Why does ChatGPT use em dashes so much?

Models are trained on text that someone edited. Books, news, reports, essays — all of it went through a publisher, an editor or at least a style guide, and American style guides have required the em dash for parenthetical asides for generations. A model that mimics the average of its corpus is therefore mimicking edited prose, and edited prose is dash-rich.

The counter-pressure that exists in real writing does not survive the same way. Nobody edits published text to say "use fewer dashes here", because that is a document-level judgement. The corpus contains the dashes; it does not contain the reaction against them.

The em dash is the most flexible punctuation mark in English. It can stand in for a comma, a colon, a semicolon, or a pair of parentheses, and it does all four without changing a word around it. For a system that picks the next token by maximising local fluency, that flexibility is exactly what it needs: reaching for the dash is nearly always grammatical, whereas reaching for a semicolon sometimes is not.

In most modern tokenizers the em dash is a single token, and so is the space around it. Writing "the plan — risky as it was — worked" costs fewer tokens than the same sentence with (risky as it was) or a semicolon and a conjunction. This is a small effect and it is not a conscious choice, but it points the same direction as everything else: the cheapest path to a fluent sentence runs through the dash.

Models are shaped by human preference data, where raters compared candidate answers and picked the one they preferred. Text that uses dashes tends to read as more confident and more immediate than text that uses parenthetical commas, and raters rewarded that register. This part is inference rather than published fact, but it explains a pattern you can observe directly: ask for a neutral summary and you still get the same dramatic punctuation.

Every general-purpose model does this. Claude is often cited as the worst offender, the em dash habit being one of the most common complaints about it, and Gemini and Copilot show the same pattern to a lesser degree. That is what you would expect if the cause is shared training data plus shared decoding behaviour, which it is.

Counts are from repeated sampling, not vendor figures — the exact number depends on genre, length and sampling settings, which is why we give ranges and no false decimals. For calibration, the 13 edited posts on this blog average 6.7 em/en dashes per 1,000 words with the spread shown in do em dashes mean AI; every model above sits at or above that. Treat the ordering as the finding, not the digits.

Layer the fixes, because they fail at different rates. An instruction in the prompt is obeyed most of the time and forgotten in long conversations, or the moment the model is asked to be more engaging. Custom instructions persist across chats but still cannot override the model's default register completely. Post-processing is the only step with no failure mode, because it does not depend on the model at all.

1. Ask for the constraint in the prompt: "no em dashes, no en dashes, no double hyphens".
2. Persist it in custom instructions or personal preferences so it does not need repeating.
3. Strip the characters mechanically from the finished draft with the AI dash remover, which also matches typed -- and ---.