Turning AI into a regulatory sandbox: exploring information disorder mitigation strategies with ABM and deep reinforcement learning

Information disorder, understood as the creation, dissemination, and amplification of false, misleading, or harmful information, has raised growing concern in recent years. The convergence of multiple interconnected factors—planetary spread of online information, network interactions that amplify cognitive biases, and algorithms fostering echo chambers—has made the phenomenon difficult to predict and challenging for public authorities to counteract. In such a scenario, a major obstacle to developing effective regulatory and control policies lies in the difficulty of grounding decisions in sufficiently realistic models and representations accounting for the many mechanisms influencing the phenomenon at individual, collective, and institutional levels. Actually, research in computational social science has made huge strides in illuminating the core dynamics of information disorder, primarily through mathematical models drawing from graph theory, complex network analysis, and operations research. Despite their strengths, such approaches rely on abstract representations. Such representations often struggle to capture the complexity of the reality which is shaped by emotional, psychological, and adaptive dynamics. Therefore, they are hard to integrate into purely mathematical descriptions. This paper presents ongoing research aimed at overcoming such limitations through a hybrid strategy for the in silico exploration of mitigation strategies. We integrate two components: (1) an agent-based simulation module to reproduce the core dynamics of information disorder in a controlled environment, leveraging well-established theoretical models; (2) a deep learning module in charge of piloting a super-agent (metaphorically, the public authority) whose task is to autonomously experiment with increasingly complex mitigation strategies. This approach allows for a gradual increase in model complexity, leveraging machine learning to identify the most effective solutions. Suitable to being extended to other social issues, such an approach allows for a gradual increase in model complexity, enabling the creation of progressively more realistic models while entrusting the machine learning component with the task of identifying the most effective strategies to act on them.