How to Stop ChatGPT From Using Em Dashes

Quick answer: there is no toggle. You do it in Custom Instructions, and even then one still slips through now and then.

I write every day and I have a hard rule against em dashes, so I have been fighting this one for months. Most of the advice you will find online is out of date, because it was written back when ChatGPT ignored the instruction no matter how you phrased it.

Here is what actually works now.

Where the setting lives

Open Settings, then Personalization, then Custom Instructions. That is the box where you tell ChatGPT how you want it to write, and it carries across your chats instead of resetting every time you start a new one.

Put your rule about em dashes in there. Not in the chat window. A rule you type at the top of a conversation only lasts for that conversation, which is why so many people feel like they are repeating themselves forever.

Your change takes effect right away. OpenAI’s own documentation says updates to custom instructions are applied immediately across all chats, including conversations you already have open, so you do not need to start a fresh one to test it.

There is a reason this matters more than it used to. On November 14, 2025, Sam Altman said OpenAI had finally made the model respect the instruction when you put it in custom instructions. Before that it would agree with you politely and then keep right on doing it. So if you tried this a year ago and gave up in disgust, it is worth another go.

Do not just tell it what to stop doing

"Do not use em dashes" is a weak instruction. You are telling the model to stop reaching for something without telling it what to reach for instead, so it has to guess, and it guesses badly.

Give it the replacement. This is the version I use:

Never use em dashes. Replace them with a period to start a new sentence, or a comma to continue the sentence.

That one change did more for me than any prompt trick I tried. The model is not working out what you meant, it just swaps in the punctuation you named.

It still leaks

Custom instructions are not a filter. They are a preference the model tries to honour, which means some get through anyway. In my experience it happens most in long outputs, inside bullet lists, and deeper into a long conversation where that original instruction has a lot of text competing with it.

So check the output before it goes anywhere. The fastest way is to search the text for the character itself. Three of them look similar and only the long one is the em dash:

hyphen -

en dash –

em dash —

Copy the long one and use it as your search term. Ctrl+F to find them, or Ctrl+H if you want to replace the lot in one pass.

Two things that quietly undo your setting

If you use Projects, check that project’s own instructions. OpenAI’s documentation says project instructions override your global custom instructions inside that project, so a project with its own set will ignore your global rule unless you repeat it there.

And if you reach ChatGPT through some other tool or through the API, your custom instructions do not travel with you. That box lives on your account, it is not a property of the model.

Why anyone bothers with this

Somewhere along the way the em dash turned into the tell. Readers spot a few of them in a paragraph and decide a machine wrote it, whether that is fair or not. (It is not especially fair. It is a perfectly good piece of punctuation that was doing nothing wrong.)

But that is where we are, and if you are publishing anything with your name on it, you are better off writing around it than arguing about it.

None of this makes the em dash impossible. It makes it rare, which is the realistic goal.

Spend the two minutes on the custom instruction, then get in the habit of searching the output before you publish. The instruction does most of the work now. The search catches the rest.

By the way I have been quiet around here for a while and I am getting back into it. If you have a ChatGPT (or any other AI tool) habit that drives you up the wall, tell me in the comments and I will look into it.

Edit, September 9, 2026: JD pointed out that two things I wrote here were out of date, and he was right. OpenAI’s documentation says custom instruction changes now apply immediately across all chats including existing conversations, and that project instructions override your global ones inside a project. I have corrected both. Thanks JD.