An em dash remover that reads each dash firstRuns in your browser, no account
Em dash remover that reads each dash before it removes it
Paste your text. This em dash remover shows how many dashes you use per 1,000 words next to the measured figures for people and for models, then handles each dash by the job it is doing in the sentence. Number ranges stay as they are.
- Published human prose runs 6.43 dash constructions per 1,000 words. Zero is not the target.
- GPT-4.1 measured 10.62 per 1,000, against a human control of 3.23. The tell is density, not presence.
- An em dash does several different jobs. One replacement character is right for one of them.
- Ranges like 2020 to 2024 written with a dash are correct typography and are kept.
Try one:
·per 1,000 words
Paste something and this fills in.
What your em dash density actually means
The number above is dashes per 1,000 words. A raw count would mostly tell you how long the document is, so the figure is normalised before it is set against anything. Below is what other writers and other models score on the same measure.
| Source | Per 1,000 words | Best read as |
|---|---|---|
| Llama 3.1 8B | 0.00 | Language model |
| Human control | 3.23 | Human writing |
| Gemini 2.5 Pro | 3.53 | Language model |
| Published prose (pooled) | 6.43 | Human writing |
| DeepSeek V3 | 6.95 | Language model |
| Claude Opus 4.6 | 9.09 | Language model |
| GPT-4.1 | 10.62 | Language model |
| SlopDetector flags above | 20.00 | Detection threshold |
Two things in that table decide how to use this page. Llama 3.1 8B produced none at all and Gemini 2.5 Pro produced 3.53, which lands among the human figures, so a low count proves nothing about who wrote the text. And individual books in the prose corpus ran as high as 10.13. Drawing an AI line at five or six would convict two celebrated American novels.
Source: SlopDetector, measured dash density across human and AI texts.
How the em dash remover classifies each dash
Each dash is read in context and sorted into one of these, then offered the repair that fits. Very long documents are capped at the first 300 cards, and the tool says so when that happens. You can override any of them, and a paired aside moves as one decision because it is one construction.
| Role | How it is spotted | Default repair |
|---|---|---|
| Number range | Digits or dates on both sides | Kept |
| Interrupted speech | Dash immediately before a closing quote | Kept |
| Attribution | Dash opening a line before a name | Kept |
| Paired aside | Two dashes bracketing a clause in one sentence | Commas, both halves together |
| Introduces a list | Followed by a list or a restatement | Colon |
| Abrupt turn | Followed by a clause that could stand alone | Semicolon |
| General break | None of the above | Comma |
Worked examples
One sentence per role, seven in all. Click any of them to load it into the tool above and see the decision it produces.
When to keep an em dash
Keep it when nothing else does the job. A range written with a dash is correct, and a replacement applied to every dash alike will change it along with the rest. A line of dialogue cut off mid-sentence needs the dash, because that interruption is what the mark means. An attribution under a quotation has used one for two centuries.
Keep it, too, when the alternative is worse. A sentence rewritten to avoid a dash and left with four commas is harder to read than the dash was. The tool counts what you kept and says so, because a report that only counts removals quietly argues that removal is always the win.
One caution worth stating plainly, since much of the writing on this topic skips it: a low dash count is not much of a disguise on its own. The tells usually described as harder to shift are uniform sentence length and stock phrasing. The four-direction rewrite on the home page is the part of this site aimed at those, and the writing notes cover which patterns are worth the effort.
Why one replacement character is not enough
On 15 August 2026 we opened three of the em dash tools that come up for this search and wrote down the options each one puts on screen. Convert Case offers a dropdown of hyphen, comma, space or nothing. GPT Cleanup Tools offers a similar set. StyledText offers six, adding semicolon, parentheses and a custom string. In each case the interface takes one choice and no per-dash control is offered, and none of the three showed a density figure on the page we saw. We read their interfaces; we did not benchmark their output.
That is a fair design for a find-and-replace and a poor one for punctuation, because the dash in between 2020 and 2024 and the dash in she left early, it had been a long week are not the same mark doing the same work. Pick comma and the range breaks. Pick hyphen and the turn reads as a typo. A paired aside is worse again: two dashes are one construction, so a tool that replaces them independently can open with a comma and close with a semicolon.
What they report back is how many dashes were replaced, which is a count of the work done rather than a fact about your writing.
Questions people ask about removing em dashes
- Should I remove every em dash from my text?
- No. Stripping all of them overshoots in the other direction. In a 702,939-word corpus of published prose the pooled rate is 6.43 dash constructions per 1,000 words, with individual books running from 3.47 to 10.13. Twain and Melville both sit above 8. A text with zero dashes is as unusual as one with twenty. Aim to land in the human range and fix the dashes that are doing a job some other mark does better.
- What is a normal em dash density?
- The human control in the Freeburg 2026 study came in at 3.23 per 1,000 words on matched prompts. GPT-4.1 produced 10.62, Claude Opus 4.6 produced 9.09, and Gemini 2.5 Pro produced 3.53, which sits inside the human range. SlopDetector itself only flags text above 20 per 1,000. The signal is density, and it is weaker than the internet believes.
- Does this em dash remover send my text anywhere?
- No. The counting and every suggestion run in your browser tab. There is no upload, no account, and no length limit imposed by a server, because no server sees the text.
- What is the difference between an em dash, an en dash and a hyphen?
- A hyphen joins words: well-known. An en dash spans a range: 2020 to 2024, written with the shorter dash. An em dash, the longest of the three, breaks a sentence open for an aside or a turn. This tool treats the range case as correct and leaves it alone, which a single replacement applied to every dash cannot express.
- Why does ChatGPT use so many em dashes?
- Because the formal edited English it trained on uses them. The dash is a marker of edited prose more than a marker of machine writing, which is why it produces so many false accusations. Reporting on newer models puts the figure lower, which would make this a tell that decays rather than one you can rely on. We have not measured that ourselves.
- Will removing em dashes get my text past an AI detector?
- Treat that as unlikely, and this tool makes no such claim. Punctuation is among the easiest things to change about a text, and a signal that a single instruction can switch off is a weak one to build a case on. Writing on detection points instead at perplexity and burstiness, meaning sentence-length variation: SlopDetector reports that text sitting between 18 and 24 words a sentence, paragraph after paragraph, is the pattern that survives a punctuation edit. We have not tested any detector ourselves and make no prediction about what one will say.
- How do I remove em dashes from ChatGPT or Claude output specifically?
- Paste it here. The source does not change what the dashes are doing, and the repair depends on the grammar around each one rather than on which model produced it. What does differ is volume: GPT-4.1 output tends to arrive around 10.6 per 1,000 words, so expect more cards than a piece of your own writing would produce.