Do LLM arguments actually argue better than humans?

LLM-generated counter-arguments score higher than human counter-arguments on the markers a rhetoric textbook would teach: they are more cogent, more explicitly justified, more respectful toward the interlocutor, and more positive in emotional tone. Humans, in contrast, score higher on three orthogonal features: greater lexical and syntactic creativity, more negative emotion, and stronger use of interactive discourse markers (turn-taking signals, addressivity, conversational repair).

The pattern is more specific than "LLMs argue better." It says LLMs argue the way an instructor wants students to argue, while humans argue the way actual people in actual disputes argue. The textbook-quality profile is a recognizable artifact of training: RLHF-style objectives reward politeness, justification, and emotional restraint; they penalize the very features that make human argumentation distinctive — disagreement intensity, creative phrasing, and the conversational micro-moves that signal a real exchange between people.

The implication for detection is uncomfortable. The features that separate LLMs from humans are precisely the features prescribed argument quality: by being good students of argumentation, LLMs become identifiable. This creates a perverse incentive in the other direction: if detection were a serious cost, the cheapest evasion would be to add lexical noise, negative emotion, and conversational disfluency — that is, to make outputs worse by textbook standards in order to look more human. The textbook–human gap is the detection surface.

The deeper finding is that argument quality and argumentative authenticity are different things. A model trained to produce good arguments will reliably fail to produce human arguments. The two targets diverge.