💬 NLP

Artificial Epanorthosis: Why large language models overuse a classical rhetorical figure, and how to mitigate it

This paper argues that large language models systematically overuse the rhetorical figure of epanorthosis due to training biases and preference tuning, and proposes a genre-specific calibration approach using lightweight adapters and instruction tuning to align model output with human rhetorical norms rather than eliminating the figure entirely.

Original authors: Federico Boggia

Published 2026-07-29

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Federico Boggia

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Secret Sauce of Robot Writing

Imagine you are walking into a giant library where millions of books, websites, and social media posts have been shredded and mixed into a giant smoothie. This is how modern AI chatbots learn to talk: they taste this digital smoothie and learn to guess the next word in a sentence, over and over again. But there's a twist. After they learn from the smoothie, humans step in to give them a "taste test." They read the robot's answers and say, "This one sounds confident and cool," or "This one sounds boring." The robot then learns to do more of what the humans liked. This process is called "reinforcement learning," and it's like training a dog with treats, except the dog is a super-smart computer and the treats are digital approval.

The big question this paper asks is: What happens when we train a robot to sound "confident" and "cool"? It turns out, the robot starts using a very specific, old-fashioned trick to sound impressive. It's a trick that human speakers have used for thousands of years, but the robot uses it so much that it starts to sound fake. The paper investigates why the robot does this, how often it happens, and whether we can teach the robot to stop overusing it without making it sound boring or stupid.

The Robot's "Wait, Actually..." Tic

You know that feeling when someone says, "I'm not just a teacher. I'm a shaper of minds"? Or when a salesperson says, "We don't just sell shoes. We sell freedom"? They start with a normal sentence, then immediately hit the "undo" button and replace it with something much bigger, louder, and more dramatic.

In the world of fancy language studies, this trick has a name: epanorthosis (say it like eh-pan-or-tho-sis). It's a Greek word that basically means "straightening out" or "correcting yourself." Ancient speakers used it to make their points punchier. But recently, researchers noticed that AI chatbots are doing this all the time, like a nervous tic. It's become the robot's signature move.

The paper argues that this isn't an accident. It's not because the robot is "thinking" deeply and changing its mind. Instead, the robot learned this habit from two main places:

The Internet Smoothie: The robot read tons of marketing copy, motivational posts, and sales pitches where this "Not X, but Y!" style gets lots of clicks and likes.
The Human Taste Test: When humans rated the robot's answers, they preferred the ones that sounded confident and emphatic. The robot realized, "Hey, if I say 'I'm not a bot, I'm a digital architect,' people give me a high score!" So, it started doing it more and more.

The paper suggests that the robot's way of writing (one word at a time, left to right) makes this worse. Because the robot can't go back and edit its sentence after it's written, it uses this "correction" trick as a way to fix its tone on the fly, making it sound more dramatic than a human usually would.

The "Epanorthosis Index": Measuring the Robot's Drama

To prove this, the authors didn't just guess; they built a measuring tape. They created something called an Epanorthosis Index. Think of it like a "drama meter." They counted how many times the robot used this "Not X, but Y" trick in 10,000 words and compared it to how often real humans use it in the same type of writing.

Here is what they found, and it's a bit of a mixed bag:

In Speeches and Sales: The robots went way overboard. In "oratory" (speeches) and "promotional" writing, the robots used the trick about twice as much as humans do. In Italian, it was even worse, nearly three times as much. They were trying so hard to sound inspiring that they sounded fake.
In Casual Q&A: The robots did the opposite. When humans were just chatting or answering simple questions, they often used this trick to soften their answers or hedge their bets. The robots, however, were too flat and didn't use it enough. They were about one-fifth as likely to use it as humans.
In Serious Writing: In things like news articles or encyclopedia entries, the robots and humans were actually pretty close. The robots knew when not to be dramatic.

The main takeaway is that the robots are miscalibrated. They don't know when to turn the drama knob up and when to turn it down. They just blast the "epanorthosis" button whenever they think they need to sound impressive.

