Open ChatGPT's settings, go to Personalization, and put a substitution rule into Custom Instructions: replace em dashes with a period to start a new sentence, or a comma to continue one. That fix persists across every new chat instead of being re-asked each time. AI Central's guide is careful to say it removes most em dashes, not all, and reports the rule holds better in Thinking mode. Then test it by asking for a story.
Reviewed August 7, 2026, against AI Central's guide of the same name, which sits in the library alongside this article.
Why anyone bothers
The em dash has become a tell. It is the punctuation mark readers now associate with machine-written copy, and once someone notices it in a draft they stop reading the argument and start reading the pattern.
Fighting it inside the chat window does not scale. The instruction dies with the conversation, so you retype it at the top of every session and forget it exactly when the writing matters.
AI Central's guide takes a different route. It moves the correction into account settings, so the rule rides along with every new chat instead of being asked for one at a time.
The setting, in two clicks
The first two steps of the guide are pure navigation, and they are the part most people skip. Open your ChatGPT settings and choose Personalization. Then turn customization on, which is what exposes the Custom Instructions field the rest of the technique depends on.
Nothing here is a hidden feature or a workaround. It is a standing preference, entered once.
The rule that does the work
The instruction itself is a single sentence, and the guide prints it in full inside a screenshot of the Custom Instructions box.
Systematically replace em-dashes (“—”) with a dot (“.”) to start a new sentence, or a comma (“,”) to continue the sentence.
That line comes from AI Central's guide. It is worth reading twice, because the wording is doing more than it looks like.
Most people write the negative version. They tell the model not to use em dashes and leave it there. A prohibition gives the model something to avoid and nowhere to go, so the punctuation reappears the moment a sentence structure calls for it.
The guide's version is a substitution with two branches. If the clause after the dash can stand on its own, it becomes a period and a new sentence. If it cannot, it becomes a comma and the sentence keeps running. That is a rewrite instruction rather than a ban, and rewrite instructions are easier for a model to apply consistently because they say what to produce, not only what to withhold.
The lines around it in the same field are not about punctuation at all. They ask for clear, straightforward English and for blunt, direct feedback. Both push in the same direction, toward plainer sentences that need fewer dashes in the first place.
The guide hedges, and that is the honest part
One word in the instruction step separates this guide from most punctuation advice, and it sits inside a parenthesis.
In “Custom Instructions”, add this simple prompt to remove (most) em dashes. Copy & paste it from my caption. Hit save
That is AI Central's own framing, parentheses included. Most, not all.
Take the hedge seriously. Custom instructions are a standing preference the model weighs while writing, not a filter applied to the finished text. Long generations, heavy formatting and unusual output shapes are where a stray dash is most likely to survive. Plan to skim your longest outputs rather than assuming the setting has sealed the leak.
The Thinking mode note
Step four of the guide is an observation rather than an instruction, and it arrives with an unusual admission.
It works better with the “Thinking” mode of ChatGPT. And I don’t know why
AI Central reports the effect and declines to explain it. That is the right way to publish something you have noticed but not tested, and it deserves to be passed on in the same shape. Try it, watch it, do not treat it as a mechanism.
The guide shows the ChatGPT model picker to make the choice concrete. In the picker's own labels, Auto decides how long to think, Instant answers right away, Thinking mini thinks quickly, Thinking thinks longer for better answers, and Pro is described as research-grade intelligence.
The one the guide points at is Thinking, the option described as thinking longer for better answers.
How to test it, and why the test is a story
Step five is the check, and the guide keeps it light.
I love asking for a story. Extra points for a fairytale about yourself.
That is AI Central's test, and it is a sharper one than it first appears.
Question and answer prompts produce short, clean sentences that rarely reach for an em dash at all, so they cannot tell you whether a rule is holding. Narrative prose is the opposite. It runs on interruption, aside and reversal, which is precisely where the dash lives. Asking for a fairytale stress-tests the setting instead of flattering it.
The output printed in the guide shows the model thinking for ten seconds, then producing a short fable about a young scribe named Ruben. It is written in clipped, standalone sentences. Where a dash would normally join two thoughts, the text simply breaks into a new line. There is not a single em dash anywhere in the excerpt.
Then the guide does something most tips never bother with. It runs the same request again with the instruction removed and describes that version as full of em dashes. One prompt, two settings, a visible difference. That is a control, and it is what makes the claim checkable rather than merely asserted.
What to do with this
The whole procedure is five moves, and only the first three are setup.
- Open ChatGPT settings and select Personalization.
- Turn customization on so the Custom Instructions field becomes editable.
- Paste the substitution rule, replacing em dashes with a period to start a new sentence or a comma to continue it, then save.
- Switch to Thinking mode when the output matters, which is where the guide reports the better result.
- Ask for a fairytale as your check, because narrative writing is where the dash would otherwise surface.
- Skim long outputs anyway, because the guide promises most em dashes, not every one.
If you write for a living, the durable lesson is not about punctuation. It is that a standing instruction phrased as a substitution beats one phrased as a prohibition. The same shape works on any habit you want gone from a model's output. Name the thing, then name what should take its place.
Does this remove every em dash?
No, and the guide does not claim it does. It says the instruction removes most em dashes. Treat it as a strong default that shrinks the problem to something you can catch in a quick read, not as a guarantee.
Do I have to paste this into every new chat?
No, and that is the point of putting it in Custom Instructions under Personalization rather than in the chat window. You enter it once and it applies to conversations from then on.
Which ChatGPT mode should I use with it?
AI Central's guide reports it works better in Thinking mode, the option the model picker describes as thinking longer for better answers. The guide is open about not knowing why, so take it as a tip to try rather than a rule to trust blindly.
How do I know it is working?
Ask for a story, ideally a fairytale, then read the result for dashes. Story writing is where a model reaches for an em dash most, so a clean fable is real evidence. Running the same prompt with the instruction removed gives you the comparison the guide itself used.
Where does this come from?
The source is AI Central's own guide, How To Remove EM Dashes in ChatGPT, filed in the library alongside this article. Every step, quote and claim reported here is drawn from that document.