AI Cliches Are Old Rhetoric on Autoplay

I started this as an audit. I was running through a stack of AI-generated content, flagging the usual tells — the triadic adjectives, the "It's not X, it's Y" framings, the "from cradle to grave" sweeps — and at some point I noticed something that reframed the whole exercise.

Every cliche I was flagging had a name in classical rhetoric.

The "It's not X, it's Y" pattern is paradiastole. The triadic adjective stack is tricolon plus isocolon. The "from X to Y" sweep is merism. The "Bad data leads to bad models leads to bad decisions" chain is anadiplosis. None of these were invented by language models. They were catalogued by Quintilian and Aristotle, used by Shakespeare and Lincoln, and they show up in good human prose all the time. Used once, used right, they are how a sentence becomes memorable for centuries.

What language models do is run them as defaults. The figures stop being load-bearing and start being wallpaper.

Here is the table I ended up with after going through the audit twice and pulling each tic apart.

The mechanism is straightforward when you look at it from inside the model. Triadic lists dominate good persuasive prose, which means triads dominate the training corpus, which means the next-token distribution favors triadic continuations almost everywhere. The structural shortcut for sounding wise is paradiastole, and the model defaults to it because it generates exceptionally low-perplexity completions — the second clause is highly predictable from the first. Anadiplosis is dramatic and easy to chain because each clause primes the next, so the conditional probability of the chain staying intact is high. The model isn't doing anything weird. It is doing exactly what its training rewards. The pathology is that what gets rewarded at the token level looks like wallpaper at the document level.

Forsyth's whole point in The Elements of Eloquence is that figures are load-bearing. You drop in a tricolon when you have something worth landing in three. Antithesis pays off when the contrast itself is the argument. Save anaphora for when momentum has been earned. Figures survive in human writing because they show up rarely and with intent. Volume kills them.

There is also a smaller category in the table that matters separately. "Let's dive in!" has no Greek ancestor. "Great question!" doesn't either. These are not rhetoric. They are pure phatic residue from RLHF — markers of warmth and engagement that got rewarded in human-feedback rounds. They do not need a fix; they need deletion.

What I want to leave you with is the small irony. Every sentence that opens with "It's not just X, it's Y" forces the next person who tries to use real paradiastole to climb out of a hole the LLMs dug. The figures are not broken. The autoplay is.

Keep the figures. Take them off autoplay.