Many of the choices we make are accompanied by a sense of confidence. Within classical Signal Detection Theory (SDT), confidence is conceptualized as the absolute distance between a decision variable and a decision criterion. The decision criterion is traditionally modelled as being stable over an experimental session. However, recent work challenges the notion of a static decision criterion, suggesting instead that the criterion undergoes trial-by-trial fluctuations. Combining SDT theory and model simulations, we predict that fluctuations in the decision criterion shape confidence. In 15 human decision-making datasets, trial-by-trial estimates of decision criterion were obtained with the Hierarchical Model for Fluctuations in Criterion (hMFC). Across all datasets, we confirmed our pre-registered hypothesis that confidence is shaped by single-trial criterion state. This effect was found in 14 out of 15 individual datasets, indicating a robust pattern across a variety of task paradigms and confidence reporting scales. Going beyond self-report, the shaping of confidence by criterion fluctuations was replicated in an implicit measure of confidence, RTs, and in two key neurophysiological markers, pupil-linked arousal and a neural signature of confidence. Our results demonstrate that variability in confidence, which has traditionally been treated as noise, actually reflects genuine sensitivity to the current state of the (fluctuating) decision criterion.