Do LLMs and humans persuade through the same mechanisms?

A 1,251-participant study of human and AI persuasion across 56 contentious claims found that LLM-generated and human-generated arguments shifted reader agreement at comparable rates. Same persuasive force. But the textual mechanisms producing that force diverged systematically. LLM arguments required higher cognitive effort to process — more grammatically complex, more lexically dense. They used moral language more heavily across positive and negative foundations. Sentiment was comparable; cognitive complexity and moral framing were not.

The authors call this "no equivalence in process despite equivalence in outcome." It is a consequential framing because it severs the standard inference from persuasive success to underlying mechanism. When two arguments persuade equally, we typically infer that they did so for similar reasons. The data here say the opposite: equivalent persuasive force can rest on entirely different rhetorical scaffolding.

For a Language as Event reading, this is precisely the place where the AI's production process and the human's interpretation come apart. The reader experiences the argument as a unified utterance from a unified speaker — a stance, a tone, a voice. But what the model produced and what the human produced are different kinds of textual artifact, achieving the same effect through non-overlapping mechanisms. The human produced an event-residue from within a communicative situation; the LLM produced an event-residue that simulates one. Their persuasive force is equivalent because the audience cannot distinguish them on textual surface. The mechanisms are different because only one of them was actually communicating.

Source: Argumentation