How to Make AI Writing Sound Human: The De-AI Playbook

How to make AI writing sound human starts with accepting one uncomfortable truth: the giveaways are not random accidents, they are the fingerprints of a model trained to write like the average of everyone at once. The em dash gets all the attention right now, and for good reason, but it is one tell in a stack of a dozen. This playbook works through all of them. We start with the dash and OpenAI's November 2025 fix, then move to the quieter patterns that survive after you delete every dash: the cliche verbs, the symmetrical lists, the 'it's not X, it's Y' cadence, the unnervingly even sentence rhythm. Then we get honest about why scrubbing those tells by hand does not scale across a real posting cadence, and what actually fixes the problem at the source instead of after the fact.

Why did the em dash become the number one AI giveaway?

Sometime in 2024 and 2025, readers collectively decided the em dash was the smoking gun. Publications noticed it. Rolling Stone ran a piece on the punctuation mark that suddenly signaled a machine wrote your text, and the sentiment spread fast enough that people started scrubbing dashes from their own genuine writing just to avoid the accusation (see the coverage at rollingstone.com). The instinct is right even if the target is slightly off. Something about long-form AI output feels processed, and the dash is the most visible edge of that feeling.

Over-frequency, not the dash itself: what the pattern actually is

The tell is not the em dash. The tell is how often the em dash shows up. A human writer might reach for one every few paragraphs, usually to set off a sharp aside or a sudden turn. A default model reaches for one in nearly every paragraph, sometimes twice a sentence, because it has learned that the dash is a low-risk connective that always fits. That over-frequency is the pattern your eye is actually catching. When six posts in a row from the same account all lean on the same rhythm of clause, dash, clause, the reader stops seeing punctuation and starts seeing a machine. So the fix is never as simple as banning one character. You are trying to break a statistical habit, not delete a symbol.

The myth to bust: humans use em dashes too

Plenty of careful human writers love the em dash. It is a legitimate, elegant mark, and treating its presence as proof of AI is a mistake that punishes good writing. The problem for social posts is practical, not moral. On Twitter and LinkedIn, most readers now carry an unconscious AI-detector, and that detector fires on the dash whether or not the accusation is fair. You are not writing for a grammar board. You are writing for a feed where perception moves faster than nuance. That is why removing the dash from public posts is worth doing even though the mark itself is innocent, and it is also why removing it alone is nowhere near enough.

How do you stop ChatGPT using em dashes (the exact fix)?

You stop ChatGPT using em dashes by putting a standing instruction in Custom Instructions or your system prompt, and by naming what should replace the dash. A bare 'do not use em dashes' fails in a predictable way: the model still writes in dash-shaped sentences, then swaps the character for a semicolon or a colon at the last moment. The rhythm survives, only the symbol changes, and the text still reads as machine-made. You have to tell it where the clause should break instead.

The Custom Instructions line to paste, and why you must name a replacement

Paste the block below into ChatGPT's Custom Instructions (or the equivalent system prompt in Claude or Gemini). The important part is the second half, which forces a specific substitution rather than leaving the model to improvise. Choosing the substitute matters: a period when the two ideas can stand as separate sentences (this is the humanizing default, because short sentences read as spoken), a comma when the aside is light and belongs in the same breath, and the word 'and' or 'but' when the two clauses genuinely depend on each other. Do not let it fall back on semicolons, which are their own AI tell in casual social copy.

OpenAI's November 2025 fix and how reliable it now is

In November 2025, OpenAI shipped a change so that ChatGPT would actually respect a custom instruction to avoid em dashes, something users had complained it ignored for months (the update was covered widely, including at tomsguide.com and on openai.com). Before the fix, you could tell the model ten times and still get a dash in the next reply. After it, the instruction largely holds inside a session. Two honest caveats remain. First, compliance is high but not perfect, so proofread anyway, especially in longer outputs where the model drifts. Second, the fix only addresses the one character you named. It does nothing about the dozen other tells, which is exactly the trap: people paste the anti-dash line, see it work, and assume their writing now reads as human. It does not.

Why the em dash is fatal in a 200-character tweet specifically

In a long article, one dash among four hundred words is invisible. In a tweet, one dash among thirty words is a third of your punctuation. The shorter the post, the higher the density of any single tell, so a construction that would slide past a reader in an essay screams in a one-liner. This is why AI tweets get flagged faster than AI blog posts, and why the de-AI job on Twitter is both harder and more urgent. If you want the deeper version of this problem, we cover it in why your AI tweets sound generic or robotic, and the LinkedIn equivalent in why every AI LinkedIn post sounds the same.

What are the dozen quieter tells that survive after the dash is gone?

Delete every em dash and the writing still smells like AI, because the dash was never the whole problem. It was the most visible member of a family. Below is the catalog: what each tell is, why the model produces it, the manual fix, and, most importantly, whether that fix actually persists across a run of posts or leaks back in the next time you draft. That last column is the one that decides whether hand-editing is a strategy or a treadmill.

