Do LLMs and humans persuade through different cognitive routes?

Bilstein's qualitative synthesis across the 7 studies in the meta-analysis surfaces an empirical pattern that maps cleanly onto the Elaboration Likelihood Model. LLM-generated persuasive messages rely on analytical reasoning and informational coherence — the central route, which works best under high motivation and ability to elaborate. Human-generated messages remain more emotionally vivid and personally engaging — the peripheral route, which works best under heuristic processing, low elaboration, identity-driven attitudes, and source credibility.

This is not a competitive framing but a complementary one. The same recipient in different states is reachable by different speakers. Under high motivation (relevant decision, sufficient cognitive resources), LLMs' fact-based, coherent, multi-step argumentation finds purchase. Under low motivation (skim reading, identity-charged topics, fatigue), humans' emotional resonance and affective signals find purchase. Under most real-world conditions, both routes are partially active and the persuasive winner depends on which route dominates for that recipient on that topic.

This sharpens large language models are as persuasive as humans but how — cognitive effort moral emotional language. That note frames the question; ELM gives the theoretical scaffolding for the answer. LLMs produce text that demands cognitive effort to process and delivers analytical density that rewards the effort — a central-route profile. Humans produce text whose moral and emotional language is read fast and acts on identity rather than argument — a peripheral-route profile.

It also connects to Do humans and LLMs differ fundamentally or just superficially?. The route asymmetry is most visible from the observer perspective (analytical vs emotional features are detectable in surface text); from the participant perspective, both routes can move the dial, and the participant typically does not know which route is doing the work.

For writing about persuasion, ELM gives an empirical handle for separating persuasion-as-argument from persuasion-as-rapport. They are not graded points on a single dimension. They are different cognitive routes, served by different speakers, optimal under different recipient states. AI's distinctive capability set — coherence, analytical density, factual recall — privileges one route. Its distinctive deficit — affective embodiment, identity grounding — under-serves the other.

Source: Argumentation Paper: A meta-analysis of the persuasive power of large language models