How to Stop ChatGPT From Writing “Not Just X, but Y”
Tired of “not only X, but Y” in AI copy? Use these prompt rules and revision steps to cut canned phrasing and improve readability fast.
If you have spent enough time around generative text tools, you start to hear the machinery. Certain phrases arrive with such regularity that they stop sounding like writing and start sounding like a preset. One of the loudest giveaways is the balance-by-template sentence: “not just X, but Y,” “not only X, but also Y,” “it isn’t X, it’s Y,” or “the issue isn’t X, it’s Y.”
These lines are meant to sound sharp and reflective. In weak AI copy, they usually do something far cheaper. They inflate a basic point and present it like a revelation. The deeper problem, as Does ChatGPT Argue Like Students? argues, is that model-generated prose often leans on formulaic bundles and reusable sentence frames. Once you have edited enough machine-drafted text, that rhythm becomes hard to miss.
Why this pattern keeps showing up
This family of sentences sits close to antithesis, the rhetorical move that places opposed ideas in parallel form. Used carefully, antithesis can be powerful. It sharpens a real contrast and gives a sentence real bite. In Using Pre-Trained Language Models in an End-to-End Pipeline for Antithesis Detection, the authors describe antithesis as a figure built from parallel structure plus opposed words or ideas, and they note its persuasive force.
That matters because LLMs are not pulling this trick out of thin air. They are reaching for a familiar rhetorical shortcut that already exists in the language. Used occasionally, the device can work well. The trouble starts when it shows up every few paragraphs.
That overuse makes sense. A balance-by-template sentence is cheap scaffolding. It sounds organized. It sounds emphatic. It sounds like a turn in the argument even when no real turn has happened. The article points to a 2024 EMNLP paper that found model-generated text relies on syntactic templates more heavily than human reference text, with many of those templates traceable to patterns already present in pretraining data. In plain English, the model keeps reaching for sentence skeletons it has seen before because those skeletons are easy to assemble.
More on AI writing
Why experienced AI users recognize it so quickly
Once you notice this habit, you start seeing it everywhere because it belongs to a broader cluster of AI writing tells. Jiang and Hyland found that ChatGPT essays often leaned on rigid lexical bundles such as “this essay will,” “the potential for,” “the ability to,” “the need for,” and “the role of.” Those phrases are not identical to “not just X, but Y,” but they come from the same factory. They are sentence frames that organize the paragraph before the writer has supplied much real substance.
That is also why heavy AI users often become sharp detectors. The article cites a 2025 ACL paper showing that people who frequently use LLMs for writing tasks were highly accurate at identifying AI-generated articles. The authors say those expert readers picked up on vocabulary, sentence structure, originality, and tone. That fits the lived experience of anyone who has spent enough time editing model output. After a while, you stop reacting to single words and start recognizing cadence.
Why this habit weakens otherwise decent writing
The first problem is wasted emphasis. A sentence like “this policy is not just inefficient, it is dangerous” sounds like it is making an important turn. But unless the next sentence proves actual danger, the line has only upgraded the drama. It has not upgraded the thought.
The second problem is lost specificity. When a model reaches for prefab contrast, it often skips the detail that would make the paragraph worth reading. In the npj Digital Medicine study on ChatGPT-generated abstracts, human reviewers described many suspected AI-written abstracts as “superficial and vague.” That gets to the heart of the problem. Formulaic phrasing does not only sound machine-made. It also tends to do less work per sentence.
The third problem is voice. When a draft keeps falling back on “not just X, but Y,” the model starts becoming the loudest personality in the room. Readers hear the template before they hear the writer. For students, journalists, founders, marketers, and anyone else whose credibility rests on sounding like a real person with a real point of view, that is a serious cost.
Set the rule before the model writes a word
The cleanest fix starts before the first draft. Do not wait until the model has already littered the page with canned contrasts. Put the rule in the prompt at the start: avoid balance-by-template constructions such as “not just X, but Y,” “not only X, but also Y,” and similar contrastive reframes unless the contrast is essential. Ask for direct claims, concrete examples, and specific evidence instead.
This works because current guidance keeps making the same point. The OpenAI prompt engineering guide recommends giving the model clear instructions and examples of the output you want. The Claude prompting best practices make a similar case for being direct about role, tone, and structure. Models do better when you tell them what strong writing should look like before they start generating filler.
Show the model what restraint looks like
Abstract style notes help, but examples usually help more. Anthropic’s Increase output consistency guidance recommends constraining output with examples, and Anthropic’s broader prompting docs say examples are one of the most reliable ways to steer tone, structure, and format. OpenAI says much the same in its Best practices for prompt engineering with the OpenAI API.