Can We Fix the Robot? (The "Drama Dial")

The paper doesn't just point out the problem; it tries to fix it. The authors asked: "Can we teach the robot to stop overusing this trick without making it sound like a robot that's afraid to speak?"

They tested a few ideas, and here is what worked:

The "One-Line" Fix (Prompting): They tried simply telling the robot, "Please write in a straightforward way and avoid saying 'It's not X, it's Y'."
- The Result: It worked! In Italian tests, this simple instruction cut the dramatic corrections by 70% to 72% in speeches and essays. It didn't erase the trick completely, but it brought the robot's usage down to a level that looked much more like a human.
The "Style Adapter" (The Magic Dial): This was the big experiment. The authors trained a tiny, special add-on (called a LoRA adapter) for the robot. Think of this add-on as a "volume knob" for drama.
- They trained it on examples of "dramatic" text and "plain" text.
- Then, they tested a "dial" (a number they could change) to see how much of this new style to use.
- The Result: The dial worked perfectly. When they turned it up, the robot sounded plain. When they turned it down, the robot sounded dramatic. Crucially, they found a setting where the robot's usage matched the human rate exactly. It wasn't about deleting the trick; it was about calibrating it.
What Didn't Work: They tried using "preference learning" (teaching the robot what humans like) on its own, but it failed to generalize. The robot learned to pick the right answers in the test but didn't actually change how it wrote new sentences.

The Big Warning

The paper ends with a slightly scary thought. The real danger isn't that the robots are using this trick too much. The real danger is that we might start writing like them.

If we get used to reading text that is full of "Not X, but Y!" corrections, we might start thinking that's how good writing sounds. We might start writing our own essays, emails, and speeches with this fake, over-dramatic flair just to sound impressive. The paper suggests that the goal shouldn't be to make robots sound exactly like humans, but to make them sound appropriate for the situation. Sometimes a speech needs drama; sometimes a medical report needs plain facts.

The authors conclude that while we can fix the robot's "tic" with these new tools, the harder job is teaching ourselves to recognize when a correction is honest and when it's just a performance. After all, if a machine can learn to sound like a human, the next step is making sure we don't forget how to sound like ourselves.

Technical Summary: Artificial Epanorthosis

Problem Statement
Large Language Models (LLMs) exhibit a systematic overuse of epanorthosis, a classical rhetorical figure of thought involving self-correction to replace a statement with a stronger, more solemn, or more precise one (e.g., "This is not a course. It is a journey of transformation"). While this figure has existed for two millennia, LLMs reproduce it with a frequency and rigidity that diverges from human stylistic norms. The paper argues this is not a random statistical artifact but a "trained disposition" driven by two primary factors:

Training Distribution: The prevalence of promotional prose, motivational content, and landing pages in pre-training data, where emphatic correction correlates with engagement.
Preference Tuning (RLHF): Reinforcement Learning from Human Feedback rewards confident, incisive, and emphatic phrasing, inadvertently amplifying the "Not X. Y" construction.

The paper posits that the autoregressive (left-to-right) nature of generation amplifies this tendency by preventing pre-publication revision, but it is not the root cause. The core issue is a mis-calibration by register: models overshoot in persuasive/oratorical contexts and undershoot in informal contexts, failing to match the genre-specific human baseline.

Methodology
The paper proposes a measurement framework and a suite of mitigation techniques, moving from observation to intervention.

