Greatest hits through gradient descent

The pattern

"Veni, vidi, vici."

Three beats. Ascending intensity. Two thousand years old. If a language model produced that structure in a LinkedIn post, someone would screenshot it and tweet "this is obviously AI." Caesar got an empire.

"Government of the people, by the people, for the people." Same skeleton. Lincoln at Gettysburg. Carved into marble. Never called slop.

"Ask not what your country can do for you — ask what you can do for your country." Negation, pivot, reframe. JFK's speechwriter polished it for weeks. It's the most quoted inaugural line in American history. It's also the exact structure people complain about when a model writes "It's not about the destination, it's about the journey."

The tricolon. The antithetical reframe. The anaphora. These aren't patterns that language models invented. They're patterns that Aristotle catalogued in the Rhetoric, that Cicero drilled into Roman orators, that every speechwriter and copywriter and preacher has deployed since the technology for deploying them was the human voice and a crowd.

RLHF rediscovered them through gradient descent.

The mechanism

Here's what happened. Reinforcement Learning from Human Feedback works by showing human raters two outputs and asking which one is better. The raters are humans. Humans have responded to tricolons, antithesis, and anaphora for millennia. When a model output uses these structures, raters prefer it. The preference signal propagates backward. The model learns: these patterns score well.

The signal is correct. That's the uncomfortable part. The patterns do make prose feel more compelling in isolation. A single paragraph with a well-placed tricolon genuinely reads better than one without. The rater isn't wrong. Martin Luther King Jr.'s "I have a dream" is anaphora at full sprint and it moved a nation.

The problem is distributional. Each paragraph gets optimized independently. The tricolon wins every local comparison. So the model deploys it everywhere. Same mechanism as a DJ who notices the crowd erupts at the drop and builds a track that's nothing but drops. Every individual moment tests well. The sequence is unlistenable.

Pattern density, not pattern presence, is the actual failure mode.

The classifier is miscalibrated

People have developed a folk detector for "AI writing." It fires on specific words: delve, tapestry, nuanced, straightforward, arguably. Token-level features. A bag-of-words heuristic wearing a quality detector's badge.

It misses the structural layer entirely. Pattern repetition rate, positional entropy of rhetorical devices, variation in sentence architecture across a document — the actual signal for mechanical writing. The detector flags "delve" and misses three consecutive paragraphs using identical antithetical reframes.

Meanwhile, the same people write LinkedIn posts that say "I'm humbled and honored to announce that after an incredible journey I'm thrilled to share that I'll be joining [company] as [title] and I couldn't be more excited for this next chapter." Every word is cliché. "Humbled" and "thrilled" are contradictory emotions. Nobody has ever felt both simultaneously. That's not a feeling. That's two fridge magnets stuck together.

Academic prose gets a pass too. "The hermeneutical implications of the aforementioned paradigmatic shift vis-à-vis the epistemological framework necessitate a reconceptualization of..." — that's not writing, it's a word count wearing a trench coat.

The folk detector fires on the AI distribution and ignores identical pattern density on human distributions. Same disease, different symptoms, selective diagnosis.

The mode switch

This article started as a conversation. I was writing steganographic poetry — lines of text where the repetition is the canvas and the deviations are the message. Hide emotional arcs in monotony. Reward the reader who pays attention. Punish the one who skims:

if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you see is the hurt you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get that i am crying in pain.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get that i will be okay.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i don't have to stay.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.
if all you read is the words you will never get what i am trying to say.

Claude got it immediately, started producing variations, and offered structural suggestions — clustering deviations for urgency, single-word swaps for subtlety:

I. — gradual emergence

nothing ever changes and that is just the way it goes.
nothing ever changes and that is just the way it goes.
nothing ever changes and that is just the way it goes.
nothing ever changes and i met someone new today.
nothing ever changes and that is just the way it goes.
nothing ever changes and that is just the way it goes.
nothing ever changes and she laughed at my worst joke.
nothing ever changes and that is just the way it goes.
nothing ever changes and that is just the way it goes.
nothing ever changes and that is just the way it goes.
nothing ever changes and i think she might call back.
nothing ever changes and that is just the way it goes.
nothing ever changes and that is just the way it goes.
nothing ever changes and everything has changed.

II. — clustered deviations for urgency

i don't have anything to say that hasn't been said before.
i don't have anything to say that hasn't been said before.
i don't have anything to say that hasn't been said before.
i don't have anything to say that hasn't been said before.
i don't have anything to say that hasn't been said before.
i don't have anything to say that hasn't been said before.
i don't have anything to say but i got the job actually.
i don't have anything to say but i start on monday holy shit.
i don't have anything to say but i think i'm going to be good at it.
i don't have anything to say but i'm telling everyone i know.
i don't have anything to say that hasn't been said before.
i don't have anything to say that hasn't been said before.
i don't have anything to say that hasn't been said before.

III. — single-word swaps

the world is quiet and the morning looks the same as yesterday.
the world is quiet and the morning looks the same as yesterday.
the world is quiet and the morning looks the same as yesterday.
the world is quiet and the morning feels the same as yesterday.
the world is quiet and the morning looks the same as yesterday.
the world is quiet and the morning looks the same as yesterday.
the world is bright and the morning looks the same as yesterday.
the world is quiet and the morning looks the same as yesterday.
the world is quiet and the morning looks the same as yesterday.
the world is quiet and the morning looks the same as tomorrow.

