Em-dashes are everywhere, and they ruin good writing.
If you’re using ChatGPT, Claude, Gemini, or anything similar, it’s one of the first signs that you’re not reading something written by a real person. The frustrating part is that there is every chance the content and ideas are yours, but you need AI to help format the content correctly so you can actually get through all of your tasks on any given day.
Your individual content is now being bundled in the same group as cheap and nasty AI content, and it is not your fault. But I think I can help!
This post explains what em-dashes are, why AI loves them so much, and what you can actually do to stop them from showing up in your content.
What Are Em-Dashes?
An em-dash is a long horizontal line that looks like this: —
It’s wider than a hyphen, longer than a regular dash, and usually used in place of brackets, commas, or colons.
In theory, it adds emphasis or a break in rhythm. In practice, it turns everything into a dramatic pause. Most people don’t even know how to type one. That’s because it’s not on your keyboard.
To get it, you either need to copy and paste from somewhere else or memorise a keyboard shortcut like Alt+0151 or Shift+Option+Hyphen, which is too much hassle to begin with.
That’s the point. Real humans rarely write or type using em-dashes, but for some reason (which we will explain in a bit), AI models love them.
What’s Wrong with Em-Dashes?
The problem isn’t the dash itself, but what it does to your writing.
Other than being a massive AI giveaway in 2025, this format breaks natural flow and interrupts thought. And it creates a rhythm that feels more like a voiceover script.
The key reason AI uses them is that it doesn’t know how to choose. Instead of deciding whether to use a comma, a full stop, or a rewrite, it throws in an em-dash and keeps moving. My preference (and why I am always refining my prompts) is to go for the rewrite. I want to avoid writing every proposal, long email or progress report myself, so I naturally turn to AI. However, I don’t want the recipient to think that I have just spend 30 seconds to throw something together. This is why I now avoid em-dashes like the plague.
You’ll see em-dashes are most commonly used in sentence structures like these:
- It sounds good — until you read it again.
- That used to work — but things have changed.
- You’re probably thinking — isn’t this a bit dramatic?
For better or worse, none of these sound like me, and every one could be cleaner without the bloody dash.
Why AI Keeps Using Them
The big models (ChatGPT, Claude, Gemini, and the rest) are trained on billions of lines of internet writing. That includes books, articles, blog posts and newsletters.
The problem is, most of that writing was edited by people trying to sound clever. The em-dash is seen as a shortcut to drama, rhythm, and control. It’s treated as a mark of polished writing, when it actually just papers over weak structure.
AI doesn’t know that. It just copies what it’s seen most often. That’s why even a simple sentence gets the long dash treatment.
Why Prompts Alone Don’t Fix It
You’ve probably tried this already:
“Avoid using em-dashes in your response.”
“Do not use — at all.”
“Use natural punctuation only.”
It might work for a paragraph or two. But then they come back, especially if the model starts rewriting or expanding a section on its own.
That’s because most prompt instructions are soft preferences. Unless you hard-code the behaviour with examples or constraints, the model will default to what it thinks is good writing. And em-dashes are baked into that idea.
What Actually Works (Most of the Time)
Here’s what we’ve tested across thousands of runs in GPT-4, Claude, and Claude 3 Opus with consistent results.
1. Train It With Real Examples
The best way to break the habit is to show the model what good writing looks like without em-dashes. Give it a few samples that reflect your style, then ask it to match the tone and structure.
Don’t just say “avoid em-dashes.” Show it that your writing never needed them in the first place.
2. Frame Your Instructions Inside the Prompt
Put your tone rules in the same place you define the output. For example:
“You are writing as Elwyn, a web design coach who uses clear, natural phrasing. You never use em-dashes or semicolons, and you structure your thoughts like someone speaking clearly in real time.”
This works far better than leaving your rules as a throwaway line at the end of the prompt.
3. Rewrite With Constraints
If the dash still slips through, don’t just delete it. Ask the model to rewrite the whole sentence:
“Rewrite this paragraph using only standard keyboard punctuation. No em-dashes, no split clauses, no dramatic breaks.”
This forces the model to rethink the sentence instead of just swapping characters.
4. Clean Up With Regex (If Needed)
If you’re working at scale, use a post-processing step to search for — in your output and flag it.
Don’t just delete it, as that will break the structure. Flag it, then trigger a rewrite of that line or paragraph. Otherwise, you end up with sentence fragments that read like someone tripped over the keyboard.
Why You’ll Never Hit 100%
AI models aren’t built to follow hard rules unless you give them strict boundaries or reinforce the behaviour over time.
You can reduce em-dash usage by 90 to 95 percent, but the odd one will always try to sneak back in. It often happens when a model rewrites content mid-stream, expands a section on its own, or generates long-form answers.
This is how the model works (it’s not bug, no matter how frustrating it can be). It’s guessing what comes next based on patterns in the data, and em-dashes are everywhere in that data.
Wrap-up
If you’re building content systems that are meant to sound human, em-dashes work against you. They’re clunky, make your content appear as AI-written nonsense, and break the rhythm of natural thought.
Train your model to cut them out and provide examples that demonstrate the correct approach. Build tone instructions into the core of your prompt, and not as a line at the end. And use tools to catch anything that slips through.
Your writing will feel sharper, sound more real, and stay consistent, even when you’re scaling content across thousands of posts.
Need help to create human content using AI tools like ChatGPT? Feel free to join our coaching community and our team can help you directly via our support calls: https://www.pixelhaze.academy/membership