← Back to recent questions

Three books and counting, and the same fight every time. I am in the middle of the manuscript of Fragile by Design at the moment, and the em-dash habit of the AI tools I use in the editing pass keeps trying to insert itself back into prose I have spent years stripping it out of. Most readers would not notice. Some readers, increasingly, do.

Three places to fix it. Add one line to the bottom of every editing prompt, put the same line in ChatGPT's custom instructions so it applies to every conversation, or build a small Custom GPT with the rule baked in. The first is the cheapest, the third is the most reliable, and none of them is perfect. The model has been trained on millions of pages that use em dashes, and one will still slip through.

Why it matters

You are right that em dashes are correct in professional writing. The problem is perception. A lot of readers now read an em dash as the AI's signature, the way they once read "delve into" or "in conclusion". If your writing is being judged on whether it looks AI-written, the dashes are worth removing. The same is true on this site, where the no-em-dashes rule has been in the style guide since the start. Not because em dashes are wrong. Because they have become a tell.

The three fixes

The one-line fix. At the end of every editing prompt, paste the line below. The "hyphens are fine" sentence stops it overcorrecting and stripping the hyphen out of "well-written" or "AI-generated".

Do not use em dashes (—) anywhere in the output. Do not use en
dashes (–) either. Use commas, full stops, parentheses, or colons
instead. Hyphens (-) in compound words like "well-written" or
"AI-generated" are fine and should stay.

If you are tempted to use a dash, restructure the sentence.

The permanent fix. Open ChatGPT, go to Settings, then Personalization, then Custom Instructions. In the box that asks how you want ChatGPT to respond, paste the same sentence. From then on, every new chat starts with that rule in place. Claude and Gemini have the equivalent setting in their own preferences pages, and the same wording works in both.

The reliable fix. If editing is a regular job rather than a one-off, build a Custom GPT. Name it "My Editor", paste your usual editing brief into the instructions box, and put the no-em-dashes rule near the top. You open that GPT instead of plain ChatGPT and the rule is already there. Claude Projects does the same job in Claude, with the bonus that you can load in a few examples of your own writing to tighten the editing pass on more than just punctuation.

The honest catch. Even with all three in place, a dash will leak through on long edits. A final pass with your text editor's find-and-replace closes the gap. Search for the em dash character, replace each one with a comma, full stop, or pair of brackets depending on the sentence. The same trick catches the en dash, which some models default to as a workaround once you have banned the em dash.

What I would not do

I would not ask ChatGPT to "write more like a human". Specific instructions outperform vague ones. "Do not use em dashes" is specific. "Sound human" is not, and the model's guess at what that means may not match yours.

The bigger thing

The em dash is not really the question. The question is what other tells the AI tools have, and how long any specific one of them will last. Five years ago people noticed "delve into". Now it is the dash. In a year it will be something else, possibly something none of us has yet identified. Tools fix obvious tells once they are publicly named, drift into a new habit, and the cat-and-mouse of "does this read AI-written" continues.

The defensive move, if you write for readers who care, is not to memorise the current list of tells. It is to develop your own voice and protect it from the tools, rather than reading what the tools wrote and trying to make it sound less like a tool wrote it. The em dash rule is a small example of a bigger discipline. Write your own first sentences. Use AI to tighten them. Never the other way around.

As of May 2026. Tool features and pricing change quickly; if you are reading this much later, check the current state before relying on the specifics.