Why does ChatGPT use so many em dashes?

What is on record

Three things are documented.

The pattern was real. Wikipedia's editors, who clean up AI-written text on the encyclopedia, describe chatbots using the long dash far more than non-professional writers in the same genre, often in spots where a person would use a comma, parentheses or a colon, and often to punch up a clause the way sales copy does. They also noticed the dashes usually had spaces around them.

OpenAI changed the behavior. In mid-November 2025, days after GPT-5.1 arrived, chief executive Sam Altman posted on X that ChatGPT now does what it should when your custom instructions tell it not to use em dashes, as Ars Technica reported. OpenAI's release notes for GPT-5.1 say the models follow custom instructions and style preferences more reliably. Neither statement says why the dashes were there in the first place.

The habit has faded in newer GPT models. Graphite, a growth agency, compared 10,000 human web articles with 90,000 articles from nine models on the same topics, published in September 2026. Em dash use fell sharply from GPT-5 to GPT-5.6 Sol and stayed low in GPT-6 Astra, at about one-eighth of the human rate. ChatGPT's paid plans began moving to GPT-6 Sol on 7 October 2026, and no measurement of that model had been published when this page was written.

One small irony: OpenAI's own help page on ChatGPT personalities, read for this page, still shows sample replies with long dashes in them.

What is only guessed

Ars Technica's reporter put it plainly: no one knows precisely why language models overuse the dash. The explanations in circulation rank roughly like this.

Most plausible: models learn from huge amounts of edited writing, such as news, essays and editorials, where the dash is common, and a request for a professional tone pulls them toward that average style. It matches the one thing that is certain about training, that models reproduce patterns common in their data, but nobody has tested it for the dash in particular.

Possible, unproven: during training on human feedback, raters may have preferred answers with dashes because they read as more sophisticated. No rating data has been published to check this.

Weaker: the 19th-century theory, which leans on a 2018 study (cited by Ars) finding that dash use in English peaked around 1860, and the Medium theory, that models copied a blogging site that converts two hyphens into a long dash.

If you read a confident one-line answer somewhere else, it is one of these guesses stated as fact.

How to stop it in ChatGPT

Put the rule in Custom Instructions, not in a single chat. On the web and desktop app, open Settings, choose Personalization, make sure Enable customization is on and type into the Custom Instructions field. On iOS and Android it is Settings, then Customize ChatGPT. OpenAI's help page says the instructions apply to all your chats, and the field holds 1,500 characters on the Free and Go plans and 5,000 on Plus, Pro, Business, Enterprise and Education.

A line that works, 139 characters long: “Never use em dashes or en dashes. Use a comma, colon, period or parentheses instead. Hyphens only inside compound words such as well-known.”

Then test it. Ask for a 300-word piece and search the result for the long dash by copying one into the Find box. As Ars Technica explained, an instruction makes a dash less likely, not impossible, so check anything you plan to send.

The Base style and tone setting won't do this for you. OpenAI's personality page says that for an email draft or a social post, ChatGPT matches tone to your instructions and the context, not necessarily to the personality you picked. If you call the model through the API, put the same line in the system message.

How to fix a draft that already has them

Look at what each dash is doing, then pick the mark that does that job. In the examples, [long dash] stands for the character.

An aside in the middle of a sentence takes commas, or parentheses if it is minor. “The report [long dash] all 40 pages of it [long dash] arrived late” becomes “The report, all 40 pages of it, arrived late.”

An explanation or a list at the end takes a colon. “We changed one thing [long dash] the price” becomes “We changed one thing: the price.”

Two full sentences joined by a dash become two sentences. “Sales doubled [long dash] nobody expected it” becomes “Sales doubled. Nobody expected it.”

A dramatic turn is a different problem. “It isn't a phone [long dash] it's a lifestyle” was built for effect, and a comma keeps the effect. Rewrite it as a plain claim, or cut it.

A range of numbers, which takes the shorter en dash in print, reads fine in words: “pages 10 to 14”, “9 a.m. to 5 p.m.”

A rule of thumb for a chatbot draft: if more than one sentence in ten has a dash, fix them all, then put back the one you would really have written.

Is the long dash still a giveaway in 2026?

Less than it was. Wikipedia's field guide now files the em dash under historical indicators, dated 2022 to September 2026, and says it means something only alongside other signs. Its editors add, without a source so far, that newer models avoid the dash so strictly that a page of nothing but commas and periods has started to look machine-made as well. Graphite's authors reach a similar view from their numbers: GPT and Gemini may have overcorrected.

So a dash proves nothing about who wrote a text, and neither does the lack of one. Writers who have always liked the dash can keep it. What reads as generated is a dash in every paragraph, doing the same dramatic job each time.

As an example of a house rule, HidenGPT tells its drafting model to avoid the long dash almost entirely, and its red pen flags a draft with three or more at a density above one per 120 words, then swaps them for commas, periods or colons when it rewrites. That is a style choice for readability, not a detector rule.