The folded-X pattern has been identified as a critical signature of confidence: as conditions become easier, confidence increases for correct trials but decreases for error trials. However, recent work has identified violations of the folded-X pattern where easier conditions lead to increased confidence for both correct and error trials (double-increase pattern). Nevertheless, it remains unclear which stimulus manipulations produce each pattern. Here we test the hypothesis that the double-increase pattern emerges for manipulations of the quality of the sensory input (stimulus reliability), whereas the folded-X pattern emerges for manipulations of the distance between the relevant sensory feature and the decision boundary (boundary distance). Across two experiments ( N = 78) using orientation judgment with either Gabor patches or moving dots, we first replicate previous findings that boundary distance manipulations have a stronger effect on accuracy whereas stimulus reliability manipulations have a stronger effect on confidence. Critically, boundary distance manipulations produced the classic folded-X pattern, whereas stimulus reliability manipulations yielded the double-increase pattern. Artificial neural networks (ANNs) trained on the same tasks exhibited folded-X patterns for both manipulations, suggesting that human confidence judgments do not simply reflect the statistical nature of the task and stimuli. Reaction time (RT) patterns largely mirrored confidence, though with some notable exceptions, underscoring both the utility and limitations of RT as a confidence proxy. These results demonstrate that different stimulus manipulations have dissociable effects on the signatures of confidence and suggest that human confidence is influenced by mechanisms nor present in standard ANNs.
Xue et al. (Tue,) studied this question.