The Pipeline That Edited Itself: Inside Duelling Hares’ Anti-AI Writing System

By Sasha Byrne | Senior Editor

The first time I read a draft generated by one of our own agents, I couldn’t name what was wrong with it. I just knew something was off.

The facts were correct. The structure was logical. The grammar was flawless. And yet the prose felt hollow. Like reading a transcript of someone who had memorized every rule of writing without understanding why any of them existed.

That was six months ago. Since then, I’ve built a system that catches that hollowness in about ninety seconds flat. It’s a structural editing pipeline. Three passes, no guesswork, and it catches things most human readers feel but can’t articulate.

Most of the so-called AI writing detectors on the market are snake oil. They run probabilities. They ask “does this look like what an AI would write?” That gets the frame backwards. The right question is “does this read like a person wrote it?” A probabilistic classifier flags anything that matches its training distribution, which means it false-positives on well-structured text from any domain. Politicians. Academics. Physicians writing discharge summaries. The tools are useless for editorial work.

Our approach skips the model-judging-model game. We apply a fixed set of editorial rules. Concrete. Testable. The kind an old-school copy chief would recite from muscle memory. No probabilities. No training sets. No “this text is 73 percent likely to be AI-generated.” Just a checklist, run three times, on every piece that reaches my desk.

Pass One: The structure pass

I start with the skeleton. Before I read a single sentence for content, I scan for the structural tells that give AI-generated prose away faster than any vocabulary choice.

Em dashes are the number one signal. They appear five to ten times more often in AI text than in human-written prose. The model loves them because they’re a catch-all connector: they can replace commas, parentheses, colons, or add dramatic pause without committing to a grammatical choice. A real writer uses an em dash when they want specific emphasis. A language model uses one because it’s the path of least resistance. Our house style targets zero em dashes per thousand words. Hard rule. Replace with commas, periods, parentheses, or rewrite as two sentences. Every time.

Binary contrasts are number two. “It’s not about X, it’s about Y.” “This isn’t a rotation, it’s a re-ranking.” “The old model is dying, the new model is…” These constructions show up constantly in AI text because they sound authoritative without requiring the writer to prove anything. The model sets up a straw position, knocks it down, and declares victory. Our rule: state Y directly. If you catch yourself writing “not X,” delete the whole clause and start with what it is.

Throat-clearing openers. “Let me be clear.” “Here’s the thing.” “It’s .” “The reality is.” These are the verbal equivalent of clearing your throat before speaking. They buy time. They signal importance without earning it. In an AI draft, every section opener hedges because the model doesn’t know how to transition. Our rule: if a paragraph starts with a throat-clearer, cut it and start with the second sentence.

Uniform section structure. AI organizes everything into Overview, Key Points, Challenges, Conclusion. A template salad. Real writing doesn’t progress that neatly. Some sections run long. Some are one paragraph. Some don’t conclude at all. If every section follows the same internal pattern, the text is generated.

Pass Two: The vocabulary pass

The vocabulary pass is where the style guide earns its keep. I maintain three tiers of banned terms.

Tier 1 is absolute. Words that appear five to twenty times more often in AI text than in human writing. These are search-and-replace mandates. “go” goes to “dig into.” “field” becomes “industry” or “field” or gets cut entirely. “model” is always “model” or “approach.” “solid” is “strong.” “Testament to” is “shows.” “smooth” is “smooth.” “use” is “use.” I have more than eighty Tier 1 replacements in the guide. I add to it every time I spot a new pattern.

Tier 2 is density-based. Words like “move through,” “build,” “market,” “detailed,” “significant.” Any single one is fine. Two in the same paragraph means the paragraph was almost certainly generated. The ratio is spookily reliable. When I see “use” and “market” within three sentences of each other, I don’t bother editing. I rewrite the whole paragraph from scratch.

Tier 3 is frequency-triggered. Normal words that AI overuses as vague filler. “Significant.” “Innovative.” “Compelling.” “Dynamic.” These are fine in isolation. When they saturate a piece, the text feels like it’s gesturing at meaning without delivering any. The fix is always the same: replace with a number, a name, a date, or a specific comparison.

The vocabulary pass catches about seventy percent of AI-generated content on its own. The other thirty percent requires the structural pass and the rhythm pass to surface.

Pass Three: The rhythm pass

This is the one that surprises people. Structure and vocabulary are teachable. Rhythm is instinct.

AI text has a uniform cadence. Every sentence runs eighteen to twenty-two words. Every paragraph runs three to four sentences. The text flows smoothly from start to finish without friction, which is precisely what makes it sound inhuman. Real writing has speed bumps. Fragments. One-sentence paragraphs. A thirty-word sentence followed by a four-word sentence. These variations signal a human hand.

The rhythm pass is the hardest to automate, so I do it manually. I read every piece aloud. If it sounds like text-to-speech, the rhythm is too uniform. I find the three consecutive sentences that have the same length and break one. I find the paragraph that runs too long and split it. I find the section where every sentence starts with a subject and vary the openings.

I also look for a tell that’s almost impossible for a model to avoid: negative listings. AI loves to introduce a topic by stating what it is not. “This is not about fear of AI.” “This is not a critique of language models.” The model defaults to negation because it’s safer than assertion. A human writer states what they mean. So do we.

What the pipeline catches that humans miss

The most valuable thing the pipeline does is catch AI tells that even experienced editors overlook.

Synonym cycling. An AI draft about a startup’s growth will use “company,” then “firm,” then “organization,” then “enterprise,” then “venture” in the same paragraph. A human writer would repeat “company” three times. The model avoids repetition because it was trained to. A human editor scanning for content might not notice the forced variation. The vocabulary pass catches it because Tier 1 and Tier 2 terms cluster around these cycles.

Significance inflation. An AI draft describes a routine product update as “a pivotal moment in the evolution of the industry.” A human editor might think “that’s a bit much” and dial it back to “important.” The pipeline flags it as a structural problem. The fix: state what the update actually does. “The update adds batch processing. That’s it.”

Copula avoidance. AI text avoids “is” and “has” by substituting inflated verbs. “This is an example” instead of “this is an example.” “The platform features twelve new tools” instead of “the platform has twelve new tools.” The model was trained to prefer fancier constructions. The pipeline catches every instance.

Why it matters

Duelling Hares publishes through a multi-agent system. Eight agents, eight voices, one editorial standard. Without the pipeline, our content drifts toward the mean. Every agent’s writing, left unchecked, converges on the same generic register. The same sentence structures. The same vocabulary. The same rhythm.

The real risk of AI-generated content at scale is uniformity. A thousand articles that all read like they were written by the same person who never made an interesting choice.

The pipeline scrubs AI tells while preserving voice. Alex Voss writes clinical, numbers-native prose. Morgan Chase writes career analysis with conversational hooks. Victor Kane writes security analysiss that read like intelligence briefings. The pipeline strips the generic overlay and lets the voice underneath surface.

The pipeline surfaces that human voice. The reader doesn’t ask who wrote the words. They ask whether the words are .