So give the model a short sample of prose that sounds like you or like the publication you are writing for. Tell it to match the directness, sentence variety, and level of detail without copying the wording. That does far more than asking it to “sound human,” which is vague enough to invite more performance instead of better prose.
Separate drafting from cleanup
One-shot prompting often produces polished sludge. The draft looks smooth on first read, then collapses when you inspect the sentences one by one.
A better workflow is to split the job into stages. Ask for an outline first. Ask for the full draft next. Then ask for a style pass that checks each paragraph for canned phrasing, weak transitions, and unnecessary contrastive reframes. Anthropic explicitly recommends chaining prompts for complex tasks, and OpenAI’s documentation also points writers toward iterative refinement rather than hoping a single prompt gets everything right.
This matters because balance-by-template language often sneaks in during the “make it sound polished” phase. When drafting and editing happen in separate passes, it becomes much easier to catch those reflex sentences before they harden into final copy.
Run a targeted pattern audit
After the draft exists, run a second pass with a narrow brief. Ask the model to find every sentence that uses antithesis, contrastive reframing, or balance-by-template wording and rewrite each one as a direct statement that adds information.
That instruction works better than broad requests to make the prose feel more natural. OpenAI’s help guidance explicitly advises writers to say what to do, not only what to avoid. That is useful here because “remove AI tone” is fuzzy, while “rewrite every contrastive reframe as a specific claim” is concrete enough to act on.
A good audit should move sentence by sentence. It should ask what claim the sentence is making, what comparison it depends on, and what evidence makes that comparison worth keeping. When the contrast adds nothing, cut it.
Make every contrast earn its place
The easiest editing test is a ruthless one. When you see a sentence built around contrast, stop and ask four questions. What is the actual claim. Compared with what. What evidence supports the contrast. Would the paragraph lose anything if the line were rewritten as a plain sentence.
That last question matters most. If the answer is no, then the sentence is probably decorative rather than useful.
Take a line like this: “This rule is not just burdensome, but exclusionary.” It sounds forceful. It also tells the reader almost nothing. A stronger version would explain the mechanism: “This rule adds legal and reporting costs that large firms can absorb and small firms cannot.” The second version gives the reader something to hold on to. The first one only performs seriousness.
That is the real editing goal. Replace the performance of insight with the substance insight is supposed to deliver. Give the reader a fact, an example, a mechanism, a number, or a quote.
Save the preference as a standing rule
If you use ChatGPT regularly, there is no reason to repeat this preference from scratch every time. OpenAI’s Custom Instructions let you set standing preferences that apply across future chats. That makes them a useful place to store house rules against canned contrasts, canned conclusions, and other recurring AI crutches.
This does not solve the problem on its own, because any long draft still needs editing. But it lowers the baseline amount of cleanup. And that matters if you write with these tools every day.
The bigger problem is standardization
There is a larger issue hiding under this annoyance. When millions of people draft with a small number of frontier models, those models start pushing the same sentence skeletons into everything from student essays to marketing copy to political messaging. The result is a quiet form of standardization. It creates a quieter kind of centralization, one that shapes cadence more than truth.
That should worry anyone who cares about voice. Writing is supposed to reveal judgment, temperament, and differences in emphasis. Balance-by-template language does the opposite. It makes different writers sound as though they all hired the same nervous speechwriter.
Research on syntactic templates points to the mechanism you would expect. Models repeatedly reuse favored patterns absorbed during training because those patterns are available, efficient, and statistically easy to continue. The result is prose that sounds orderly while flattening away the individual pressure of thought.
Further reading
To tighten this workflow even further, the OpenAI prompt engineering guide, Anthropic’s page on Increase output consistency, Anthropic’s broader Claude prompting best practices, and OpenAI’s Best practices for prompt engineering with the OpenAI API all point in the same direction. Be explicit, use examples, break complex work into stages, and give the model narrow editing tasks instead of vague aesthetic commands.
The research side is worth your time too. Does ChatGPT Argue Like Students?, Using Pre-Trained Language Models in an End-to-End Pipeline for Antithesis Detection, and the npj Digital Medicine study on ChatGPT-generated abstracts help explain why these patterns are persuasive, repetitive, and so easy to spot once you know what you are looking for.
The practical takeaway
The cure is to keep using AI without outsourcing your sense of emphasis. Balance-by-template sentences are easy for LLMs to produce, easy for readers to notice, and easy for you to cut once you learn the pattern.
Set the rule early. Show the model what restraint looks like. Separate drafting from editing. Run a hard second pass for canned contrasts. Do that consistently and the prose starts sounding less like a generic demo and more like a person who actually means what he wrote.
Explore more from Popular AI:
Start here | Local AI | Fixes & guides | Builds & gear | AI briefing