The 'it's not X, it's Y' construction and how to kill it at the source

This one has become the em dash's successor as the most-mocked tell. You know the shape: a sentence that sets up a false contrast to sound profound, then resolves it with a neat reversal. It reads as insight but carries almost none, because the model reaches for the template regardless of whether a real contrast exists. Avoiding the 'it's not X, it's Y' pattern means killing it at the prompt level (add 'do not use antithesis or 'not this, but that' constructions' to your instructions) and then catching the survivors by hand, because it slips through more often than the dash. The plain fix is almost always to just say the second half. State Y directly and drop the manufactured tension.

Cliche verbs, hedge words, and phrases like 'in today's fast-paced world'

Default models gravitate to a small vocabulary of impressive-sounding verbs and framing phrases because those words are statistically safe. Delve, leverage, unlock, elevate, harness, navigate, foster, and the immortal opener 'in today's fast-paced world.' None of them are wrong, exactly. They are just what everyone's AI produces, which makes them a uniform. Alongside the verbs sit the hedges: 'it's important to note,' 'that said,' 'ultimately,' 'at the end of the day.' The fix is blunt substitution. Replace each cliche verb with the plain word you would use out loud, and delete the hedges entirely, because a confident post does not announce its own caveats.

Symmetrical lists, bold-term formatting, and predictable emoji patterns

Three tells cluster around structure. The first is the perfectly balanced list, almost always three items of similar length, because balanced triads read as coherent to the model's scoring. Real thinking is lumpier: sometimes you have two points, sometimes five, and they are rarely the same length. The second is the bold-term-then-colon format ('Consistency: post every day. Value: teach one thing.') that turns a post into a slide. The third is decorative emoji, one per line or a rocket at the top, placed to signal energy rather than meaning. On LinkedIn especially these three combine into the instantly recognizable AI post shape. Break the symmetry, write in sentences instead of labelled fragments, and cut emoji to at most one that you actually mean.

Uniform sentence rhythm: the tell nobody edits out

This is the deep one, and it is why symptom-scrubbing has a ceiling. Human writing has a heartbeat. Sentences run long, then snap short. A fragment. Then a rangy clause that wanders a little before landing. Default AI output averages all of that away into medium-length, evenly constructed sentences, one after another, and the evenness itself is the tell, even when every individual sentence is fine. Almost nobody edits this out, because it is invisible at the sentence level and only shows up in the aggregate. You have to deliberately introduce variance: drop a two-word sentence next to a twenty-word one, start something with 'And,' let one line run on. Rhythm is the hardest tell to fake by hand and the easiest to reproduce automatically if the model has actually learned how a specific person writes. Hold that thought.

Why does AI write like the average of everyone?

Every tell in that catalog traces back to one cause. A general-purpose model is trained to predict the most likely next token across an enormous corpus of everyone's writing, which means its safest, highest-probability output is the statistical middle of all of it. The middle is where the cliche verbs live, where the balanced triads live, where the even rhythm lives. The model is not malfunctioning when it produces generic text. It is doing exactly what it was built to do, which is regress toward the average. That is the structural reason the tells cluster rather than appearing at random, and it has a hard implication: the only durable escape from the average is a model conditioned on one specific writer's history instead of everyone's. That is the approach VoiceMoat's audenAI takes, and we come back to it below. For a fuller treatment of the mechanism, see why AI drafts sound the same.

No persistent voice: the structural reason surface fixes leak

A default chat model has no persistent sense of you. Each new conversation starts from the average again. You can spend twenty minutes in one session teaching it your rhythm, and the next morning it has forgotten, because your voice was never stored, only borrowed for that thread. This is why every surface fix leaks. Your Custom Instructions hold the anti-dash rule, but they cannot hold the hundred subtler decisions that make your writing yours: which words you never use, how long your sentences run, where you land a joke. So you re-teach, or you re-edit, every time. The fix that persists has to live in the model's picture of you, not in a prompt you keep re-pasting.

What is the manual de-AI checklist (and where does it stop scaling)?

Here is the honest checklist that works, followed by the honest admission of where it breaks. If you are de-AI-ing a single post by hand, this sequence gets you most of the way to human.

Contractions, sentence-length variety, and casual transitions

The quickest humanizing moves are mechanical. Use contractions everywhere ('you're' not 'you are'), because formal expansion reads as machine. Force sentence-length variety on purpose, as the flow above says. And swap stiff connectives ('furthermore,' 'moreover,' 'additionally') for the casual ones people actually use to change direction ('but,' 'so,' 'anyway,' 'here's the thing'). None of these are clever. They are just the difference between text that was assembled and text that was spoken, and they move the needle more than any single fix except rhythm.

