You scroll through LinkedIn and see: "This isn't a course. It's a transformation journey." Then: "We don't sell products. We craft experiences." Your instinct says: AI wrote this. You're right—but what you may not know is that this "not X, but Y" pattern has a 2,000-year-old name: epanorthosis, a classical rhetorical figure catalogued by Cicero and Quintilian in ancient Rome. Why would a language model that knows nothing about rhetoric systematically reproduce a technique from two millennia ago?
Federico Boggia's July 2026 paper, *Artificial Epanorthosis*, digs to the bottom of this phenomenon.
A Tell You Can't Unsee
Epanorthosis comes from Greek *epi* (upon) plus *anorthosis* (correction)—literally "correcting again." The speaker makes a statement, then immediately revises it, replacing the original word with something stronger, more solemn, or more elevated. Classical rhetoric distinguishes three uses:
- Emphatic correction: replacing a word with a stronger one — "He is good, nay, excellent."
- Mitigated correction: replacing with a milder one — "I'll kill you—or, well, at least scold you."
- Redefining correction: overturning and restating the whole claim — "This isn't a course, it's a transformation journey."
- Speeches: models overshoot ~2x (nearly 3x in Italian), concentrated in larger model tiers
- Informal Q&A: models undershoot, using it less than humans
- Argumentation, news, encyclopedia: roughly human-level
The third is the LLM favorite. Boggia notes it has proliferated across chatbot answers, LinkedIn posts, and marketing copy—"impossible to unsee once seen."
The Cause: Not Generation Mechanics, but Training Preferences
The natural guess is that it's a byproduct of autoregressive left-to-right generation: the model says a word, then revises it. But Boggia argues this is only an amplifier, not the root cause. There are two real sources:
First, the training distribution is saturated with marketing prose. The highest-density "not X, but Y" corpora on the internet are ads, marketing copy, LinkedIn posts, and TED talk scripts. The model learned that this pattern "sounds powerful."
Second, RLHF rewards it. Preference tuning has human annotators scoring confident, emphatic, compelling answers highly. Epanorthosis is the cheapest tool for manufacturing emphasis—one sentence upgrades a mundane statement into a quotable line. Annotator preferences become internalized generation tendencies.
The evidence: base models show lower epanorthosis density than instruction-tuned models, and density rises with model size. If it were a byproduct of generation mechanics, base models would show it equally. Instead, the later the training stage, the heavier the tell—training reinforces it; the architecture doesn't produce it.
The Epanorthosis Index: Turning "Vibes" into Numbers
One of Boggia's core contributions is a measurable metric: the Epanorthosis Index—the model's epanorthosis density in a given genre, divided by the human density in the same genre. An index of 1 means human-equivalent; above 1 means overuse.
Measured across one instruction-tuned model family at three sizes, the results reveal a genre-specific mismatch:
Mitigation: One Prompt Line Halves It, LoRA Nearly Eliminates It
Boggia tested three mitigations, experimenting in Italian:
A single instruction: adding "do not use 'not X, but Y' corrective constructions" to the prompt cut density by 50–75%. Blunt but effective—the tendency is instruction-controllable.
LoRA fine-tuning: a small supervised adapter trained to reduce epanorthosis nearly eliminated it. Crucially, LoRA's scaling coefficient acts as a knob to tune the reduction precisely back to the human baseline—not elimination, but calibration.
The goal is calibration, not elimination: Boggia stresses this repeatedly. Epanorthosis isn't bad—Cicero used it, Churchill used it, good oratory requires it. The problem is runaway density in contexts that don't need it. The right target is returning model density to the human baseline in every genre.
The Real Risk: We Start Writing Like Machines
The paper's final line is worth quoting: "The real risk is that we start writing like machines."
This isn't alarmism. As LLM text floods the internet—students using ChatGPT for essays, marketers for copy, job seekers for cover letters—epanorthosis density will compound through a feedback loop of "humans imitating AI text." We already see LinkedIn posts resembling chatbot output: it's not humans using epanorthosis, it's humans imitating the machine's epanorthosis.
Boggia's paper does something rare: it turns a "something feels off" phenomenon into a measurable, attributable, mitigable scientific problem. The Epanorthosis Index is a simple tool, but it converts "AI flavor" from vague feeling into a concrete number. Once you can measure it, you can control it.
If the rhetoricians of two millennia ago knew their catalogued technique was being systematically reproduced by a statistical model that understands no rhetoric, they might find it an epanorthosis in itself: it's not humans using rhetoric—it's rhetoric using humans.
---
Paper: Federico Boggia. *Artificial Epanorthosis: Why large language models overuse a classical rhetorical figure, and how to mitigate it.* arXiv:2607.21498, 2026. Link: https://arxiv.org/abs/2607.21498