tindall (1994) and some new data reported here suggest that the confidence criteria fan out on the decision axis as d' decreases. This result is qualitatively consistent with the predictions of a likelihood ratio model, although the data did not support the stronger quantitative predictions of this account.