Did ChatGPT stop using em dashes?
Short answer: the ChatGPT app mostly did. OpenAI's API models didn't. We sent the same 45 prompts to 11 OpenAI models on 2026-11-13. ChatGPT's default model used 0.75 em dashes per 1,000 words; GPT-5.5 through GPT-6.1 in the API used 3.78 to 9.16. Humans answering the same questions: 0.21.
Em dashes by model
| Model | Released | Em dashes / 1,000 words | Answers with ≥1 em dash | Words |
|---|---|---|---|---|
| chat-latest | ChatGPT default | 0.75 | 16% | 22,607 |
| gpt-4o-mini | 2024 | 1.37 | 16% | 9,502 |
| gpt-5.4-mini | Mar 2026 | 1.81 | 29% | 10,518 |
| gpt-5.5 | Apr 2026 | 3.78 | 47% | 10,310 |
| gpt-5.6-sol | Jun 2026 | 7.02 | 67% | 7,696 |
| gpt-5.6-luna | Jun 2026 | 5.38 | 56% | 9,288 |
| gpt-6-astra | Sep 2026 | 6.97 | 73% | 9,039 |
| gpt-6-sol | Sep 2026 | 5.91 | 44% | 4,909 |
| gpt-6-luna | Sep 2026 | 9.16 | 71% | 5,674 |
| gpt-6.1-sol | Sep 2026 | 7.98 | 73% | 8,519 |
| gpt-6.1-terra | Oct 2026 | 4.3 | 40% | 5,344 |
| Humans (HC3 answers, same questions) | n/a | 0.21 | n/a | 65,983 |
| ChatGPT, Dec 2022 (HC3) | 2022 | 0.06 | n/a | 69,297 |
What this means
- Text pasted from the ChatGPT app now rarely has em dashes. 84% of its answers in our test had none.
- Text from tools built on the API (writing assistants, agents, SEO and email tools) often still does: 73% of GPT-6.1 Sol answers and 73% of GPT-6 Astra answers had at least one.
- Other studies measure differently. Graphite reported GPT-6 Astra well below the human rate with its own prompts and settings, so treat any single number as a snapshot.
What took the em dash's place in ChatGPT
ChatGPT's default leans on "rather than": 0.75 per 1,000 words, against zero for GPT-4o-mini. Wikipedia added "Y rather than X" corrective disclaimers as a mid-2026 sign of AI writing.
| Model | "Not X, but Y" family / 1,000 words | "rather than" / 1,000 words | Answers with ≥1 |
|---|---|---|---|
| chat-latest | 0.49 | 0.75 | 22% |
| gpt-4o-mini | 1.16 | 0 | 20% |
| gpt-5.4-mini | 0.67 | 0.19 | 11% |
| gpt-5.5 | 0.78 | 0.39 | 13% |
| gpt-5.6-sol | 0.13 | 1.04 | 2% |
| gpt-5.6-luna | 0.75 | 0.43 | 13% |
| gpt-6-astra | 0.44 | 0.89 | 7% |
| gpt-6-sol | 0.81 | 1.02 | 9% |
| gpt-6-luna | 0.7 | 0 | 9% |
| gpt-6.1-sol | 0.35 | 0.47 | 4% |
| gpt-6.1-terra | 0.37 | 0.56 | 4% |
See "it's not X, it's Y" and Wikipedia's signs of AI writing.
Still have em dashes to remove?
Older chats, API tools and Claude output still carry them. The free em dash remover fixes them in your browser. For everything else, run the AI slop detector or let the unslop tool rewrite the text.
Never use em dashes (—) or en dashes (–). Use commas, periods, colons or parentheses instead.
Free: 3 AI rewrites a day (up to 300 words) and unlimited quick clean. Need more? A $5 pack of 50 rewrites never expires, or go Pro at $12/month · $79/year. See plans.
FAQ
Did ChatGPT stop using em dashes?
In the ChatGPT app, mostly yes: its default model used 0.75 em dashes per 1,000 words in our November 2026 test, and 84% of its answers had none. But the same-generation models in OpenAI's API did not: GPT-5.5 through GPT-6.1 used 3.78 to 9.16 per 1,000 words. Humans answering the same questions used 0.21.
Why does ChatGPT use fewer em dashes than the API models?
OpenAI tunes the ChatGPT default (and its system prompt and personality settings) separately from the raw API models. Since GPT-5.1 in November 2025, ChatGPT also follows a "no em dashes" instruction in Custom Instructions. Apps and tools built on the API (writing assistants, agents, SEO tools) get the dash-heavy default unless the developer forbids it.
Which AI still uses em dashes?
In our test, GPT-6 Luna (9.16/1k), GPT-6.1 Sol (7.98), GPT-5.6 Sol (7.02) and GPT-6 Astra (6.97) used the most. Third-party studies in 2026 reported Claude near or above the human rate and Gemini nearly at zero, and Graphite reported Claude Opus 5.5 cut them about 99% versus Opus 5.
Is an em dash still a sign of AI writing?
Weaker than before. Wikipedia's Signs of AI writing guide now lists em dash overuse under historical indicators (2022–September 2026). Our data says that's true for text pasted from the ChatGPT app, not for text from API-based tools. Look for clusters of signs instead.
How do I remove em dashes from text I already have?
Paste it into the free em dash remover at gounslop.com/remove-em-dashes. It picks commas, periods, colons or parentheses from context and keeps number ranges like 2019–2023.