Bias or Optimality? Disentangling Bayesian Inference and Learning Biases in Human Decision-Making

2025-05-12

AI Summary

This study investigates whether “confirmation bias” and “positive bias” observed in human performance on a bimanual Bernoulli bandit task reflect genuine cognitive biases or are merely epiphenomena arising from decaying learning rates. Using Bayesian inference modeling, stochastic dynamical analysis, and master equation theory, we formally demonstrate that symmetric yet decaying learning rates alone suffice to reproduce the canonical statistical signatures of these biases; moreover, Bayesian updating is mathematically equivalent to Q-learning with a decaying learning rate. We show that standard asymmetric Q-learning models readily misattribute learning-rate decay to cognitive asymmetry. To resolve this confound, we propose a novel experimental paradigm and a parameter-identifiability framework enabling empirical discrimination between true cognitive biases and learning-rate dynamics. Our results establish rigorous model-selection criteria for causal interpretation of cognitive biases, advancing computational psychiatry and decision neuroscience.

Technology Category

• Machine Learning: Bayesian Learning
• Cognitive Modeling & Cognitive Systems: Simulating Human Behavior
• Game Theory and Economic Paradigms: Adversarial Learning

Application Category

• Economics, Online Markets and Human Computation: Social networks and social learning
• User Modeling, Personalization and Recommendation: Psychology-informed user models and recommender systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Recent studies claim that human behavior in a two-armed Bernoulli bandit (TABB) task is described by positivity and confirmation biases, implying that humans do not integrate new information objectively. However, we find that even if the agent updates its belief via objective Bayesian inference, fitting the standard Q-learning model with asymmetric learning rates still recovers both biases. Bayesian inference cast as an effective Q-learning algorithm has symmetric, though decreasing, learning rates. We explain this by analyzing the stochastic dynamics of these learning systems using master equations. We find that both confirmation bias and unbiased but decreasing learning rates yield the same behavioral signatures. Finally, we propose experimental protocols to disentangle true cognitive biases from artifacts of decreasing learning rates.

Problem

• Disentangle Bayesian inference from learning biases in human decisions
• Analyze behavioral signatures of confirmation bias versus unbiased learning
• Propose protocols to distinguish true cognitive biases from artifacts

Innovation

• Bayesian inference as effective Q-learning algorithm
• Master equations analyze stochastic learning dynamics
• Experimental protocols distinguish biases from artifacts