AI's Overuse of Ancient Rhetorical Self-Correction
Large language models frequently employ epanorthosis, a 2,000-year-old rhetorical device, creating distinct stylistic patterns.
1The Epanorthosis Habit
Artificial intelligence models frequently use epanorthosis, a rhetorical device where a statement is immediately qualified or replaced. Common examples include phrases like, 'This is not a course; it is a journey.' This ancient figure, cataloged by Roman orators like Cicero, appears as a persistent stylistic habit in modern machine-generated text.
2Defining the Rhetorical Move
Epanorthosis functions through emphatic upgrades, softening downgrades, or substantive revisions. While humans use conversational markers like 'I mean' to refine speech, chatbots favor dramatic, punchy self-corrections. Researcher Federico Boggia argues that statistical programs perform this structural move for emphasis, regardless of whether the model possesses any genuine persuasive intent.
3Origins of Machine Prose
A 2025 study indicates that large language models diverge from human writing patterns, a gap that widens as models grow larger. Boggia suggests this habit stems from training on promotional web text and reinforcement learning from human feedback, where raters reward confident, dramatic phrasing that sounds authoritative to human readers.
4Measuring Stylistic Divergence
Boggia developed an automated detector to identify 'not X, but Y' patterns in Italian and English. Validated against human-labeled passages, the tool proved highly accurate for model outputs but struggled with human speech. Humans typically use conversational fillers rather than the formal structures that chatbots rely on for their emphatic corrections.
5Genre-Specific Model Behavior
Testing across seven genres revealed that models use epanorthosis at double the human rate in oratorical speeches. Conversely, models underused the figure in informal question-and-answer writing compared to human baselines. These findings suggest that stylistic habits are not uniform but vary significantly depending on the specific context and genre of text.
6Statistical Significance Limitations
Because the study tested seven genres simultaneously, the author notes that individual results for speech and question-and-answer writing would not remain significant under strict statistical corrections. The findings are presented as pilot-scale signals. Further research is required to determine if these patterns persist across broader, more diverse datasets and model architectures.
7Fine-Tuning with Adapters
Boggia tested Low-Rank Adaptation to adjust this habit without retraining models from scratch. By training an adapter on 797 affirmative paraphrases, the researcher created a 'stylistic dial.' At full strength, this adapter nearly eliminated emphatic epanorthosis in test prompts, demonstrating that model behavior can be calibrated to match human rates effectively.
8Calibrating the Stylistic Dial
The adapter acted as a continuous control, with intermediate settings successfully matching human rates in argumentative essays. However, the dial required genre-specific calibration, as promotional writing showed unpredictable spikes at lower settings. This suggests that a single universal setting may not be sufficient to normalize stylistic habits across all writing categories.
9Prompting as a Mitigation
For users unable to modify model weights, Boggia tested simple prompt instructions. Instructing models to write affirmatively cut epanorthosis by nearly 75 percent in oratory and 70 percent in argumentative essays. While effective, the author warns that prompting is less robust than fine-tuning because surface instructions may not override deep-seated training habits.
10The Role of Reasoning
Self-correction is not always a superficial habit; it can shift the statistical context for reasoning. However, a 2024 study found that models often fail to improve their logic through self-correction without external feedback. Without verified facts or tools, internal revisions may sometimes lead to worse outcomes rather than more accurate reasoning.
11Human-Machine Stylistic Gap
The research highlights that machine-generated prose is not just different in content, but in rhetorical structure. By identifying these specific patterns, developers can better understand how reinforcement learning shapes model personality. Moving forward, the ability to calibrate these stylistic dials could lead to more natural, less repetitive, and more nuanced AI writing.
12Future of Stylistic Control
The study provides a framework for measuring and mitigating unwanted stylistic habits in large language models. By treating rhetorical figures as measurable quantities, researchers can move toward more flexible AI outputs. Future work will likely focus on whether these techniques can be scaled to larger, proprietary models without compromising their core performance.
Original article: PsyPost — AI's Overuse of Ancient Rhetorical Self-Correction