Confidence is detection-like in high-dimensional spaces

Confidence estimates are often “detection-like” - driven by positive evidence in favour of a decision. This empirical observation has been interpreted as showing human metacognition is limited by biases or heuristics. Here we show that Bayesian confidence estimates also exhibit heightened sensitivity to decision-congruent evidence in higher-dimensional signal detection theoretic spaces, leading to detection-like confidence criteria. This effect is due to a nonlinearity induced by normalisation of confidence by a large number of unchosen alternatives. Our analysis suggests that detection-like confidence is rational when computing confidence in a higher-dimensional evidence space than that assumed by the experimenter.

💡 Research Summary

The paper tackles the long‑standing observation that confidence judgments in perceptual tasks often appear “detection‑like,” meaning they are driven primarily by evidence supporting the chosen option rather than by the balance of evidence between alternatives. This pattern, known as the positive evidence bias (PEB), has traditionally been interpreted as a suboptimal heuristic or bias in human metacognition. The authors propose instead that PEB can arise naturally from optimal Bayesian confidence computation when the observer’s internal evidence space is high‑dimensional, even if the experimental task is low‑dimensional.

First, the authors review existing explanations of PEB, including response‑congruent evidence heuristics, selective neural read‑out, and stimulus‑variance effects. They note recent work showing that convolutional neural networks trained on MNIST also exhibit a PEB, suggesting a more general computational origin.

The core theoretical contribution is an extension of classic signal detection theory (SDT) from the usual one‑ or two‑dimensional formulation to a k‑dimensional space. In the standard 2‑dimensional SDT model, the observer receives a noisy evidence vector x =