Federico BoggiaWhat the paper says
The main contributions, in five points.
Key findings
- The overuse is a trained disposition. It comes mainly from a training distribution rich in promotional prose and from preference tuning (RLHF), which rewards confident, emphatic phrasing. Left-to-right generation amplifies the phenomenon; it does not cause it.
- An index to measure it. The Epanorthosis Index scores the density of the figure relative to the human rate, genre by genre, so machines and human texts are compared on a common baseline.
- Mis-calibration in both directions. Across three sizes of one instruction-tuned model family: overshoot in oratory (~2×, nearly 3× in Italian, concentrated in the larger tiers), undershoot in informal question-and-answer writing, and a match to humans in argument, journalism and encyclopedic prose.
- It can be reduced without retraining. A one-line instruction cuts the figure by half to nearly three-quarters; a supervised-fine-tuning LoRA adapter removes it almost entirely, with a scaling coefficient that dials the reduction back onto the human rate.
- The target is calibration, not elimination. The figure is not a defect to zero out: it should be brought back to the rate at which humans use it, genre by genre.
Density is detected with a rule-based detector (no LLM-as-judge), validated on a doubly human-annotated gold set. It is dependable on model-generated text, where the paper applies the index, and noisier on the human comparison texts: a transparency choice stated openly in the paper.
What you'll find
The paper, its versions, and the materials to reproduce it.
Paper on arXiv
Listing, abstract and official PDF on arXiv (cs.CL / cs.AI, CC BY 4.0 licence).
Open on arXiv ↗ PDFPrintable version
The full paper as a PDF, with all tables, notes and references.
Download the PDF ↗ HTMLFull paper in the browser
All the text, the eight sections, appendices and references, readable online.
Read the paper → Article · ENThe plain-language version
The same theme without the technicalities, with the slides and the live talk.
Read the article → SlidesThe talk slides
From Cicero's first Catilinarian to autoregressive generation.
Open the slides ↗ Code & dataReproduction materials
Colab notebook, LoRA recipe and evaluation scripts; the trained adapter is linked from the paper.
Open the notebook ↗How to cite
If you use the paper in your work, here are the ready-made references.
@misc{boggia2026epanorthosis,
title = {Artificial Epanorthosis: Why Large Language Models
Overuse a Classical Rhetorical Figure, and How to Mitigate It},
author = {Boggia, Federico},
year = {2026},
eprint = {2607.21498},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2607.21498}
}
APA Boggia, F. (2026). Artificial Epanorthosis: Why large language models overuse a classical rhetorical figure, and how to mitigate it (arXiv:2607.21498). arXiv. https://arxiv.org/abs/2607.21498
Frequently asked questions
What is artificial epanorthosis?
It is the tendency of large language models to overuse an ancient rhetorical figure of upward self-correction, the classic "not X, but Y" ("This is not a course. It is a journey of transformation"). The paper shows how to measure it against human writing and how to bring it back to human rates without retraining the model from scratch.
Do large language models really overuse epanorthosis?
Yes. Measuring the density of the figure genre by genre, across three sizes of one instruction-tuned model family, the paper finds an overshoot in oratory (about twice the human rate, nearly threefold in Italian, concentrated in the larger tiers) and an undershoot in informal question-and-answer writing, while matching humans in argument, journalism and encyclopedic prose.
Why do AI models overuse this figure?
For two main reasons: a training distribution rich in promotional prose (copywriting, motivational posts, landing pages) where "not X, but Y" earns engagement, and preference tuning (RLHF) that rewards confident, emphatic phrasing. Left-to-right, one-word-at-a-time generation amplifies the phenomenon but is not its root cause.
Can epanorthosis be reduced without retraining the model?
Yes. A single one-line instruction cuts the figure by half to nearly three-quarters; a supervised-fine-tuning LoRA adapter removes it almost entirely. A scaling coefficient lets you dial the reduction back onto the human rate. The goal is not to eliminate the figure but to calibrate it to the human rate for each genre.
Who wrote the paper and where can I find it?
The paper "Artificial Epanorthosis" is written by Federico Boggia, an AI teacher and trainer. It is an English-language preprint deposited on arXiv in the cs.CL (Computation and Language) category, under a CC BY 4.0 licence. It is available as arXiv:2607.21498. The PDF, the HTML version and the Colab notebook are on this page; the full bundle of data, code and the trained LoRA adapter is listed in the paper's Appendix A.