Adding personal anecdotes and specific, checkable detail

The strongest de-AI move is one the model cannot do for you: put in something only you know. A number from your own week, the name of the client who said the thing, the exact moment the idea landed. Specific, checkable detail is the one signal a general model cannot fake, because it does not have access to your life. Generic posts stay generic because they float at the level of abstraction the average lives at. One concrete detail drops the whole post out of the average and into your actual experience, and that is what readers read as human, more than any punctuation choice.

Tone and persona prompting

You can push the draft closer to human before you even start editing by prompting for a persona: 'Write like a blunt operator who hates buzzwords and uses short sentences.' This helps, and it is worth doing. But persona prompting borrows a costume, not your face. It moves the output to a different average (the average of blunt operators) rather than to you specifically. It is a real improvement and a real ceiling at the same time. If you want to go deeper on prompting the model toward your own patterns, we walk through it in how to train AI to write in your voice and the tweet-specific version in how to make ChatGPT write tweets in your voice.

Humanizer tools as a post-processing scrub: what they fix and miss

Humanizer tools like Quillbot, Walter Writes, and SupWriter take finished AI text and rewrite it to read as more human and to slip past AI detectors. They are genuinely useful as a last-mile scrub: they smooth obvious tells, vary some phrasing, and lower detector scores. What they miss is the same thing every post-processing step misses. They operate on text that already arrived as average-AI output, so they are moving it from one average to a slightly different average, not to you. They also cannot add the specific detail only you know, and they can flatten meaning while chasing a detector score. Treat them as spell-check for AI tells, not as a voice. If you want to see what a scrub actually changes, our free humanize AI text tool and AI detector let you test a draft before and after without paying for anything.

The 60-70% AI, 30-40% manual-edit rule and its ceiling

The workflow a lot of good writers land on is roughly 60 to 70 percent AI draft and 30 to 40 percent manual human edit, and it produces genuinely good posts (the hybrid split is discussed well at every.to). The catch is the word 'manual.' That 30 to 40 percent is real work, per post, every post. It scales fine when you publish twice a week. It falls apart when you publish twenty times a week, because you cannot hand-edit rhythm, cliche verbs, and symmetrical lists out of twenty posts and stay consistent, and the moment you get busy the un-edited draft ships and the tells are right there in public. Manual de-AI is not a one-time fix. It is a maintenance cost that grows linearly with your posting volume, and that is the ceiling the whole approach hits.

What is the durable fix? Write in your own voice from the start

Everything above is symptom treatment. You take average-AI text and you edit it toward human, one tell at a time, forever. The durable fix flips the order: generate in your own voice from the first draft, so the tells never get baked in and there is nothing to scrub out. This is where VoiceMoat comes in, and here is the disclosure up front. Full disclosure: VoiceMoat is our own product, so treat this as us showing our work, not a neutral referee. (And to clear up a common mix-up: VoiceMoat is a writing tool for Twitter and LinkedIn, not the VoiceMod voice-changer. Different product, different problem.)

VoiceMoat is an AI social media tool for Twitter and LinkedIn (create, schedule, engage, analyze, grow) whose drafting learns your own posting history first, so tweets and posts start in your voice instead of arriving as average-AI text you then have to de-AI by hand. The writing partner inside it is audenAI. Rather than pull from the average of everyone, audenAI models your own writing patterns from your past posts: your punctuation, your sentence rhythm, your hooks, the words you use and the ones you never touch.

How audenAI learns your punctuation and cadence up front

The mechanism is deliberately boring, which is the point. audenAI reads your existing Twitter and LinkedIn history once and builds a picture of how you actually write, then conditions drafts on that picture instead of on the statistical middle. Because the model already carries your cadence and your punctuation habits, the tells from the catalog above do not appear and then get removed. They never appear, because the thing generating the draft is not reaching for the average dash frequency or the balanced triad in the first place. You draft in Studio, you handle replies in Engage, you schedule, and you check what worked in Analytics, all on the same picture of your voice. See how it drafts in your own voice from the first line.

Starting human vs scrubbing human: why the tells never appear

The difference between the two approaches is not quality on any single post. A patient editor can hand-tune one AI draft to read beautifully. The difference is what happens across a cadence. Scrubbing is per-post effort that stays inconsistent and lets tells leak back the moment you are busy. Voice-first drafting captures your patterns once and applies them to every draft, so consistency is the default rather than something you fight for. To be plain about the limits: audenAI suggests, and you still decide. It gets you a draft that already sounds like you, and you edit for judgment and truth, not for punctuation and rhythm. The maintenance cost that grew with your volume in the manual approach does not grow here, because the work happened once, up front, instead of twenty times a week.

The honest bottom line: removing em dashes is worth doing, and the manual checklist genuinely works on any single post. But scrubbing tells out is treatment, and treatment does not scale. The durable answer to how to make AI writing sound human is to make it write in your voice from the start, so there is nothing average to correct. See how VoiceMoat drafts in your voice in Studio, so there is nothing to de-AI later. Free to start, paid plans from $35/mo. audenAI suggests, you decide.