Artificial Epanorthosis: Why LLMs overuse a classical rhetorical figure

Epanorthosis is a rhetorical figure in which a speaker replaces or sharpens an initial statement - for example, saying something is not merely a course but a transformative journey. The paper argues that language models use this device too often because their training data contain abundant promotional writing and preference tuning rewards emphatic, confident phrasing. Sequential text generation may intensify the tendency, but is not its main cause. The author frames the problem as a mismatch between model and human rhetorical habits, and proposes measuring it with an Epanorthosis Index that compares a model’s use of the figure with human rates in the same genre.

A preliminary assessment of three sizes of one instruction-tuned model family finds that the mismatch depends on register. Models use epanorthosis at roughly twice the human rate in oratory, and nearly three times the rate in Italian oratory, especially in larger models. They use it less than people in informal question-and-answer writing, while aligning more closely with human rates in argument, journalism, and encyclopedic prose. In Italian experiments, a brief instruction reduced the figure’s frequency by about half to nearly three-quarters; a supervised-fine-tuning LoRA adapter nearly eliminated it, with a scaling control able to restore usage toward the human baseline. The proposed goal is not to ban a legitimate rhetorical device, but to calibrate its use by genre. The broader concern is that habitual exposure to machine-generated style could make people adopt its rhetorical quirks.