The metacognitive paradox of OCD: confidence is globally reduced but shows increased sensitivity to local evidence

Abstract

Confidence is a critical metacognitive signal that guides performance. Biases in confidence, such as excessive doubt, are hallmark features of mental health disorders, especially obsessive-compulsive disorder (OCD). Yet, the underlying neurocognitive mechanisms and how they link to learning and decision-making remain elusive. We asked patients with OCD and matched healthy controls to perform a novel rule-shifting task incorporating trial-by-trial confidence ratings. Using a Bayes-optimal model, we identified two distinct confidence biases: while patients with OCD indicated lower overall confidence, their trial-by-trial confidence ratings more accurately tracked task-relevant information, rendering their confidence reports more Bayes-optimal than those of controls. These findings challenge the idea of a simple, unified metacognitive impairment in OCD. Instead, they suggest that OCD is linked to an enhanced responsiveness to environmental evidence and feedback during decision-making.

Keywords: Cognitive Flexibility; Confidence; Decision Making; Learning; Metacognition; Obsessive-Compulsive Disorder; Rule-Shifting.

Figure 1.. Rule-shifting task.

(A) Participants used a touchscreen device to complete a new rule-shifting task (the “Cryptic Creatures Task”, implemented in the citizen science app Brain Explorer, www.brainexplorer.net) involving two or three multi-dimensional creatures. Each dimension (e.g., color, shape, presence of “hair”) has specific features (e.g., orange vs. blue, star vs. round, hair vs. no-hair). On each trial, participants chose one creature and then rated their decision confidence in that choice using a slider. As they moved the slider, anchor labels updated (e.g., “totally guessing,” “pretty unsure,” “pretty certain,” “totally certain”). Feedback was shown as either a coin (correct choice) or a cross (incorrect choice), with coins accumulating in the top-right corner of the screen. (B) Unannounced rule shifts occurred every 6–11 trials and required participants to flexibly update their decision strategy. During intra-dimensional (ID) shifts, the currently relevant dimension remained the same but switched from one feature to another (e.g., from “hair” to “no-hair”). During extra-dimensional (ED) shifts, the relevant dimension itself changed (e.g., switching from “hair vs. no-hair” to “star-shape vs. round-shape”), requiring participants to learn which new dimension was now rewarded. (C) The task progressed through three levels of increasing difficulty. At Level 1, participants were shown two creatures differing on two dimensions. At Level 2, additional dimensions were introduced (still two creatures). Finally, at Level 3, a third creature was added, making it more difficult to identify the correct dimension and feature.

Figure 2.. Average accuracy and confidence ratings overall and relative to rule shifts in the task.

Participants played a new intra-/extra-dimensional rule-shifting task that included confidence ratings. (A) Participants’ average accuracy dropped when a shift happened (vertical dashed lines) but recovered afterward (right panel). Contrasting behavior between the two groups, we observed no significant difference in overall accuracy between patients with OCD and controls (left panel). (B) Confidence ratings dropped after the shifts and increased with more trials under the new rule (right panel). Patients were significantly less confident than controls throughout the task (left panel). The box plots on the left of each panel show the means of all trials across the entire task for each measure. The right panels display the mean behavioral measures from five trials before to five trials after each shift. Error bars represent standard errors. Statistical significance was determined using two-sided two-sample t-tests (***p<0.001, n.s. = non-significant).

Figure 3.. Confidence ratings in patients with OCD are more sensitive to external feedback.

(A) A mixed-effects model indicated that, overall, patients with OCD were less confident than controls and that confidence was positively associated with accuracy on the preceding trial. (A-B) There was a significant interaction between group and preceding accuracy, showing that the influence of feedback on next-trial confidence differed between groups. Follow-up comparisons examining this interaction revealed that participants with OCD exhibited heightened feedback sensitivity: (C) they showed a significantly greater average increase in confidence following positive feedback and (D) a significantly greater average decrease following negative feedback compared to controls (Bonferroni-corrected). Panels (A) and (B) display model estimates or effects illustrating the main effects and interaction. Box plots in (C) and (D) represent mean confidence levels or changes on relevant trials. Error bars represent standard errors and *p < 0.05, **p < 0.01, ***p < 0.001.

Figure 4.. Bayesian-observer certainty is linked to confidence fluctuations.

(A) Time course of the Bayesian observer’s Global Certainty (GC; dashed blue) and Choice Certainty (CC; purple) around an extra-dimensional (ED) rule shift, with the x-axis ranging from pre-shift trials (−2, −1), to the shift trial (extra-dimensional rule shift; dashed line at trial =0), and post-shift trials (1–4). Pre-shift, GC and CC are high under the learned “round-shape” rule. At the ED shift (0), the latent rule changes to “hair”. A negative outcome (‘X’ over the chosen stimulus) at the shift renders the previous rule incompatible, collapsing GC toward zero while CC 50, reflecting the sum of priors for the chosen features (‘hair’, ‘star-shaped’) under a belief state now divided equally between all present features (‘hair’, ‘no-hair’, ‘star-shaped’, ‘round’). On trial 1, a correct choice (yellow coin) and trial 2, an incorrect choice, allow the model to perform hypothesis elimination by reinforcing the correct features and discarding the irrelevant one. This resolution causes GC to rise, and with a final confirmatory choice on trial 3, GC and CC converge at maximal certainty, signifying that the new correct feature has been identified. (B) Across both groups, confidence ratings were significantly predicted by GC and CC. A significant interaction effect revealed that the relationship between CC and subjective confidence was significantly stronger in participants with OCD compared to controls, indicating more optimal tracking of choice-specific evidence in the patient group. Error bars represent standard errors and **p < 0.01, ***p < 0.001.