Adaptive Dual-Channel Diffusion of Disinformation and Correction with LLM-Calibrated Influence on Networks

Abstract

Disinformation rarely spreads as a single and isolated process. It competes with corrective information for the same audience, and the outcome depends on both how widely false narratives propagate and when corrective interventions are introduced. Existing diffusion models typically assume fixed transmission probabilities, model only a single narrative, and initialize edge influence independently of the content being communicated. We propose an adaptive dual-channel diffusion framework in which disinformation and fact-check narratives evolve as coupled, history-dependent processes with adaptive edge influence and cross-channel belief inhibition. Instead of assigning initial influence from an arbitrary prior, a locally-hosted large language model generates a claim and rebuttal for each directed edge and estimates their persuasive strength, producing content-grounded initial influence weights. The fact-check process is introduced as a delayed, targeted intervention, reflecting realistic correction dynamics. Experiments on a Congress Twitter network show that delayed fact-check interventions substantially reduce the final spread of disinformation, with cross-channel belief inhibition and adaptive link reinforcement jointly driving this reduction. Initial edge influence is derived from LLM-generated claim/rebuttal pairs for a sampled subset of edges, with the remaining edges initialized from the resulting score distribution. Within this content-grounded subset, the LLM consistently scores disinformation as more persuasive than its paired rebuttal. These findings demonstrate that coupling adaptive, competing narratives captures correction dynamics beyond the scope of traditional, static single-narrative models, with content-grounded initialization offering a promising direction for a subset of the network.