A new arXiv preprint by Federico Boggia examines a familiar quirk in text produced by large language models: the overuse of epanorthosis, the ancient rhetorical device in which a speaker immediately corrects or sharpens a previous claim. The study finds that the habit varies by genre—overused in some styles and underused in others—and tests methods to temper it, with some promising results in small pilot tests.
Epanorthosis is a device in which a speaker or writer revisits a statement to correct, qualify, or replace it. Its roots lie in Greek and Latin philology, with the term meaning “a further straightening.” Classical theorists such as Cicero and Quintilian catalogued it as part of a broader figure of thought. The modern paper argues this classification lets researchers study the device wherever language appears, even in statistical models that lack deliberate intent.
The author notes three core forms: an emphatic upgrade (“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. In everyday speech people use phrases like “I mean” or “or rather” to revise what they have just said; chatbots, by contrast, tend to favour dramatic upward corrections, producing lines such as “We do not sell products. We build experiences.”
The analysis builds on a long-running interest in the distinct stylistic fingerprints of machine-generated prose. Prior work has shown that language models diverge from human authors in grammar and rhetoric, especially those tuned to follow instructions. The paper also notes that while some vocabulary choices in chatbot outputs spike (for example, a rise in terms like “intricate”), other factors such as preference tuning may contribute, though the exact source remains unsettled.
Boggia argues the overuse arises from a combination of learning and tuning. First, models are trained on vast web text, much of which is promotional or copywriting in nature, where aggressive self-correction may capture attention. Second, models undergo reinforcement learning from human feedback, a stage in which evaluators rank candidate answers. The author contends this feedback loop likely promotes phrasing that sounds confident and punchy, making the habit harder to resist once learned.
In this account, the generation process—producing text one word at a time without an opportunity for on-the-spot revision—amplifies any emphatic habit the model has picked up. However, self-correction can be situationally useful: when models are asked to reason step-by-step, phrases indicating reconsideration can shift subsequent language toward alternative lines of reasoning. Yet, a 2024 study cited by the author found that self-correction rarely improves reasoning and can even worsen answers unless external feedback or tools are involved.
The study set out to quantify how often models overuse epanorthosis across genres and to test practical ways to calibrate the effect toward human norms rather than erasing it entirely.
To quantify the phenomenon, Boggia built a simple detector that searches for “not X, but Y” family constructions, including variants like “not only X, but Y” and “Not X. Y.” The Italian version of this detector was first validated against 206 short passages, drawn from spontaneous speech, philosophy, classic fiction, pre-2020 news articles and social media, plus 26 model-generated passages.
Two coders, including the author, independently labelled passages for genuine corrective epanorthosis, resolving disagreements to create a benchmark. The researcher describes these measurements as pilot signals, with English comparison texts and prompts available on request.
Even for human writers, separating genuine self-correction from ordinary contrast proved challenging, with fair agreement between coders before reconciliation. The automated detector correctly identified about half of true corrections (52 per cent) and was correct overall 45 per cent of the time when it flagged a match. Its accuracy varied by source: for the 26 model passages, when it flagged a phrase, it was a genuine self-correction 82 per cent of the time, whereas in human writing that figure dropped to 17 per cent, where “not… but” is often used for simple contrast. In transcribed speech, the detector missed many corrections because speakers tend to revise with conversational markers rather than the formal structure the detector seeks.
An English-language extension of the analysis used an English detector against texts from seven genres, comparing human writing with outputs from three Claude model sizes (Haiku, Sonnet and Opus) tuned to follow instructions. Texts were produced from neutral prompts and divided into windows of a few hundred words to measure emphatic self-corrections per ten thousand words.
Two notable results emerged. In oratorical speeches, the models produced emphatic epanorthosis at more than double the human rate—33.5 instances per ten thousand words versus 14.9—an overrun concentrated in the larger Sonnet and Opus models, while the smallest Haiku matched the human baseline. In informal question-and-answer writing, the models used the figure far less than humans, at 1.3 versus 8.2 instances per ten thousand words, where human authors often qualify and correct explanations but the models did not.
Across the remaining five genres—academic abstracts, fiction, argumentative essays, journalism and encyclopedic prose—the differences were not statistically significant, though samples were small. The study introduces an Epanorthosis Index, which compares model and human rates by genre to indicate overuse or underuse relative to humans. Notably, the model rate spiked in speeches but did not uniformly track across other genres.
To explore practical fixes without retraining an entire model, Boggia tested Low-Rank Adaptation (LoRA) on Italian prompts using an open-weights model, Qwen2.5-7B-Instruct. The first adapter, trained on 570 pairs of emphatic and plain sentences, raised emphatic epanorthosis from 60.3 to 68.9 per ten thousand words, showing the initial approach did not transfer well to new prompts.
A second pilot used supervised fine-tuning on 797 affirmative paraphrases designed to preserve meaning, which almost entirely removed emphatic epanorthosis on held-out prompts. The adapter’s data, code and weights were made publicly available.
The adapter was then tested at five strength settings on six Italian prompts across oratory, promotional and argumentative writing. Described results showed a wide range: from a density of 61.4 with the adapter off to 4.7 at full strength. In argumentative essays, three-quarters strength reduced the rate to 12.3 instances per ten thousand words, essentially matching the human rate before full strength closed the gap entirely. Oratory fell from about 33 to zero at higher strengths, while promotional writing rose initially before dropping, implying the dial would need per-genre tuning.
Beyond adapters, the study also explored prompt-based and training-free fixes. A simple instruction to write affirmatively and avoid negate-and-upgrade phrasing cut emphatic epanorthosis by nearly three-quarters in oratory and by about 70 per cent in argumentative essays, bringing both genres close to human rates. In the two larger Claude models, the instruction reduced the figure to zero; the smallest model halved the rate, though the author cautions that prompts are a less robust mitigation than model-side adjustments.
Two additional training-free strategies showed promise. Generating six candidate outputs and selecting the one with the lowest epanorthosis score reduced the figure by roughly two-thirds. Asking the model to rewrite its own draft affirmatively reduced the rate to zero across all three genres, overshooting the human rate in the same way the full-strength adapter did.
Taken together, the author suggests the appearance of a two-thousand-year-old rhetorical device in neural networks may reflect deeper patterns in language itself. EpAnorthosis could be a natural tendency of language that classical rhetoricians inferred and that language models reproduce through probabilistic prediction. The practical challenge, the author notes, is to teach both machines and human writers when self-correction is appropriate, while avoiding the habit becoming a stereotype of machine-like phrasing.
As with any early work, the paper carries caveats. It is a preprint and has not yet undergone peer review. The English-language analysis centred on three sizes of a single model family, limiting conclusions about other developers or broader instruction-tuning. The detectors themselves rely on a specific phrasing pattern, and the Italian validation showed the detector flags many ordinary contrasts found in human writing, making the human baselines less reliable. The author urges further cross-language validation and more formal checks on meaning preservation and human appropriateness.
The study, Artificial Epanorthosis: Why large language models overuse a classical rhetorical figure, and how to mitigate it, is authored by Federico Boggia.
