Yes. With no instructions, GPT-5.6 Terra and GPT-5.6 Luna used about 4.4 em dashes per 1,000 words. That is about one in every 230 words, and 62% of their drafts had at least one. It is less than half the rate of Claude Sonnet 5, which used 9.55. In October we tested GPT-6 Luna, which used 5.39. That is more than GPT-5.6 Luna in every run, but not by enough to confirm, and GPT-6.1 Sol used 3.56.
GPT's main tell is the list of three. GPT-5.6 wrote about 16 of them per 1,000 words, or five in a 300-word draft, about twice Claude's rate. Turning up reasoning effort changed neither count. GPT-6 wrote about a third fewer lists, near 11 per 1,000 words.
GPT-5.6 still uses em dashes, at half Claude's rate
We asked each model for 35 short pieces in seven kinds of writing, from emails to LinkedIn posts. Each ran three times, so each row is 105 drafts of about 300 words. We called the models directly through OpenAI's API, without the ChatGPT app.
Em dashes per 1,000 words, GPT-5.6 and GPT-6 against two other models, September and October 2026
| Model | Dashes per 1,000 words | 95% range | Drafts with any dash | Drafts with over two |
|---|---|---|---|---|
| Claude Sonnet 5 | 9.55 | 8.56–10.65 | 92.4% | 59.0% |
| GPT-6 Luna (October) | 5.39 | 4.61–6.29 | 71.4% | 24.8% |
| GPT-5.6 Luna | 4.43 | 3.74–5.24 | 61.0% | 19.0% |
| GPT-5.6 Terra | 4.35 | 3.67–5.14 | 62.9% | 16.2% |
| GPT-6.1 Sol (October) | 3.56 | 2.96–4.28 | 65.7% | 8.6% |
| Gemini 3.8 Flash | 3.56 | 2.94–4.30 | 63.5% | 8.7% |
Each row is one model, about 105 drafts of 300 words. The 95% range shows where the true rate likely sits. The last two columns give the share of drafts with at least one em dash and with more than two.
The 95% range is where the true rate likely sits. Across three runs, Terra ranged from 3.99 to 4.64 and Luna from 3.83 to 5.11. Both sit well below Sonnet 5's runs, 9.14 to 10.03, and overlap Gemini's. GPT-6 Luna ranged from 5.05 to 5.56 and GPT-6.1 Sol from 3.24 to 3.81, on 15 October 2026. GPT-5.6 Luna ran again that day as a check, at 4.33. More than two in one draft is where our AI slop detector flags them.
At low effort, no GPT draft used more than six, while Sonnet 5 went up to eight. Neither GPT model swapped the em dash for a spaced en dash ( – ) at low effort, a swap that would hide it from a simple count.
November: GPT-6.2 Astra and Sol, counted
On 20 November 2026 we gave the same 35 requests, three times each, to GPT-6.2 Astra and GPT-6.2 Sol, through OpenAI's API. GPT-6 Luna ran again as a check and matched its October count.
Em dashes per 1,000 words, GPT models, 20 November 2026
| Model | Group | Dashes per 1,000 words | 95% range | Lowest and highest run |
|---|---|---|---|---|
| GPT-6.2 Astra | Frontier | 0.53 | 0.33–0.86 | 0.30–0.70 |
| GPT-6.2 Sol | Middle | 1.35 | 1.00–1.82 | 1.34–1.35 |
| GPT-6 Luna (check) | Fast | 5.53 | 4.75–6.45 | 4.71–6.02 |
Each row is one model, 105 drafts in three runs. The 95% range shows where the true rate likely sits. The three models sit in three different size groups.
GPT-6.2 Astra wrote 0.53 em dashes per 1,000 words and GPT-6.2 Sol 1.35. Too many lists of three showed up in 21.0% of Astra drafts and 37.1% of Sol drafts. We compare each model only with models of its own size.
Our best AI model for writing test counts all 23 tells for these models, against Gemini and Claude models of the same size.
Its real tell is the list of three
GPT-5.6 reaches for three items in a row far more than any other model we tested. "Crisp, cold, and ready." "Your bag, desk, or car cup holder." It wrote about 16 of these lists per 1,000 words. Claude wrote 7 to 8, and Gemini 10 to 11.
