Scroll LinkedIn and you keep seeing it: "This isn't a course. It's a transformation journey." "We don't sell products. We craft experiences." Your instinct says: AI-written. Correct. What's surprising is that this "not X, but Y" pattern has a two-thousand-year-old name — epanorthosis — a rhetorical figure catalogued in ancient Rome by Cicero and Quintilian. Why would a language model that knows nothing about rhetoric systematically reproduce a technique that old?
Federico Boggia's July 2026 paper *Artificial Epanorthosis* digs into exactly this.
A pattern you can't unsee
Epanorthosis comes from Greek *epi* (upon) + *anorthosis* (a setting right): literally, "correcting once more." The speaker states something, then immediately amends it with a stronger, weightier word. Classical rhetoric distinguishes three uses:
- Emphatic correction: replace with a stronger word — "He is good, nay, excellent."
- Understatement correction: replace with a milder one — "I'll kill you, well, at least berate you."
- Redefining correction: overturn the whole statement — "This isn't a course, it's a transformation journey."
- Speech register: models overshoot ~2x, nearly 3x in Italian, concentrated in larger model tiers
- Informal Q&A: models undershoot, using it less than humans
- Argumentation, news, encyclopedic text: roughly human-level
The third is the LLM favorite. It is now ubiquitous in chatbot answers, LinkedIn posts, and marketing copy — "once seen, impossible to ignore."
The cause: training preferences, not generation mechanics
The obvious hypothesis — that it's a byproduct of left-to-right autoregressive generation (say a word, then revise it) — is, Boggia argues, only an amplifier, not the root cause.
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 learns that the pattern "sounds powerful."
Second, RLHF rewards it. Preference tuning pushes human annotators to rate confident, emphatic, engaging answers highly. Epanorthosis is the cheapest possible tool for manufacturing emphasis — one sentence upgrades a flat statement into a quotable line. Annotator preferences get internalized as generative tendencies.
Evidence: base models show *lower* epanorthosis density than instruction-tuned models, and density rises with model size. If it were a byproduct of the architecture, base models would show it equally. Instead, the further along the training pipeline, the heavier the trace — training is reinforcing it.
The Epanorthosis Index: turning "vibes" into numbers
Boggia's key contribution is a measurable metric: the Epanorthosis Index — the model's density of epanorthosis in a genre divided by the human density in the same genre. 1 means human-level; 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 cuts it in half; LoRA nearly eliminates it
Three mitigations, tested in Italian:
One-line instruction: adding "Do not use 'not X, but Y' corrective constructions" to the prompt reduces density by 50-75%. Crude but effective — the tendency is steerable by instruction.
LoRA fine-tuning: a small supervised adapter trained to reduce epanorthosis nearly eliminates it. Better yet, LoRA's scaling coefficient acts like a knob, dialing the reduction until density lands exactly at 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 uncontrolled density. The right target is bringing per-genre density back to the human baseline.
The real risk: we start writing like machines
The paper's closing line is worth quoting: "the real risk is that we start writing like machines."
It isn't alarmism. As LLM-generated text floods the internet — students' essays, marketing copy, cover letters — epanorthosis density compounds through a feedback loop of humans imitating AI text. LinkedIn posts increasingly resemble chatbot output not because humans use epanorthosis, but because humans are imitating the machine's epanorthosis.
Boggia's achievement is rare: he turned 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 ignorant of rhetoric, they might call it an epanorthosis in itself: it isn't 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