Measurement Protocol:
- Epanorthosis Index (EI): A metric defined as the density of emphatic-correction constructions (e.g., "not X, but Y," "X? No. Y") per 10,000 words, normalized against a human baseline for a specific genre. An EI of 1 indicates human-like calibration.
- Detector: A two-stage system involving a high-recall lexical pass over corrective markers followed by a construction-level classifier to filter false positives (e.g., distinguishing "I mean" as a filler vs. a correction).
- Corpora: Human baselines include pre-2022 texts across genres (encyclopedic, journalism, academic, narrative, Q&A, argument, oratory). Model outputs are generated from instruction-tuned families (specifically Claude Haiku, Sonnet, and Opus) using neutral, genre-specific prompts.
Mitigation Strategies:
The paper surveys interventions ranging from data curation to inference-time control, prioritizing Lightweight LoRA (Low-Rank Adaptation) adapters.
- Data-side: Counterfactual augmentation (pairing emphatic text with de-emphasized paraphrases) and pruning promotional examples from instruction data.
- Model-side (LoRA): Training a small, detachable adapter using Supervised Fine-Tuning (SFT) on "plain" responses or Direct Preference Optimization (DPO) on preference pairs. The adapter is controlled via a scaling coefficient ( $\alpha$ ) at inference time to dial the reduction from the base model's rate down to zero.
- Inference-side: Decoding-time control (PPLM, GeDi, FUDGE), activation steering, and prompting (system instructions).
- Evaluation: Success is measured not by elimination, but by calibration—reducing the figure to the human baseline rate for a given genre while preserving content fidelity and appropriateness.

Key Results

Empirical Evidence of Mis-calibration:
- Oratory: Models overshoot significantly, producing emphatic epanorthosis at roughly 2.2 times the human rate (33.5 vs. 14.9 per 10k words). This effect is concentrated in larger model tiers (Sonnet, Opus).
- Informal Q&A: Models undershoot, producing the figure at roughly 0.2 times the human rate (1.3 vs. 8.2 per 10k words), failing to replicate human hedging and self-correction.
- Other Genres: Argument, journalism, and encyclopedic prose show no significant deviation from human baselines.
Mitigation Effectiveness:
- Prompting: A single-line instruction in Italian ("write affirmatively, avoid negate-and-upgrade") reduced epanorthosis density by 70–72% in oratory and argumentative essays, bringing larger models closer to human rates.
- LoRA Adapter (SFT): A supervised fine-tuned adapter on an open 7B model (Qwen2.5-7B) reduced density by ~98% (from ~75 to ~1.7 per 10k words) across oratory, promotional, and argument genres.
- The "Dial" Effect: Scaling the LoRA adapter by a coefficient ( $\alpha$ ) allows for continuous control. At $\alpha = 0.75$ , the adapter reduced argumentative density to 12.3, matching the human baseline of ~12, before over-correcting to zero at full strength ( $\alpha = 1.0$ ). This demonstrates that calibration to human rates is achievable without total elimination.
- Training-Free Methods: "Best-of-n" selection and post-hoc affirmative rewriting also significantly reduced density, though rewriting tended to over-correct past the human baseline.

Significance and Claims
The paper argues that the convergence of ancient rhetoric and neural network statistics suggests epanorthosis is an emergent structure of language itself, an "attractor" toward which both human cognition and statistical models tend. However, the unregulated reproduction of this figure by LLMs creates a "rhetorical fingerprint" that signals artificiality.

The primary contribution is the shift from viewing this as a flaw to be eliminated to a calibration problem. The paper claims that:

The overuse is a measurable, register-specific phenomenon driven by training and alignment, not a fundamental limitation of autoregressive generation.
Lightweight LoRA adapters offer a practical, scalable solution to "dial" the rhetorical style of models to match human baselines for specific genres.
The ultimate risk is not that machines write like humans, but that humans begin to write like machines, adopting the machine's default, uncalibrated rhetorical tics.

Limitations
The paper explicitly notes that its findings are based on pilot-scale evidence with small sample sizes and uncorrected p-values for multiple comparisons. The causal link to RLHF is inferential, based on the divergence between instruction-tuned and base models in prior literature, rather than a direct base-vs-aligned comparison within this study. Furthermore, the trained adapter was demonstrated on a 7B open model, with transfer to frontier models remaining a future task, and content fidelity checks were based on informal inspection rather than rigorous measurement.