AI Contagion study shows how AI feedback loops destabilize social networks

A new theoretical economics paper by Olivier Bos and Stefano Bosi examines how artificial intelligence (AI) interacts with social communication networks to destabilize collective knowledge. In their model, agents exchange information through a network while simultaneously receiving AI-generated content. Critically, the AI systems retrain on the aggregate social information they helped generate. This creates two feedback forces: an "AI contagion channel," where distortions propagate across the network like a virus, and an "AI social distortion multiplier," where retraining amplifies previous errors. The authors show that even with high-dimensional agent states, the long-run dynamics collapse to a simple two-dimensional system. The stability of the entire information ecosystem is determined by the spectral radius of this reduced system — a measure of how quickly distortions grow or decay.

Bos and Bosi mathematically derive a sharp regulatory frontier: a minimum level of AI filtering required to keep the system stable. If filtering falls below this threshold, initial distortions explode into systemic informational instability — essentially, an AI-generated misinformation cascade. Importantly, the required filtering depends heavily on the topology of the underlying social network. Highly connected or clustered networks require stricter filtering than sparse, random ones. The paper, published on arXiv as 2606.15206, runs 49 pages with 2 figures and spans economics, AI, and social network theory. It offers a formal foundation for understanding why AI-generated content can lead to runaway echo chambers and provides policymakers with a quantitative tool to set safety standards for AI-mediated communication.