In November 2025, Sam Altman announced a small victory — ChatGPT could finally follow instructions not to use em dashes. For a company working towards artificial general intelligence, getting a chatbot to stop using a particular punctuation mark seems like a rather modest achievement. But the em dash (—) had become one of the most recognisable signs of AI-generated writing. Spot a few in an article, LinkedIn post or email, and someone will probably suggest ChatGPT wrote it. Some people have even started replacing them with shorter dashes to make their writing appear more human — which is quite funny, considering that AI learned to use them from us.
The em dash was here first
There are three punctuation marks worth knowing about:
- Hyphen (-): The shortest one, used to connect words like well-known or AI-assisted.
- En dash (–): A little longer, commonly used for ranges such as 2020–2026.
- Em dash (—): The longest, used for interruptions, pauses or additional thoughts.
Their names come from traditional typesetting. An em was a unit of measurement related to the size of the type, and an en was half that width. Matthew Butterick's Practical Typography has a useful explanation.
The dash has a long literary history. Laurence Sterne used it extensively in his eighteenth-century novel Tristram Shandy, to the point that he became known for his unusual punctuation. Emily Dickinson was equally fond of dashes in her poetry, using them to create pauses and connections between thoughts. Publishing conventions differ, too: American publishers generally favour an em dash without spaces, while British publishers often use a spaced en dash – like this. Both are perfectly legitimate.
An em dash can also do things that commas and brackets don't quite achieve. Compare "I thought it was a terrible idea, until I saw the results" with "I thought it was a terrible idea — until I saw the results." The second version gives the change of thought more emphasis. Whether that's desirable depends on the sentence.
Then came ChatGPT. AI models were trained on enormous collections of text, including books, websites and articles. The original GPT-3 paper explicitly listed two large book datasets among its training sources. Naturally, the models learned punctuation habits from human writing — and apparently became rather enthusiastic about the em dash. A 2026 Pew Research Center study analysing 490,000 English-language webpages found that em dashes appeared about twice as frequently as in 2023. Certain words favoured by AI, including delve, interplay and testament, had also become much more common.
Of course, an em dash alone proves nothing about who wrote a sentence. Neither does a suspiciously enthusiastic use of the word delve. Still, I understand why people have become irritated by AI-generated writing. I've had to rewrite more of it than I'd care to admit.
I use AI to write. And I think you should too.
I use AI quite extensively when working on articles, including this one. It's excellent for research. I can ask it to find the original source of a statistic, explain something unfamiliar, suggest examples or challenge an argument. It also helps with structure, grammar and finding better ways to express an idea.
What I don't particularly enjoy is reading an entire article written by AI without much human intervention. The writing often looks fine at first — grammatically correct, carefully structured, full of explanations. Then you realise you've read five paragraphs that could have been expressed in two sentences.
I've written a PhD thesis, so I know it's possible to write an awful lot of words. Writing something short, while preserving everything important, is considerably harder. That's where I find AI particularly useful. I ask it for a TLDR, then a TLDR of the TLDR, and on a brave day a TLDR of that. If an idea survives all three rounds, it probably belongs in the article. If it quietly disappears and nobody misses it, it probably never did. But you still need to read what it produces. AI can misunderstand your argument, invent facts, misrepresent sources or remove the very thing that made your writing interesting. The responsibility remains with whoever publishes it.
A few things that have helped me
1. Show it what good writing looks like. Telling AI to "write naturally" or "sound human" isn't particularly helpful. Everyone has a different idea of what that means. Give it examples of articles you like. Better still, provide examples of your own writing, including passages you think are good and ones you don't like, and explain why.
2. Give your AI some context. Create a dedicated project for your writing, with instructions describing your preferred tone, sentence length, structure and things to avoid. Include examples of your writing and links to authors you enjoy. Tell it who your audience is, what you're trying to achieve and how much research you expect. The more useful context you provide, the less time you'll spend correcting the same problems.
3. Try different models. ChatGPT, Claude, Gemini and other models don't necessarily write the same way. Even different versions of the same model can produce noticeably different results.
4. Keep asking for a TLDR. This is probably one of my favourite tricks. Write something, then ask AI to summarise it. Ask again, but shorter.
5. Don't publish the first draft. Read it, ideally aloud. Check the facts and follow the links. Ask yourself whether you agree with the argument. I also find it useful to ask AI to criticise what it has written, identify unsupported claims or explain what it might have missed. And if a paragraph sounds like AI wrote it, rewrite it. Changing the em dash to an en dash won't make it any more interesting.
Even the title came from AI
Speaking of useful AI assistance, I asked ChatGPT to suggest titles for this article. Here are five of its suggestions:
- The Em Dash Is Innocent Until Proven AI
- The Em Dash Didn't Ask for This
- The Em Dash Was Here First
- A Dash of AI
- The Em Dash and the Art of Looking Human
I mean, come on — those are pretty good!
As for the em dash, I intend to keep using it whenever it makes sense. I'd go one step further and thank ChatGPT for actually educating me about our lovely punctuation choices and the centuries of history behind them.