Adaptive yet suboptimal integration of advice in decision-making

Area of Science:

• Cognitive Science
• Decision Science
• Artificial Intelligence

Background:

• Effective advice utilization is crucial for decision-making, but individuals often underutilize external input.
• Understanding the limitations in integrating advice is key for improving both human and AI-assisted decision-making.

Purpose of the Study:

• To investigate how individuals integrate advice from artificial agents with varying cognitive and metacognitive profiles.
• To quantify the optimality of advice integration and identify the primary sources of suboptimality.

Main Methods:

• Two experiments involving 89 participants performing a perceptual decision-making task.
• Interaction with seven artificial agents possessing distinct cognitive and metacognitive characteristics.
• Application of a Bayesian modeling approach combined with experimental data to establish a normative benchmark for advice integration.

Main Results:

• Participants adapt their reliance on advice based on adviser profiles, improving performance but remaining suboptimal.
• Key errors identified: egocentric bias (overweighting own judgment) and global-to-local deficit (inconsistent trait translation).
• Suboptimality persists even with full information and is exacerbated with higher-quality advisers, varying with individual metacognitive abilities.

Conclusions:

• Human advice integration is fundamentally constrained by flawed information processing, not just ignorance.
• Egocentric biases and difficulties in applying global knowledge to specific decisions limit effective advice utilization.
• Findings have significant implications for optimizing human-human and human-AI decision-making systems.