Why ChatGPT uses so many em dashes

Every heavy ChatGPT user has had the same moment: scrolling a draft and realizing the em dashes (—) are doing structural work on every paragraph. It is not your imagination, and it is not a setting you left on. Models genuinely write this way, for reasons that fall out of how they are built.

What the em dash actually is

Three horizontal marks, three jobs. The hyphen (-) joins: well-known, up-to-date. The en dash (–) marks ranges and joints: 1914–1918, Bose–Einstein. The em dash (—) interrupts: it starts asides, delivers punchlines, and stitches two half-thoughts into one long sentence. Good prose uses it sparingly because each one asks the reader to hold the sentence open a beat longer. A draft where every paragraph leans on that trick reads breathless, which is exactly the ChatGPT texture people have learned to spot.

Why models reach for it

A language model predicts the most likely next token, and its sense of likely comes from its training diet: books, journalism, essays, marketing copy, and a great deal of internet writing where the em dash is the favorite tool for sounding thoughtful. Fine-tuning sharpens the habit further. Human raters consistently score fluent, rhythmic, slightly elaborate answers above plain ones, so the tuning process rewards the dash-heavy cadence and the model learns that this rhythm pleases. The em dash is also a safe choice under uncertainty. When the model is unsure how two clauses relate, an em dash connects them without committing to the logic a conjunction would require. It is the punctuation of confident vagueness, and a prediction machine finds it constantly useful.

The filler words that travel with it

The dash rarely arrives alone. The same training and tuning pressures promote a small court of favorite words: delve, tapestry, landscape, crucial, moreover, furthermore, in today's fast-paced world. None is wrong on its own. Together, at model frequency, they are as recognizable as the dashes. If you are editing AI text back into your own voice, search these along with the dashes: replacing the characters takes seconds, but the words are what make a paragraph sound generated after the punctuation is fixed.

What cleaning can and cannot do

Normalizing the characters is the easy half. Paste the draft into the text cleaner and the long dashes become hyphens, commas, or stops as their context dictates. The word choice is the other half, and no cleaner touches it: swapping delve for dig into is editing, not formatting. And neither half hides authorship. Detectors judge how writing reads, not which dash it uses, as what AI watermarking actually is explains in detail. Fix the dashes so the text fits your style guide. Rewrite the tells so it sounds like you.