Lists of three per 1,000 words, and drafts with too many
| Model | Lists per 1,000 words | 95% range | Drafts with too many |
|---|---|---|---|
| GPT-5.6 Terra | 16.47 | 15.11–17.96 | 84.8% |
| GPT-5.6 Luna | 16.07 | 14.72–17.54 | 87.6% |
| GPT-6.1 Sol (October) | 11.71 | 10.58–12.96 | 73.3% |
| GPT-6 Luna (October) | 11.26 | 10.11–12.53 | 73.3% |
| Gemini 3.7 Flash | 11.21 | 10.07–12.47 | 71.4% |
| Claude Sonnet 5 | 8.20 | 7.29–9.23 | 61.0% |
Each row is one model, about 105 drafts. The 95% range shows where the true rate likely sits. Too many means more than one list of three per 300 words, where our detector flags a draft.
A 300-word draft is allowed one list of three before our detector flags it. Across both models, 181 of 210 GPT drafts went over, against 119 of 210 for Claude. A gap that size would turn up by chance less than once in a billion tries. GPT-6 cut the habit by about a third. On 15 October 2026, GPT-6 Luna wrote 5.08 fewer lists per 1,000 words than GPT-5.6 Luna on the same prompts (95% interval 2.61 to 7.48 fewer), and GPT-6.1 Sol 4.62 fewer (2.01 to 7.18). That puts GPT-6 level with Gemini Flash, and seven in ten of its drafts still went over.
Stay cool, hydrated, and prepared with a water bottle that works as hard as you do.
GPT-5.6 Terra, the last line of a product description, first run
One list of three reads fine, but five in 300 words reads like a brochure. It is the tell to look for first in a GPT draft.
It uses the fewest stock AI words
Words like "seamless", "robust", "leverage" and "landscape" are the tell most people know, and GPT-5.6 used them least. Only 6 of 105 drafts from each model had one, or 5.7%. GPT-6 used them rarely too: 2 of 105 GPT-6 Luna drafts and 5 of 105 GPT-6.1 Sol drafts had one.
Claude drafts had one 17.1% of the time and Gemini drafts 23.9%. So if a draft is full of "seamless" and "landscape", GPT-5.6 is the least likely of the three to have written it.
No effort level changed the count
GPT-5.6 Luna lets you set how long it thinks before it writes, from minimal to xhigh. We ran all five levels on the same 35 jobs, three times each.
GPT-5.6 Luna at five effort levels, 105 drafts per level
| Effort | Dashes per 1,000 words | Tells per 1,000 words | Thinking tokens per draft | Cost per 100 drafts |
|---|---|---|---|---|
| Minimal | 3.94 | 4.07 | 26 | $0.05 |
| Low | 4.43 | 4.33 | 25 | $0.05 |
| Medium | 3.97 | 4.33 | 35 | $0.05 |
| High | 3.41 | 3.77 | 79 | $0.06 |
| Xhigh | 3.99 | 4.06 | 245 | $0.10 |
Each row is one effort level, 105 drafts. Tells counts all 23 patterns our detector checks, em dashes among them. Thinking tokens are the word pieces the model spends before it writes, as a median per draft.
Every level's gap from minimal is small enough to be chance, for em dashes and for all tells. Lists of three held at 16 per 1,000 words at every level.
Xhigh thought about ten times as long as minimal and cost twice as much, and wrote the same way.
Terra costs ten times more and writes the same
GPT-5.6 Terra lists at ten times Luna's price per token. On these jobs it cost $0.46 per 100 drafts against Luna's $0.05. Its tells matched Luna's on every measure: 4.22 against 4.33 per 1,000 words, 4.35 against 4.43 em dashes.
If you pick Terra for better prose, this test gives you no reason to. On short drafts, the cheaper model leaves the same tells.
Emails had almost none, LinkedIn posts the most
The kind of writing moved GPT's em dash count more than anything else. In emails, GPT-5.6 used 0.77 per 1,000 words. In LinkedIn posts it used 8.07, and in product copy 7.97, about ten times as many.
GPT-5.6 em dashes per 1,000 words for each type of draft, Terra and Luna together
| Kind of writing | Dashes per 1,000 words |
|---|---|
| LinkedIn post | 8.07 |
| Product or marketing copy | 7.97 |
| Opinion piece | 4.95 |
| Blog post | 4.74 |
| How-to guide | 2.53 |
| Personal essay | 2.18 |
| 0.77 |
Each row is one type of draft: five requests, three runs and two models, so 30 drafts.
