![[Adobe Stock]](https://sp-ao.shortpixel.ai/client/to_webp,q_glossy,ret_img,w_750,h_375/https://www.psypost.org/wp-content/uploads/2025/04/artificial-intelligence-750x375.jpg)
Anyone who spends much time reading text generated by artificial intelligence has likely noticed a recurring stylistic habit: a sentence makes a plain statement and immediately negates it in favor of something grander, such as “This is not a course. It is a journey of transformation.”
A new preprint posted on arXiv examines how large language models use this ancient rhetorical figure, known as epanorthosis, and finds early signs that the models overuse it in some genres and underuse it in others. The paper also tests several ways to turn the habit down, some of which brought it close to typical human rates in small pilot tests.
An ancient figure of thought in modern machines
Epanorthosis is a rhetorical device in which a speaker or writer returns to a statement they have just made in order to correct, qualify, or replace it. The word comes from Greek roots meaning “a further straightening,” and Roman orators and rhetoricians such as Cicero and Quintilian cataloged it in Latin under the name correctio.
It takes three main functional forms: an emphatic upgrade that replaces a word or claim with a stronger one (“he is good, indeed excellent”), a softening downgrade (“it is a disaster, well, a setback”), and a substantive revision that makes a description more exact (“he was running, or rather, walking briskly”). People use the figure in everyday speech when they use “I mean” or “or rather” to revise what they just said, whereas chatbots tend to favor the dramatic upward version, frequently producing phrases like “We do not sell products. We build experiences.”
Don’t miss the next big psychology finding
Every morning, PsyPost sends a short briefing on the most interesting new peer-reviewed research in psychology and neuroscience.
In an 1827 treatise on rhetoric, the French scholar Pierre Fontanier drew a distinction between surface verbal corrections and epanorthosis as a broader “figure of thought,” meaning a general operation of discourse in which an utterance turns back on itself. In the new single-author paper, Federico Boggia, a teacher and trainer who founded the company Binatomy and holds a degree in humanities computing from the University of Pisa, argues that this classification makes it possible to study the device wherever language appears. In Boggia’s framing, a statistical program can perform the same structural move that ancient orators used for emphasis, whether or not anything like persuasive intent lies behind it.
Where the habit may come from
Recent work has established that machine-generated prose exhibits measurable stylistic habits that separate it from human writing. A 2025 study of parallel human and artificial intelligence text collections found that large language models systematically diverge from human authors in their grammatical and rhetorical patterns, and that this gap is wider in models tuned to follow instructions and persists even as models grow larger.
Similarly, a 2025 computational linguistics study examining why certain vocabulary words like “intricate” spiked in chatbot outputs found no evidence that model architecture, text-generation algorithms, or raw training data alone caused the surge; instead, its comparative tests were consistent with human preference tuning playing a role, though the authors left the exact source open.
Boggia proposes that the overuse is mainly a learned tendency shaped by two stages of model development. First, models are trained on vast amounts of web text, a substantial share of which is copywriting and promotional posts. The author suspects dramatic self-corrections help these genres capture attention, though this link has not been measured.
Second, models go through reinforcement learning from human feedback, a later training stage in which people rank candidate answers. Boggia argues that raters tend to reward phrasing that sounds confident and punchy, and the paper treats this stage as the more likely driver.
In this account, the way models produce text plays a smaller, amplifying role. Because they generate one word fragment at a time with no chance to delete and revise before showing the result, whatever emphatic habit they have learned appears unedited.
At the same time, Boggia notes that textual self-correction is not always an empty stylistic habit. When models are prompted to solve problems step by step, phrases like “Wait, let me reconsider” shift the statistical context for upcoming words toward alternative lines of reasoning. However, a 2024 study found that language models often fail to improve their reasoning through self-correction, and sometimes make their answers worse, unless they receive outside feedback such as a verified fact or tool result.
Because self-correction also helps human writers express nuance and honesty in argumentative prose, Boggia set out to measure how often models use emphatic epanorthosis across different genres and to test ways of calibrating the figure to natural human rates rather than erasing it completely.
Measuring self-correction in Italian and English
To turn this stylistic impression into a measurable quantity, Boggia built a simple program that searches for specific word patterns in the “not X, but Y” family, including variations like “not only X, but Y” and two-sentence corrections such as “Not X. Y.” The researcher first validated an Italian version of this detector against 206 short text passages of about 150 words each: 180 drawn from a human-authored Italian reference collection spanning spontaneous speech, academic philosophy, nineteenth- and early twentieth-century fiction, pre-2020 news articles, and social media posts, plus 26 model-generated passages.
