The em-dash problem
ChatGPT reaches for the em dash more than most human writers do, and that habit has become one of the most recognizable tells of AI-generated text. Here is why it happens, when em dashes are completely fine, and how to stop a draft from reading as machine-written.What an em dash actually is
An em dash is the long dash used in English to mark a stronger break than a comma, but less terminal than a period or semicolon. It sets off parenthetical remarks, creates dramatic pauses, or signals a shift in thought. The Chicago Manual of Style treats it as a tool of emphasis, and the APA style guide advises using it sparingly.In practice, most human writers use em dashes irregularly. One em dash every page or two is normal. Em dashes in every paragraph, or several in one paragraph, is unusual outside specific stylists.Why ChatGPT reaches for it
Training data is the proximate cause. Large language models learn from a corpus that over-represents polished web content, such as journalism, long-form blogs and marketing copy, where em dashes appear more often than in casual writing. When a model optimizes to produce clean, professional prose, it pulls toward patterns that reviewers and raters scored as clean and professional. Em dashes are one of those patterns.There is also a structural reason. Language models tend to generate sentences that build toward a point and then tack on a polishing clause. The em dash is the perfect connector for that rhythm: it lets the model append a refining thought without risking a grammatical mistake. Paragraphs end with em dashes because the structure of the paragraphs pulls toward them.Why it matters for detection
Em-dash density is one of the writing signals used in AI scoring. The count alone is not enough, since plenty of human writers use em dashes, but em dashes appearing regularly as the dominant non-period punctuation is a strong signal. A piece with three em dashes in one paragraph and none in the next, where the dashes always append a parallel clause, reads as AI even before you look at vocabulary.Leap's free AI detector scores em-dash density alongside other signals such as burstiness, stock phrases and hedging words, and highlights the sentences that weigh most. It is one of the clearest instant-tell patterns because it is visible to the naked eye. We cover the other signals in our guide to how AI detectors work.The specific em-dash patterns that flag
- The polish pattern. "X does A, and it also does B" written as "X does A, with B tacked on behind a dash." Appending a parallel clause after a dash is a very ChatGPT move.
- The hedging pattern. "This is a great approach, though it has limitations," where the hedge is appended via em dash. This shows up constantly in LLM output.
- The emphatic pattern. "The answer is simple: do nothing," delivered instead as a dramatic dash reveal, repeated paragraph after paragraph.
- The paired em-dash pattern. Parenthetical dashes setting off a clause mid-sentence are fine in isolation; en masse, they are a tell.
When em dashes are fine
There is nothing wrong with em dashes. Great human writers use them liberally: Emily Dickinson, Cormac McCarthy, John McPhee. The problem is not the punctuation itself but the frequency and predictability of use. If you genuinely write with em dashes, keep writing with em dashes. If you noticed em dashes appearing in a draft that were not part of your voice, that is the AI showing through.How to fix it
In a quick pass, convert most em dashes into alternatives:- Comma or semicolon for softer pauses.
- A period plus a new sentence for harder breaks.
- Parentheses for actual asides.
- A colon for reveals, as in "The answer is simple: do nothing."