Do LLMs and humans persuade through the same mechanisms?

Human-generated and LLM-generated arguments have been shown to achieve equivalent persuasive force across many studies. The standard interpretation is that LLMs have closed the gap with human writers. The persuasion-strategies study suggests a different reading: equivalent outcomes can arise from non-overlapping production mechanisms, and the two types of text are persuading through different rhetorical pathways.

Human arguments tend to be emotionally vivid and personally engaging — drawing on lived experience, narrative authenticity, identity-based framing. LLM arguments substitute different ingredients: higher grammatical and lexical complexity (which signal substance), more frequent moral language across foundations (which heightens stakes and engages moral reasoning), and comparable sentiment without the personal vividness. Both produce persuasion. They produce it through different routes.

This matters for at least two practical questions. First, detection: if human and LLM persuasion mechanisms differ, then forensic features distinguishing them may be reliable enough to support source attribution even when the persuasive effects are indistinguishable. Second, vulnerability: the two production pathways have different failure modes. Human persuasion's reliance on personal engagement makes it brittle when the speaker's authenticity is questioned. LLM persuasion's reliance on cognitive complexity and moral framing makes it brittle when readers learn to recognize and discount these specific signals.

For a Language as Event reading, the deeper point is that persuasive equivalence is a measurement artifact at the level of effect, not a claim about communicative parity. The LLM is not doing the same thing as a human writer and arriving at the same place. It is doing a different thing that the audience cannot, on current detection abilities, distinguish.

EMNLP 2025 CMV analysis sharpens the divergent-mechanisms picture along three dimensions. First, the emotional gap is not "more emotion overall" but a specific emotion-class signature: LLM counter-arguments show elevated anticipation and trust in particular, not generalized affect. Second, LLM replies stylistically converge with the original post — across named entities and psycholinguistic features — while human replies do not mirror the post in the same way. This convergence is itself a mechanism of equivalence: LLMs partially produce persuasion by tracking the input style. Third, the detection corollary: the production-mechanism gap is large enough that lightweight interpretable features achieve ~99% accuracy distinguishing LLM from human counter-arguments. Equivalent persuasive outcomes do not entail equivalent forensic signatures; the latter remain distinguishable even when the former converge. Source: Argumentation.