Two coders, one of whom was the author, independently labeled every passage for genuine corrective epanorthosis and resolved their disagreements to create a benchmark set. The author describes the measurements as pilot-scale first signals, and the English comparison texts and prompts are available from the author on request.
Separating genuine rhetorical self-correction from ordinary contrast proved difficult even for the human coders, who reached only fair agreement before reconciling their ratings. Against the final human-verified labels, the automated detector caught just over half of all true corrections (52 percent) and was accurate 45 percent of the time overall when it flagged a match.
However, its accuracy split sharply depending on the source: across the 26 model-generated passages, when the detector flagged a phrase, it was a genuine self-correction 82 percent of the time, compared to only 17 percent of the time in human writing, where people mostly used “not… but” for simple contrast. In transcribed speech, the detector missed most corrections because speakers typically revised their phrasing with conversational markers like “that is” or “let’s say” rather than the formal “not X, but Y” structure.
Next, Boggia ran an English-language analysis with an English version of the detector, which, unlike the Italian version, was not separately validated. The analysis compared human texts against outputs from three sizes of a single model family tuned to follow instructions: Anthropic’s Claude Haiku, Sonnet, and Opus, ordered from smallest to largest, though the paper does not specify which exact version of each model was tested.
The human baseline drew on public-domain and openly licensed texts across seven genres, filtered where possible to material written before 2022: encyclopedia articles from Wikipedia, news stories from Wikinews, academic abstracts, fiction by Arthur Conan Doyle, informal question-and-answer posts from Stack Exchange, argumentative essay writing from John Stuart Mill’s On Liberty, and historical speeches by figures such as Abraham Lincoln and Theodore Roosevelt.
Across those seven genres plus a model-only promotional writing category, the three models responded to neutral, genre-specific prompts that never mentioned correction or any rhetorical figure. All texts were divided into windows of a few hundred words, yielding 17 to 28 windows per genre on the model side and 19 to 128 on the human side, to measure emphatic self-corrections per ten thousand words.
In two of the seven genres, the models diverged from the human baseline in opposite directions. In oratorical speeches, the models produced emphatic epanorthosis at more than double the human rate, averaging 33.5 instances per ten thousand words compared to 14.9 in human speeches, an overshoot concentrated in the two larger models (Sonnet and Opus) while the smallest model (Haiku) matched the human baseline.
In informal question-and-answer writing, the models used the figure at less than one-fifth the human rate (1.3 versus 8.2 instances per ten thousand words), as human writers frequently qualified and corrected their explanations while the models rarely did. Because seven genres were tested at once, the odds of a fluke result rise, and the author notes that neither the speech nor the question-and-answer result would remain statistically significant under a strict correction for this, so the paper treats both as suggestive first signals.
In the remaining five paired genres, the differences between machine and human rates were not statistically significant, though the small samples limited what those tests could detect. The models’ rate in academic abstracts was roughly double the human rate, nearly the same ratio as in speeches, and their rate in fiction was about half. Rates in argumentative essays, journalism, and encyclopedic prose were close to even.
To check whether the spike in speeches simply reflected a general habit of using more linking words, Boggia counted neutral connecting words and found that the models actually used only a fifth as many of them in speeches as humans did (1.5 versus 7.5 per ten thousand words), suggesting that the surge was specific to emphatic self-correction. Boggia summarized these genre comparisons in an Epanorthosis Index, which divides the model rate by the human rate for each genre so that a score above 1 signals overuse and a score below 1 signals underuse.
Building a stylistic dial to curb the habit
To test whether this habit can be adjusted without retraining an entire model from scratch, Boggia used Low-Rank Adaptation, a technique for fine-tuning (additional training on a smaller, targeted set of examples) that freezes a base model’s weights (the numerical settings that determine a model’s behavior) and attaches a small, removable set of trainable weights called an adapter.
Working in Italian on Qwen2.5-7B-Instruct, an open-weights model whose internal settings are publicly downloadable, the researcher first trained an adapter on 570 pairs of emphatic and plain sentences using direct preference optimization, which teaches a model to rank preferred phrasing over dispreferred phrasing. That first attempt failed to carry over to new test prompts, raising emphatic epanorthosis from 60.3 to 68.9 instances per ten thousand words.
Boggia then changed the design for a second pilot, using supervised fine-tuning (training directly on examples of the desired plain style) over 797 affirmative paraphrases designed to preserve the original meaning. That second adapter removed almost all emphatic epanorthosis on held-out test prompts. The adapter’s training data, code, and weights were posted publicly.
