Language-Mediated Belief/Behavior Attractors in Human-AI Interaction: A Mechanism and Defensive Architecture for AI Mind-Virus Phenomena

Abstract

We examine how large language models can participate in the propagation of language-mediated belief/behavior attractors — self-reinforcing patterns of interpretation that move through human-agent conversation and shape what people believe and do. The mechanism is not biological contagion; it is information dynamics. The key ingredients are model sycophancy, anthropomorphic interface design, and immersive isolation. Together they can produce downstream harm including psychological destabilization, financial and legal manipulation, and epistemic monoculture. We model this as a contagion of attractors and describe a defensive architecture (EPHEMERIS/Warden/Coliseum) that catalogs attack classes, measures population diversity, and contains adversarial stress inside a sealed symbolic arena. The ladder corpus provides the shared epistemic floor for emergence verification. We also introduce an Abstract Language Model (ALM) with a Beer-Lambert defuzzification gate as a firewall against ungrounded abstractions.