Speaker
Description
Recent advances in machine learning have raised ethical concerns in both industry and academia regarding the uncontrollable diffusion of AI and the diminishing human capacity to oversee its impacts. These concerns underscore the need for regulatory and design approaches that maintain human oversight in AI-driven decision-making. Keeping humans in the loop is essential for auditing, fairness, trust, and—critically—ongoing human learning. Without continuous human learning, the knowledge gap between ever-learning machines and static human experts widens, risking the exclusion of humans from meaningful oversight. We focus on future human-AI configurations for high-stakes environments, particularly those that support mutual learning. One such configuration is Reciprocal Human-Machine Learning (RHML), in which humans and machines learn both independently and from each other to reduce their knowledge gap. While RHML holds promise for improving decision quality, it also entails significant operational costs. Its effectiveness depends on the learning capacities of both agents and the allocation of effort across different learning modes.
To formalize RHML, we develop an optimal control model that maximizes a performance objective—balancing decision accuracy against the cost of learning efforts. In this model, decision accuracy evolves as a function of effort allocated to independent and reciprocal learning. We assume that learning contributes to knowledge, which in turn drives accuracy—modulated by knowledge complexity, explainability, and decay. We derive both an open-loop and a feedback solution, where the latter is characterized by an optimal allocation of learning efforts that directly depends on the system’s current level of accuracy.
Our analysis explores how system parameters influence learning effort, decision accuracy, model complexity, and maximized performance. The model also identifies optimal stopping times, where continued learning is no longer cost-effective. We provide insights on inter-channel allocation and inter-agent allocation. We conclude with implications for AI design, policy, and management in decision-critical settings.