How to Stop ChatGPT's Writing From Sounding Like ChatGPT
"Make it sound more human" is one of the least useful instructions you can give ChatGPT. It's not that the model refuses, it's that "human" isn't a style, it's an absence of a very specific set of tells, and the model has no way to know which ones you're reacting to unless you name them. People can spot AI writing almost instantly now, not because it's badly written but because certain habits repeat across nearly every model's default output. Once you know what those habits actually are, you can tell ChatGPT to remove the specific thing, and it will.
If you haven't used ChatGPT much beyond basic prompts, the Complete Beginner's Guide to ChatGPT is worth reading first. This article assumes you already get usable drafts back and just want them to stop reading like drafts.
The structural tells, which matter more than the word choice
Word choice is the easiest thing to catch and the least important thing to fix. The structural habits are what actually give AI writing away at a glance, before anyone's read a sentence closely.
The rule-of-three list. Almost everything comes in threes: three benefits, three reasons, three tips. Real writing is uneven. Someone explaining something they actually care about lists two points because that's all that mattered, or five because the third one led to two more. A paragraph that resolves into exactly three parallel items, every time, is a tell.
The symmetrical paragraph. AI-generated paragraphs tend to be near-identical in length, each one setting up a point, developing it for two sentences, and landing a soft conclusion. Human writing is lumpier. A quick aside gets one sentence. Something the writer actually has an opinion about gets four.
The "not just X, but Y" construction. "This isn't just a marketing tactic, it's a way of thinking about your customer." It shows up so often because it's a genuinely good rhetorical move, once. Used in nearly every paragraph, it becomes a fingerprint.
The summary that restates the intro. A closing paragraph that says, in slightly different words, everything the piece already said. Real writing trusts the reader to have read the middle; it ends on a new thought, a next step, or just stops.
The hedge-then-assert pattern. "While there are many factors to consider, ultimately, the most important thing is..." This construction shows up so often that its presence alone is now a signal, independent of what follows it.
The characteristic phrases
These aren't inherently bad words. They're overused specifically by AI writing, at a rate real writers don't match, which is what makes them recognizable rather than the words themselves being wrong.
"Moreover," "furthermore," "additionally" doing the work a comma or a new sentence would do more naturally
"It's important to note that" or "it's worth noting" before a point that could just be stated
"In today's [fast-paced/digital/competitive] world" as scene-setting
"In conclusion" or "in summary" opening the last paragraph
"Whether you are a beginner or an expert" as a false-inclusive opener
Hype verbs that promise more change than the sentence actually delivers, the kind of marketing-brochure language that claims something reinvents or permanently upgrades a category, when what it actually does is one useful but modest thing
The em dash, used constantly as a substitute for a comma, colon, or period
Note
None of these are wrong in isolation. A single "moreover" in a long piece is invisible. Six of them, plus a rule-of-three list, plus an em dash every other sentence, is what actually reads as generated.
A real before and after
Here's an unedited ChatGPT paragraph, the kind you'd get from a plain "write about why small businesses should have a content calendar" prompt, followed by a rewrite using the instructions further down.
Before:
In today's fast-paced digital landscape, having a content calendar isn't just a nice-to-have, it's essential for success. Moreover, a well-structured content calendar helps ensure consistency, saves time, and reduces stress. Additionally, it allows businesses to plan ahead, align content with key dates, and maintain a cohesive brand voice across platforms. Ultimately, whether you are a solo entrepreneur or part of a larger marketing team, a content calendar can transform your content strategy.
After:
Most small businesses don't fail at content because they run out of ideas. They fail because Tuesday shows up and nobody planned anything, so whatever gets posted is whatever's fastest to write, not what actually needed saying. A content calendar fixes that by moving the "what do we post" decision to a slow afternoon instead of a rushed one. It also catches the obvious stuff early, a product launch and a competitor's sale landing on the same week, before it's a scramble instead of a plan.
The rewrite is shorter, uses no em dashes, has no rule-of-three list, and every sentence is doing a different kind of work: one names the real problem, one states the mechanism, one gives a concrete example. Nothing in it announces "in conclusion." It just stops.
