Artificial Epanorthosis: Why LLMs overuse a classical rhetorical figure (arxiv.org)

A recent study highlights how large language models (LLMs) exhibit a notable overuse of the rhetorical figure known as epanorthosis—self-correction aimed at emphasizing a point—often seen in advertising and promotional contexts. This phenomenon stems from LLMs being trained on datasets rich in promotional language and the reinforcement learning from human feedback (RLHF) that favors confident expressions. The study introduces an Epanorthosis Index to measure the frequency of this rhetorical feature across different genres, revealing that while LLMs exceed human rates in formal oratory, they underperform in informal styles like conversational Q&A. The significance of these findings lies in their implications for AI-generated text, as the overemphasis on certain rhetorical devices could skew the quality of machine-generated prose, potentially leading to a homogenization of writing styles. Notably, the research proposes practical solutions, including lightweight LoRA adapters and targeted fine-tuning of models, which effectively reduce epanorthosis by adjusting outputs closer to human norms without completely eliminating its presence. As the study concludes, calibrating LLM outputs to align with human linguistic patterns across various genres is crucial, highlighting the ongoing need for careful oversight as technology continues to shape our communication styles.