Confirmation Bias Found Adaptive: Active Inference Study from Friston et al.

In a paper posted on arXiv on June 22, 2026, researchers Dorje C. Brody, Karl J. Friston, Bernhard K. Meister, and Emmanuel M. Pothos present a new theoretical framework for confirmation bias. Rather than treating it as a flaw, they model decision-making on the space of square-root probabilities using quantum probability structures. Here, observations are represented as matrices, not random variables. In the classic binary hypothesis testing problem, the optimal evidence choice that minimizes expected error probability is shown to lead naturally to confirmation bias—a surprising result that frames bias as a rational adaptation. The authors demonstrate two key evolutionary benefits: (a) the decision maker requires only the smallest memory capacity, and (b) the error probability can be reduced exponentially in sample size. A complementary approach using active inference—where the decision maker seeks evidence that provides maximum information—yields the same optimal evidence choice. The work thus bridges cognitive bias with optimal information processing and provides a protocol for active quantum inference. The 31-page paper includes 5 figures and is classified under Neurons and Cognition and Quantum Physics on arXiv. Karl Friston, renowned for the free energy principle, brings further weight to the findings.</p><p>For AI and machine learning professionals, this paper offers a fresh lens on designing efficient decision-making agents. The implication is that what appears as irrational bias may actually be computationally optimal, especially under resource constraints. By adopting evidence selection strategies that mirror confirmation bias, AI systems could achieve lower memory footprints and faster convergence in sequential decision tasks. The quantum probability framework also suggests new ways to model uncertainty and evidence in reinforcement learning and active learning scenarios. While the work is theoretical, it opens avenues for practical algorithms that embrace bias as a feature, not a bug. The alignment with active inference, a principle behind many modern AI architectures, makes this research particularly relevant for developers building autonomous agents that must balance accuracy and efficiency.