Costello, Pennycook & Rand's 2024 paper offers its 'central-route' reasoning-shift explanation as a hedged inference, and names the mechanism itself as unconfirmed

conspiracy-beliefLLM-persuasionAI-epistemicselaboration-likelihood-modelmechanism-vs-correction

The 2024 paper's Discussion ties together the durability of its effect and its cross-topic spillover and offers an explanation drawn from the Elaboration Likelihood Model: participants processed the AI's arguments deeply enough to internalize them, "consistent with the 'central route' to persuasion... which is known to promote durable belief change." Central-route (effortful, argument-based) processing is contrasted in that theory with peripheral-route (cue-based) persuasion, which does not generalize or last.

The paper states this as consistent with its data, not established by it. Its Limitations section is explicit: "the causal mechanisms underpinning our results remain unformalized." On a direct read, the 2024 paper does not claim to have confirmed that participants underwent a general shift in how they weigh evidence — it offers that reading as the authors' own best-fit interpretation of a result (durability + spillover) that is also compatible with other mechanisms, and says in its own text that it could not formally distinguish between them.

This qualifies the optimistic reading in claim-tailored-ai-dialogue-durably-reduced-conspiracy-belief: the "deeply entrenched beliefs were reachable after all" result stands on its measured numbers, but the stronger claim that followers' underlying reasoning changed is the authors' hedge, not their finding.