Thinking in Vain: An Evidence Accumulation Account of Biased Reasoning

Reasoning biases are often explained as a failure to detect conflict between an erroneous intuitive response and logical information —or, when conflict is detected, an inability to correct the initial error during deliberation. We propose that both failure modes share a common computational origin: a distorted latent evidence signal that jointly drives choice, confidence, and response time. In a classic base-rate neglect task, participants (N = 151) judged which of two groups an individual most likely belonged to, given base-rate and stereotypical information about each group. Stereotype strength was quantified using large language models, allowing us to independently manipulate both sources of information across trials. We found that for most participants, choices, confidence, and RTs were driven by a shared latent evidence signal that strongly overweighted stereotype relative to base-rate information. We formalized this signal using an evidence accumulation model in which the strength of the stereotype and base-rate influenced the strength of evidence accumulation. We show that biased drift dynamics produce misplaced deliberation, leading reasoners to deliberate too little where it would help most and too much where it would not. The model further predicts that increasing decision caution to deliberate longer yields sharply diminishing accuracy gains. Consistent with this prediction, a large out-of-sample individual-participant meta-analysis of reasoning datasets (total N = 2,428) showed that "thinking longer" did not substantially increase error correction among highly biased participants. These findings show that persistent reasoning biases arise from a biased evidence accumulation process, rather than a failure to engage in deliberation.