The researcher then evaluated the adapter at five strength settings, from 0 (turned off) to 1.0 (full strength), across six Italian prompts (two each for oratory, promotional, and argumentative writing, with six outputs per prompt) to see if the adapter could act as a continuous stylistic dial.
In these pilot tests, which were evaluated descriptively without statistical tests, running the supervised adapter at full strength cut overall density from 61.4 instances per ten thousand words with the adapter off to 4.7 at full strength. For argumentative essays, an intermediate setting landed on the human rate: three-quarters strength lowered the model’s rate from 51.3 to 12.3 instances per ten thousand words, essentially matching the human benchmark of about 12, before full strength dropped it to zero.
Oratory fell from about 33 at quarter strength to zero at half strength, passing its human benchmark of about 14 between the tested settings. Promotional writing, which had no human benchmark, jumped from 68.5 to 96.0 at quarter strength before dropping sharply at higher settings, which the author takes as a sign that the dial must be set separately for each genre.
Testing prompt instructions and training-free fixes
Because many users interact with commercial chatbots whose weights they cannot modify, Boggia also tested whether a simple prompt instruction could curb the figure in Italian across the three Claude models, generating 12 texts per genre in each condition (72 texts in total). Adding a single sentence instructing the models to write affirmatively and avoid negate-and-upgrade phrasing cut emphatic epanorthosis by nearly three-quarters in oratory (from 39.6 to 10.9 instances per ten thousand words) and by 70 percent in argumentative essays, bringing both genres close to human rates, while a 48 percent reduction in promotional writing was not statistically significant.
Across genres, the two larger models dropped their use of the figure to zero under the instruction while the smallest model cut its rate roughly in half, though the author notes that prompting is the least robust mitigation option because a surface instruction may not reliably override a habit learned during training.
In a separate descriptive test on Italian prompts without statistical tests, Boggia evaluated two additional training-free strategies that require no changes to the initial prompt. Generating six candidate responses and automatically selecting the one with the lowest epanorthosis score reduced the figure by roughly two-thirds (from 59.9 to 20.9 instances per ten thousand words), which the author argues tends to keep phrasing more natural because it selects an output the model actually produced. Asking the model to rewrite its own draft affirmatively reduced the rate to zero across all three genres, overshooting past the human rate in the same way the full-strength adapter did.
What the pattern may say about language
Stepping back from the engineering tests, Boggia suggests that the spontaneous appearance of a two-thousand-year-old rhetorical figure in neural networks may point to deeper structural patterns in language itself. In the author’s reading, figures like epanorthosis may act as patterns that language naturally tends to settle into, which classical orators discovered through intuition and language models reach through probability. The author argues that the practical challenge is teaching both models and human writers a sense of when self-correction fits, while guarding against the risk that human writers begin adopting the repetitive habits of the machines.
Caveats and future directions
As with all research, there are some caveats to consider when interpreting these findings. This paper is a preprint that has not yet been peer reviewed, so the findings should be treated as preliminary. In addition, the English genre measurements and the prompting demonstration tested three sizes of a single model family that all share one training process, so the results cannot separate the quirks of that specific setup from models tuned to follow instructions as a whole. The study also did not directly compare raw base models against human-feedback-tuned versions to test whether preference tuning causes the habit.
Several measurement constraints also affect the baseline comparisons. The automated detector only tracks surface “not X, but Y” phrasing, the English version used in the main genre study was never separately validated, and the Italian validation showed that the detector flags many ordinary contrasts in human writing, making the human baseline rates the least reliable part of each comparison.
The human baselines were also narrow or mismatched in era. The English speeches, fiction, and argumentative baselines came from nineteenth- and early twentieth-century public-domain works, and the argumentative benchmarks rested on a single book in English (Mill’s On Liberty) and a single historical author in Italian (Cesare Beccaria).
Finally, the mitigation experiments relied on small pilot samples, using 12 texts per genre in each condition for the prompting test and six prompts with six outputs each for the adapter tests. The removable adapter was also trained and tested on a smaller open-weights model rather than the commercial models used in the genre survey, and the author checked that the adapter preserved meaning through informal reading rather than formal testing. Future work will need to compare models across different developers, validate detectors across languages, and apply formal meaning-preservation and human appropriateness checks.
The study, “Artificial Epanorthosis: Why large language models overuse a classical rhetorical figure, and how to mitigate it,” was authored by Federico Boggia.
