Why is AI writing full of em dashes?
Because the training data is, and because the em dash is a low-risk connector. It joins two clauses without committing to whether the relationship is causal, contrastive or parenthetical. A model optimising for fluent, inoffensive prose reaches for it constantly.
The result is the single most commented-on tell in AI writing. It became prominent enough that OpenAI shipped a control for it in ChatGPT after sustained complaints.
Why find-and-replace makes it worse
Because the em dash does at least four different grammatical jobs, and one replacement cannot serve all of them. Swap every dash for a comma and you get comma splices, run-ons, and sentences that lost the pause they were built around.
The substitution table
Identify what the dash was doing, then replace accordingly.
| The dash is doing this | Replace with | Example after |
|---|---|---|
| Continuing a thought | Comma | “It survives copy and paste, which is the point.” |
| Introducing an explanation or list | Colon | “Three things break: diffs, keys, search.” |
| Joining two independent clauses | Full stop | “The premise is wrong. The deletion is real.” |
| Wrapping a true aside | Brackets or commas both sides | “The mark (added in August) is statistical.” |
| Trailing off, or emphasis padding | Delete it | Often the sentence is stronger without. |
The last row catches more cases than people expect. A good share of AI em dashes are decorative, and the sentence reads better once the clause is either promoted to its own sentence or cut.
The mechanical steps
- Search for the characters, not the concept. Em dash is
U+2014, en dash isU+2013. Both appear. A double hyphen--also shows up in some output. - Review each hit in context. In a normal article there are rarely more than twenty. This takes minutes.
- Apply the table above. Do not batch replace unless you are willing to reread every sentence.
- Check for the second-order tell. Text that had its dashes stripped often ends up with too many short sentences in a row. Vary the length back.
Fix the prompt instead
Cheaper than cleaning output. A custom instruction along the lines of “do not use em dashes; use a comma for a continuing thought, or a full stop if it should be a separate sentence” works in most tools, and ChatGPT now honours it more reliably than it used to.
That has a real advantage over post-processing: the model composes the sentence around the constraint rather than having punctuation swapped in afterwards, so the rhythm survives.
Does removing em dashes hide AI authorship?
No. It removes one stylistic signal that humans notice. It does nothing to Anthropic’s statistical watermark, which lives in token selection rather than punctuation, and it does not defeat detectors that key on sentence structure, hedging density and topic framing.
Remove em dashes because you or your publication dislike them. That is a good enough reason and it happens to be the only honest one. See what Claude’s text watermark actually is for what punctuation changes do and do not affect.
Who this is not for
Anyone who actually likes em dashes. They are correct punctuation with a long history, and plenty of good writers use them heavily. The case for removing them is stylistic and situational, not a rule.