Confidence-accuracy dissociations in perceptual decision making

Confidence in perceptual tasks typically follows accuracy, making it difficult to study the mechanisms specific to confidence computation. However, the last decade and a half has seen a proliferation of experimental manipulations that lead to confidence-accuracy dissociations where one condition leads to higher confidence, but equal (or lower) accuracy compared to another condition. Such dissociations promise to reveal the underlying principles behind confidence ratings. Nevertheless, there is no consensus surrounding how to create confidence-accuracy dissociations or how to interpret their existence. This review first categorizes existing confidence-accuracy dissociations, highlighting that they can be produced by stimulus, task, or brain manipulations. It then examines the most common mechanistic explanations for why such dissociations occur. Finally, it explores several confusions and open questions regarding the use of the term “positive evidence bias,” the relationship between confidence-accuracy dissociations and metacognition noise, and the role of reaction time in such dissociations. Overall, this review highlights the fact that confidence-accuracy dissociations can be produced by multiple mechanisms and suggests that computational models of confidence should be evaluated based on their ability to predict new dissociations.