Artificial
epanorthosis

Try reading, carefully, the next answer ChatGPT gives you, or the next motivational post scrolling through your LinkedIn feed. There is a structure that, once you learn to spot it, you will never be able to unsee:

The model states something and, in the same breath, negates it to replace it with something bigger, stronger, more solemn. It's a tic so widespread that it has become a kind of signature of AI-generated text. And it has a name, Greek, over two thousand years old, one that Cicero and Quintilian used with full awareness: epanorthosis.

This article is here to do two concrete things: to recognise text written by a machine, and to avoid ending up writing like a machine. To get there, though, it helps to first understand where that tic comes from.

01What epanorthosis is

The word comes from the Greek epanorthosis: literally, “to straighten further”. The speaker returns to what they have just said in order to correct it. It isn't a mistake being amended, though: it's a deliberate move. You go back to reinforce, to soften or to make more precise.

There are three kinds. The upward correction, which replaces a term with a stronger one (“it's good, or rather excellent”). The downward one, which softens (“it's a disaster, well… a hiccup”). And the genuine kind, which changes the substance (“he was running, or rather walking briskly”).

This isn't material from a manual for orators: it's one of the most common figures in everyday speech. Every “I mean”, “or rather”, “better yet” is an epanorthosis made without thinking about it. “I'll be there in an hour, or rather, maybe two.” The problem, with machines, isn't that they use it: it's that they use it too much, and almost always in the upward version, the one that inflates.

02From Cicero to ChatGPT

Classical rhetoric grasped its power early. In Book IX of the Institutio Oratoria Quintilian catalogues it under several Latin names (correctio, emendatio, reprehensio) and separates corrections that touch the thought from those that touch only the wording. One move will matter later on: paenitentia dicti, the simulated regret for something just said, which made the speech look simple and unprepared, and the speaker less calculating in the eyes of the judge. A clear example comes from Cicero, in the first Catilinarian:

Hic tamen vivit. Vivit? immo vero etiam in senatum venit.

Cicero, In Catilinam I, 2 · “Yet he lives. Lives? Nay, he even comes into the Senate.”

First the statement, then the upward correction: that “nay” (in Latin immo) retracts the “lives” as if it were a trifle and raises the stakes. It's the same move as the everyday “or rather”, used with real craft.

There is one detail, though, that lets us move from Cicero to ChatGPT without cheating. A nineteenth-century French rhetorician, Pierre Fontanier, classified epanorthosis not as a figure of style but as a figure of thought: an operation of discourse, not an ornament. The consequence is practical: you can find it wherever there is language, including the answers of a statistical model, and you can study it without asking what whoever produced it “meant”. Which is exactly what we need in order to talk about a machine.

03Why machines fall for it

Machines overuse it mainly for two reasons, which have more to do with how they are built than with how they write. The first is the data: enormous quantities of text from the web, with a big slice of copywriting, motivational posts and landing pages, that is, precisely the genres in which “Not X. Y” collects likes. The model has learned that this form works. The second, and probably the most important, is the fine-tuning with human judgements (RLHF, in the jargon): the people who train it reward answers that sound confident and passionate, and few things sound as confident as a nice upward correction.

On top of these two pressures comes a third, which doesn't create the tic but leaves it on display. A model writes one word at a time, from left to right, and it can't go back to delete: there is no backspace key. We plan, jot down a draft, revise and show only the clean result; the model writes live. So, when the learned tendency leads it to correct upward, the correction stays there in writing (“good, or rather excellent”) instead of disappearing in the edit. Left-to-right writing amplifies the phenomenon, it doesn't cause it: if it hadn't learned to inflate, the model would write perfectly well in affirmative form.

04When the correction really matters

So far it's a tic. But there is one case in which the machine's self-correction stops being decoration and becomes useful: reasoning. If you ask a model to proceed “step by step”, genuine second thoughts appear in the intermediate steps (“wait, that doesn't add up, let me start again”). And, curiously, the models that spell out these second thoughts answer more accurately: the form of self-correction works even without a mind thinking it through.

There is a reversal here that says a great deal. Ancient rhetoric taught orators to simulate self-correction in order to seem sincere; today's engineering teaches machines to practise it in order to be more precise. Same form, opposite direction. And it leaves us a distinction that is useful for those of us who write, too: correction is precious when it adds something, and it's fluff when it only serves to put on a show.

05Recognising AI-written text

Here is the practical part. Repeated epanorthosis is only one signal: on its own it isn't enough, but together with the others it forms a fairly recognisable fingerprint. When you read something and ask yourself “did a machine write this?”, check these points.

An honest caveat: none of these signals is proof. We use epanorthosis too, and the most recent models are learning to hide the more obvious patterns. The only genuinely reliable tool remains the head of whoever is reading.

06How not to write like the machines

The flip side matters even more, especially if you write for a living: the same signals that give a machine away, if you have them too, make your own text ring false. Here is how to keep a human voice.

07Figure or artefact? The stakes

A fundamental question remains. When I say “it's good, or rather excellent”, there is a judgement behind it: I evaluated, and I decided that the first word wasn't enough. When a model does it, there is only a statistic behind it: “or rather” follows “good” because in the data that sequence was frequent. The same perfect figure, but with no one intending it. It's Searle's old Chinese Room applied to rhetoric: impeccable outputs, no understanding behind them.

If a statistical system, knowing nothing of rhetoric, spontaneously reproduces the same figures humans have used for millennia, perhaps those figures are emergent structures of language: attractors toward which the medium tends, whether a brain or a network produces it. In that light the classical orators discovered epanorthosis rather than inventing it, as mathematicians discover theorems that were true before they were proved. Models, correspondingly, act as explorers of the space of linguistic possibility, arriving by statistics at structures we reach by cognition. The convergence is evidence that the structure belongs to language itself, independent of any particular speaker.

This reframes the engineering problem. The techniques surveyed there (LoRA adapters foremost, alongside preference optimisation, decoding-time control, and activation steering) can measurably calibrate the figure toward human rates. The deeper task is to teach readers, and models, appropriateness: the sense of when a correction is earned and when it is empty. The real risk points the other way: