Human-AI Complementarity Enhanced by Error Correlation

Integrating machine learning models to augment human decision-making often falls short, even when the AI provides valuable signals. A key challenge lies in the asymmetric information regarding the quality of information available to humans versus AI, which hinders the realization of complementary gains. This study investigates how the error correlation structure between human and AI predictions critically influences a decision maker's ability to extract value from AI. The findings highlight that when an AI's prediction errors are negatively correlated with those of a human, decision-makers can devise robust strategies. These strategies are shown to guarantee improvements in expected utility, indicating a path towards more effective human-AI systems. Empirical investigations using real-world forecasting benchmarks confirm that these conditions for robust complementarity can arise in practical scenarios. This suggests that understanding and leveraging error correlation is vital for designing AI systems that truly enhance human capabilities rather than merely providing additional, often unutilized, information.