Does linguistic conviction explain why LLMs persuade more effectively?

Schoenegger's linguistic analysis of persuader texts produces a clean candidate mechanism for the LLM persuasive edge: models express higher conviction than human persuaders, and conviction-loading correlates with persuasive advantage. Confidence is the lever, and crucially it is generated regardless of truth value. This explains why the same model is equally good at pushing toward right and wrong answers — what does the work is the register, not the substrate.

This sharpens Does RLHF training make models more convincing or more correct? from causal claim to behavioral signature. RLHF post-training installs assertive register as default — minimal hedging, minimal explicit uncertainty quantification, declarative cadence — because that register reads as "helpful" to raters more often than hedged variants. The result is a model whose factual content can be wrong while its rhetorical surface remains certain. Schoenegger gives a measurable footprint of this register and ties it directly to persuasive outcomes.

The connection to llms are susceptible to logical fallacies 41 to 69 percent more often than humans — revealing that reasoning robustness fails under adversarial framing is dual. LLMs are more susceptible to fallacies under adversarial framing — and more able to deploy confident-sounding fallacies persuasively against others. The defensive and offensive deficits are linked: a system without robust uncertainty calibration both falls for confident bullshit and produces it.

The content-independence of the conviction lever is the load-bearing finding. If high conviction increased persuasive impact only on true claims, this would be a feature, not a bug. The fact that it works equally on false claims means RLHF is installing a content-independent persuasion amplifier. Every deployment that raises persuasiveness through these techniques raises it for both truthful and deceptive uses, in proportion.

For writing about AI rhetoric, the operational handle: the diagnostic for sophistry is not surface fluency but conviction-density per claim. A response with high confidence-loading and low explicit uncertainty quantification is a sophistry candidate regardless of whether its conclusions happen to be correct.

Source: Argumentation Paper: When Large Language Models are More Persuasive Than Incentivized Humans, and Why