The short answer
If you want AI to stop writing em dashes, asking helps. In my experience, it never gets you all the way to zero.
What got me to zero was a check at the very last step, right before the text goes out.
If you build with AI, that check is a line of code. If you write in ChatGPT or Claude, that check is you, and it takes about ten seconds. Both are in the steps at the bottom.
Three marks come up a lot here. The em dash is the long one, about as wide as a capital M. The en dash is the middle one you see in ranges of numbers. The hyphen is the short one on your keyboard, the one that holds a word like set-aside together.
Everything below happened while I was writing with AI every day, mostly Claude.
I used to like them
I liked em dashes. I'm a designer, and I learned the shortcut for one over a decade ago, when I got into typography. They looked good on the page. A little bougie, so I only used one now and then.
Then AI started putting them in everything, and they turned into a giveaway. I stopped using them and added them to my ban list.
ChatGPT never really stopped for me. I asked over and over. Even after I trained it pretty well, they kept showing up.
The em dash is probably part of why I switched to Claude. It got much better at it, like an annoying habit you can actually break. It still slipped sometimes, though. That's why the rule ended up in code.
I wrote about spotting them in The Em Dash Epidemic in November 2025. I'd started noticing AI writing everywhere in 2023, and the dashes were how I learned to see it. This is the next part: keeping them out.
It wasn't just me seeing things. A 2026 preprint, not yet peer reviewed, looked at about 70,000 medical research papers that had been posted before review. Before ChatGPT, about 4 in 100 had an em dash in the discussion section. Afterward it was closer to 12 in 100, and in 2025 alone, about 20.
The study is careful about what that means. It shows a shift across a whole field of writing. It can't tell you whether one paper was written with AI.
If a study that size can't judge one paper, one dash in your post can't either.
Nearly 60 times
My chat history with Claude goes back to early June 2025. By the middle of that month, the rule is already sitting in my prompts.
Nine days later I asked for it to be built in somehow, so it would stick.
It didn't stick. I went back through my chats and read every short message where I mentioned dashes. Nearly 60 of them tell it to drop the em dash, in 54 different chats. Most landed in October and November 2025. The last one is from May 2026.
My record has gaps, so that's a floor. It also gets much thinner in 2026. So the drop doesn't prove it learned. It could be better models, or just fewer chats on record.
My ChatGPT master prompt had the same rule by September 2025.
I wasn't the only one stuck on this. On November 13, 2025, the same day I published the epidemic piece, OpenAI's CEO posted that if you tell ChatGPT not to use em dashes in your custom instructions, "it finally does what it's supposed to do."
An engineer who wrote about it put it simply. It's surprisingly hard to prompt models out of em dashes, and nobody really knows why they love them so much.
It swapped one tell for another
When it did listen, it found a way around me.
In October 2025 hyphens started showing up where the em dashes used to be. A hyphen with a space on each side, in the middle of a sentence, is a stand-in for the dash. I had to ban those too.
In May 2026 I had a script clean up my framework library. It went through 150 frameworks and made 3,937 swaps. 2,551 became commas. 1,343 became colons.
So a third of them turned into colons.
This part is my read, not a measurement. Replace one mark with one other mark everywhere, and the tell moves with it. A page full of colons where the dashes used to be is its own pattern.
I watched the same thing happen with images that August. Telling a model to make a picture look less like AI just traded one even pattern for another. What worked was pulling rules back out and looking at the result.
The day the prompt lost
In May 2026 I built an Android app that runs one of my frameworks. It talks to Claude through a small server I control, a middle step between the app and the AI.
One of its first real answers came back with an em dash in it.
So we added a strong rule to the instructions that server sends with every request: never use em dashes.
The next answer still had one, right between a score and its label.
So we put a replace in the server's code. Every em dash gets swapped out before the text reaches the app, on every screen. That fixed it completely.
Ars Technica explained why, writing about that post. A normal program follows a rule like "don't include character X" exactly. For a language model, the same request shifts the odds. It isn't a hard rule. Its own quick test found the custom instruction did seem to work.
So the line in the prompt lowers the odds. The line in code is the guarantee. My own writing rules say that now, in so many words: enforce it in code where the text comes out.
The chat assistant on my sites carries both.
// In the prompt: it asks
STYLE: Never use the long em dash punctuation in your replies. Use a period or a comma instead. This is a firm brand rule.
// In the code: it makes sure
$response = preg_replace('/\s*(?:\x{2014}|\\\\u2014)\s*/u', ', ', $response) ?? $response;A rule I have to keep repeating is still just a request.
Then I had 25 of them
Once code worked, new tools started getting their own fix. My AI assistant. The scripts behind my daily videos. My newsletter. The tool that turns my voice into text.
By November 2026 there were 25 separate pieces of code for this one rule. Some fixed the dash. Some only checked for it. They didn't even agree on the answer. Most turned the em dash into a comma. Two turned it into a hyphen.
Some of it was funny.
- A cleanup pass rewrote the page standard for this site. Its list of characters to ban turned into four plain hyphens. It's fixed now.