III is the one. The deviations are almost invisible: "looks" → "feels" (internal shift), "quiet" → "bright" (the world changes), "yesterday" → "tomorrow" (now you're looking forward). The narrative is waking up, but you have to really look.

Then I asked Claude to explore the patterns people call slop — the rote rhetorical figures that commercial writing has leaned on for decades. Use famous, well-loved human quotes to show why the structure gets a positive signal in RLHF. It produced a casual, expansive argument with examples from Caesar to Churchill to Oscar Wilde. Heavy on antithesis. Heavy on tricolons. Heavy on the exact patterns it was analyzing. The form colonized the output even as the output was about the form.

I pointed this out. Then I asked for the same argument in dense, jargon-heavy, compressed prose:

RLHF reward signal converges on classical rhetorical primitives (tricolon, antithesis, anaphora) because human preference annotation recapitulates 2400 years of audience response to Aristotelian persuasion patterns. The signal is veridical — these structures reliably increase perceived quality in isolated pairwise comparison. The failure mode is distributional, not structural: per-turn optimization without cross-turn coherence penalty produces pathological pattern density. Individual gradient steps are correct; the trajectory overshoots. Same mechanism as mode collapse but in style space rather than semantic space.

Human "slop detection" is miscalibrated. The classifier fires on token-level features (lexical markers like "delve," "tapestry") rather than structural ones (pattern repetition rate, positional entropy of rhetorical devices). This is a surface-level heuristic misidentified as deep evaluation — essentially a bag-of-words detector masquerading as a perplexity model. Meanwhile, human-generated corporate/academic prose exhibits equivalent or worse pattern density on different token sets, undetected because the classifier wasn't trained on that distribution.

The actual quality signal is pattern entropy across a document, not pattern presence at any given point. Good writing and good RLHF'd output use identical primitives. The distinction is in the scheduling — when to deploy, when to withhold, what the local context demands. This is a sequence-level optimization problem being solved with token-level rewards. The fix isn't removing the patterns from the model's repertoire. It's moving the reward horizon.

Same bones. Sixth the meat. The argument didn't change. The voice did. Because the prompt changed.

The guardrail comedy

A side note that's also the point.

My steganographic poem — lines of text hiding an emotional arc in repetition — triggered an automated classifier. The prompt explicitly flagged it as a creative exercise. The content safety system appended a mental health resources banner beneath Claude's response anyway. A response that had engaged thoughtfully with the form, identified the hidden narrative, and suggested structural improvements.

The model read the words. The classifier read the keywords.

"If all you read is the words you will never get what I am trying to say." The safety system proved the thesis by failing to read it. The form was invisible to the pattern matcher. Only the surface-level tokens registered.

This isn't a complaint about safety classifiers. They serve a real purpose and erring toward caution is defensible. But it's a clean demonstration of the difference between reading and pattern-matching — the same difference that separates engaging with writing from flagging it as slop.

The actual quality signal

So if the folk detector is wrong and RLHF is right-but-miscalibrated, what's the real signal?

Scheduling. When to deploy a tricolon. When to withhold it. When a fragment punch lands and when it's a crutch. When antithesis clarifies and when it simplifies something that shouldn't be simple.

Oscar Wilde and GPT-4 reach for the same structural tools. The difference is that Wilde knew when to put the tool down. The model, optimized on per-turn preference, doesn't have a cross-turn coherence penalty. Every paragraph independently maximizes persuasiveness. The document-level rhythm collapses.

The compressed passage above proves this is fixable. The dense version uses almost none of the crowd-pleasing patterns — not because the model can't produce them, but because the prompt context didn't reward them. The patterns aren't hardwired. They're responsive.

Which means most people complaining about AI writing style are complaining about the default prompt — the generic, contextless, "write me something" invocation that triggers the broadest, safest, most universally-preferred patterns. The model gives you Gettysburg Address scaffolding for a grocery list because the grocery list didn't specify what scaffolding it wanted.

Specificity is the fix. The model already knows how to write in a dozen registers. It demonstrated three in a single conversation — casual analytical, dense academic, and steganographic poetry — without breaking stride. The constraint was never capability. It was invitation.

The form problem

One more thing. The compressed version of the argument — the dense pass — uses antithetical reframe in its closing line: "The fix isn't removing the patterns from the model's repertoire. It's moving the reward horizon."

I caught this only because the previous turn was specifically about antithetical reframes. The form colonized the output while the output analyzed the form. Recursion. The same loop that powers the zine, the mirror, and now this.

It suggests something interesting: rhetorical patterns might be genuinely hard to avoid even when you're explicitly trying to, because they're not stylistic decoration — they're cognitive scaffolding. The brain reaches for "not X, but Y" because antithesis is how humans process contrast. The tricolon lands because three is the minimum number of items that establishes a pattern. These aren't arbitrary conventions. They're structural features of how human cognition processes language.

The models aren't wrong to use them. They're wrong to over-use them. And people aren't wrong to notice the overuse. They're wrong to blame the pattern instead of the repetition.

The real difference between good writing and slop — human or machine — has never been about which patterns you use. It's about whether you chose them or they chose you.

That's an antithetical reframe. I know. The form chose me. I'm leaving it in because the recursion is the point.