It's worth being specific about which edit fixed which tell, since "make it sound less like AI" is exactly the vague instruction this article is arguing against.
| What changed | Why it works |
|---|---|
| Cut the scene-setting opener entirely | The original spent a full sentence establishing "today's landscape" before saying anything. The rewrite's first sentence is already the claim. |
| Replaced two rule-of-three lists ("consistency, saves time, and reduces stress" / "plan ahead, align content with key dates, and maintain a cohesive brand voice") with one concrete scenario | A list of three parallel benefits reads as filled-in template slots. A specific Tuesday, with a specific consequence, reads as something someone actually noticed. |
| Removed the transition scaffolding ("Moreover," "Additionally," "Ultimately") | Real writers connect ideas with "so," "it also," or just a new sentence. Transition words doing that job are a tell precisely because they're doing a job punctuation could do. |
| Cut the false-inclusive hedge about who this applies to | The point is true regardless of company size. Saying so out loud added a sentence without adding a fact. |
| Varied sentence length on purpose | Short claim, then one longer sentence with a clause in it, then a medium one. Four same-length sentences in a row is its own tell, independent of word choice. |
| Replaced a vague promise of transformation with a concrete pair of events | "A product launch and a competitor's sale landing on the same week" is checkable and specific. A claim that something "transforms your strategy" is neither. |
A second, shorter example: an About bio
The content-calendar paragraph is a good demonstration because it's long enough to show every tell at once. Most real editing happens on shorter text, so here's the same treatment applied to something a freelancer would actually paste into ChatGPT: a one-paragraph website bio.
Before:
Jordan Reyes is a dedicated marketing professional with a genuine passion for helping brands grow. With years of experience across strategy, content, and analytics, Jordan brings a thoughtful, results-driven approach to every project. Whether working with early-stage startups or established companies, Jordan is committed to delivering high-quality work that makes a real difference. Ultimately, Jordan's goal is simple: to help every client succeed.
After:
Jordan Reyes spent six years running paid social for two YC-backed startups before going independent in 2019. The work now splits about evenly between brand strategy for early-stage founders and hands-on campaign builds for teams that already know what they want and need someone to execute it well. Recent work includes a rebrand for a Denver furniture retailer and a six-month paid acquisition ramp for a fintech app that grew monthly ad spend from $40,000 to $310,000 without cost-per-lead climbing.
Same tells, smaller scale: the "strategy, content, and analytics" list is a rule of three doing no real work, "whether working with early-stage startups or established companies" hedges toward including everyone instead of saying anything, and "committed to delivering high-quality work that makes a real difference" is a sentence that would be true of almost any freelancer alive. The rewrite replaces all of it with two real numbers and two real clients, which is the same fix as the longer example: specificity standing in for a claim of quality, instead of a claim of quality standing in for specificity.
The prompt instructions that actually produce this
Telling ChatGPT "sound more human" gets you a vague pass at removing the most obvious tics. Naming the specific habits gets you closer in one attempt.
Rewrite this without: any rule-of-three lists, the phrase "not just X, but Y," transition words like "moreover," "furthermore," or "additionally," a summary paragraph that restates the intro, or any em dashes. Vary sentence length noticeably, some short, one or two longer. Cut the opening scene-setting sentence entirely and start with the actual point.
”For a full piece rather than a paragraph, it helps to give ChatGPT a real example of your own writing to match against, since "avoid these tics" describes an absence, not a voice. A positive model does more work than a list of negatives.
Here's a paragraph I wrote myself: [paste a real paragraph in your voice]. Match this rhythm and sentence variety when you rewrite the draft below. Don't imitate the topic or the specific words, imitate the pacing: short sentences next to longer ones, no rule-of-three lists, no transition-word scaffolding.
”Tip
If you write regularly, keep two or three paragraphs you're genuinely proud of saved somewhere. Pasting one into a rewrite prompt is a faster fix than describing your voice in the abstract, and it works from day one.
Where this still falls short
Even with specific instructions, a single rewrite pass rarely fully removes every tell, especially in longer pieces. The habits creep back in by paragraph four even when the first paragraph was clean. Two things help: rewriting in shorter chunks rather than asking for a full 1,500-word piece to be de-AI'd in one pass, and reading the result out loud. Rhythm problems that are invisible on the page are obvious the moment you try to say the sentence.
It's also worth being honest about what this doesn't fix. If the underlying idea is generic, no amount of sentence-level editing makes the piece specific. The tells covered here are about how AI writing sounds, not about whether it has anything to say. A structurally clean paragraph built on a vague, could-apply-to-anyone point is still a vague point, just a better-disguised one. The fix for that isn't a rewrite instruction, it's giving ChatGPT a real detail, a real number, a real name, to write about in the first place.