- An em dash kept slipping into the search pattern of one of my own checkers. The shared tool I built later spells every dash in code, so the file can pass its own check.
- Claude Code caught a chat putting an em dash in the subtitle of my article about AI tells.
- WordPress turned a plain hyphen into an en dash after the text left my hands. It does that by default when it shows a page, to a hyphen with a space on each side. That's a normal style in plenty of places. It just wasn't what I wrote.
On November 13, 2026, I asked how many ways we had to get rid of dashes. The answer was 25. We built one shared tool for all of them that day.
Nothing has moved over to it yet. Three days later a new mail tool shipped with its own one-line copy.
Where the rule stops
The rule is for what goes out under my name.
When I'm talking to Claude, I don't mind the dashes. If a post is written from the AI's own side, that's its voice, and I don't mind there either.
Hyphens that grammar needs stay, like the one in a 12-year habit. The model taught me where a few of those go. I thought it was making mistakes until it explained.
If you write with em dashes and someone calls your writing AI, they're wrong about you. One dash proves nothing on its own.
And I still think tells are choices. In April I wrote that if you love em dashes, use them. Just know how they read right now.
I once asked whether Emily Dickinson would pass my own human touch check, with all her dashes. As a creative tool, dashes are great. Just not in my posts.
How to make a writing rule hold
Say it once, where it loads every time
For everyone. One setting beats retyping it in every chat.
Put the rule where every chat starts. In ChatGPT on the web, that's Settings, then Personalization, then the Custom Instructions box. In Claude, it's Settings, then Your Claude, then Instructions. If you build with code, it's the rules file or system prompt your app sends.
Write it plainly: "Never use em dashes. Use a comma, a period or parentheses instead."
It lowers the count. Expect it to slip anyway.
When to do it: day one.
Check before you post
For everyone. In a chat window, you're the last step.
You can't add code to a chat window, so the check is you. Before you post, open Find in your editor and search for the long dash. The easiest way to get one is to copy it out of the AI's own reply. It'll give you plenty.
Fix each one by hand. It takes about ten seconds.
When to do it: every time, before you hit post.
Count the repeats
The third fix by hand is the signal.
Every time you fix it yourself, that's a repeat. The third time, stop asking. If you build with AI, move the rule into code. If you don't, make Step 2 a habit you never skip.
I didn't. I asked nearly 60 times.
When to do it: the third time you catch it.
Put it where the text comes out
If you build with AI. Fix it on the way out, and stop it at the door.
Find the one place every piece of text passes through. The server between your app and the AI. The script that writes your documents. The feed into a voice tool. The last step before something publishes.
Do two things there. Replace the dash on the way out. Then run a check before anything publishes, and have it stop with an error if one got through. My article build flags a single em dash before anything goes up.
# Fix it on the way out
text = text.replace("\u2014", ", ")
# Stop it at the door
if "\u2014" in page:
raise SystemExit("em dash found, not publishing")When to do it: before anything goes out that you won't read first.
Pick the swap on purpose
For everyone. Then read what it turned into.
Decide what the dash becomes. A comma for an aside. A period between two thoughts. A colon before a list. Parentheses for a real side note.
Leave code and data alone. When I asked for a cleanup in November 2025, I said not to take dashes out of the code.
After any bulk fix, read a few pages back. If one mark is suddenly everywhere, you made a new tell.
When to do it: after any bulk cleanup.
Keep one copy
If you build with AI. One rule, one piece of code.
When a second tool needs the rule, have it use the same code as the first. Every copy you add is one more place the rule can drift.
That's how I ended up with 25 versions and two different answers.
When to do it: the second time a tool needs it.
What's still open
There's no hook yet that checks every file Claude writes, the moment it writes it. The prompt rules, the fixes where text comes out and the checks before I publish cover most of what goes out. The shared tool exists, and nothing uses it yet.
A hook is the obvious next piece. Whether I build it is still open.
What's the tell you keep fixing by hand?
Turn the fix you keep repeating into a rule
How to Framework walks through writing your own rules down, one domain at a time.
Explore How to FrameworkSources
- Mike Goetz, The Em Dash Epidemic, RageDesigner, November 13, 2025
- OpenAI’s CEO, post on X about custom instructions and em dashes, November 13, 2025, read November 16, 2026
- Ars Technica, report on ChatGPT following the em dash instruction, November 14, 2025, read November 16, 2026
- Em-ergence of the em-dash: a population-level rise in em-dash frequency in medRxiv preprints at the dawn of the large-language-model era, arXiv:2606.29540, preprint submitted June 28, 2026, read November 16, 2026
- Why do AI models use so many em-dashes?, an engineer’s blog, October 30, 2025, read November 16, 2026
- WordPress Developer Resources, wptexturize(), read November 16, 2026
- OpenAI Help Center, ChatGPT Custom Instructions, read November 16, 2026
- Claude Help Center, Understanding Claude’s personalization features, read November 16, 2026
- My own chat record with Claude, June 2025 to November 2026, counted by script; and the code on my own machines and server, read November 16, 2026
