Adaptive yet suboptimal integration of advice in decision-making

Effective use of advice is critical for sound decision-making, yet individuals often fail to fully benefit from external input. Across two experiments, 89 participants performed a perceptual decision-making task while interacting with seven artificial agents characterized by distinct cognitive and metacognitive profiles. By combining a Bayesian modeling approach with experimental data, we establish a normative benchmark to quantify the optimality of advice integration. We show that individuals adaptively adjust their reliance on advice based on the cognitive and metacognitive profiles of the different advisers. While this adaptation generally improves performance, the integration process remains significantly suboptimal, leading to a substantial loss in potential accuracy. This suboptimality is primarily driven by two integration errors. First, an egocentric bias leads decision-makers to overweight their own judgments and confidence signals relative to those of the adviser. Second, a global-to-local deficit prevents them from consistently translating knowledge of each partner's global traits into trial-level adjustments. These errors persist even when participants are explicitly informed of both their own and the adviser's attributes, indicating that the bottleneck is not ignorance but flawed information processing. Suboptimality is particularly severe when interacting with higher-quality advisers and varies widely across individuals, tied specifically to their own metacognitive sensitivity and calibration. Our findings reveal fundamental constraints in information integration with implications for both human-human and human-AI-assisted decision-making.