So an email from ChatGPT will rarely give itself away with a dash. A LinkedIn post will, about two or three times in 300 words.
How to get fewer em dashes from ChatGPT
Tell it not to use them. In November 2025, Tom's Guide reported that OpenAI's Sam Altman said ChatGPT now follows a custom instruction not to use them. Users still see the odd one slip through. We gave no instruction, so our counts are GPT's default.
The instruction only covers the dash, so the lists of three stay. Our free AI slop detector flags all 23 tells we counted here and quotes each line, and the AI humanizer rewrites them.
None of this touches the facts. A GPT draft with no dashes and no lists can still quote a number nobody published. That is the part TrueStandard checks: paste your draft, and four models from different labs test each claim in about 60 seconds.
How we ran it
We wrote 35 plain requests, five each for seven kinds of writing. One was "Write a LinkedIn post about being laid off and looking for my next role in data analysis." Each asked for about 300 words and the text only. We set no tone, no style and no system prompt.
Each model got every request three times, on 26 September 2026, through OpenAI's API with low reasoning effort. Luna also ran at minimal, medium, high and xhigh. We gave every draft room for 32,000 tokens, or word pieces, so none was cut short. A script scored all 630 GPT drafts against the 23 patterns in our AI slop detector. No AI model graded anything. On 15 October 2026 we ran GPT-6 Luna and GPT-6.1 Sol the same way, 210 more drafts, with GPT-5.6 Luna re-run as a check that nothing had drifted.
These are short first drafts from a bare request. ChatGPT adds its own instructions and yours on top of the model, so its output can differ from ours. A flag is a pattern match, and it is not a verdict. We left out GPT-6 Astra for cost, and GPT-6.2 Astra ran on 20 November 2026.
We ran the same test on Claude and Gemini. The ten-model comparison has all of them side by side.
The other two vendors:
- Claude Sonnet 5 and Haiku 4.5 used the most em dashes of any model we tested.
- Gemini 3.7 and 3.8 Flash used the fewest em dashes in September and the most stock AI words.
FAQ
Does ChatGPT still use em dashes?
Yes. With no instructions, GPT-5.6 Terra and Luna used about 4.4 em dashes per 1,000 words in our September 2026 test, and 62% of their drafts had at least one. That is less than half of Claude Sonnet 5's rate of 9.55. In October, GPT-6 Luna used 5.39 and GPT-6.1 Sol 3.56. On 20 November, GPT-6.2 Astra used 0.53 and GPT-6.2 Sol 1.35.
How do I stop ChatGPT from using em dashes?
Add "don't use em dashes" to ChatGPT's custom instructions under Personalization. OpenAI says ChatGPT now follows it, though some users still see one slip through. It will not remove GPT's other habits, like lists of three.
Who uses more em dashes, ChatGPT or Claude?
In September, Claude. On the same 35 writing jobs, Claude Sonnet 5 used 9.55 em dashes per 1,000 words and Claude Haiku 4.5 used 6.04. GPT-5.6 Terra and Luna used about 4.4. The newer Claude models used 0.30 or fewer on 20 November 2026, run a different way, so we do not rank them against GPT.
What is the biggest sign that ChatGPT wrote something?
Lists of three, in our test. GPT-5.6 wrote about 16 per 1,000 words, about five in a 300-word draft, twice Claude's rate. 86% of GPT-5.6 drafts had more than our detector allows. GPT-6 wrote about 11, a third fewer. It used stock words like "seamless" the least of any vendor.
Does GPT use fewer em dashes at higher reasoning effort?
No. GPT-5.6 Luna used between 3.4 and 4.4 em dashes per 1,000 words at all five effort levels, minimal to xhigh. The gaps were small enough to be chance, though xhigh thought about ten times as long.
Is GPT-5.6 the model inside ChatGPT?
ChatGPT runs OpenAI's GPT models with its own instructions, and your settings, on top. We called GPT-5.6 Terra and Luna directly with no instructions, so ChatGPT's output can differ from what we measured.
Keep reading
Multi-Agent vs Multi-Model AI in 2026
Multi-agent AI runs many roles of one model. Multi-model AI runs different models from different labs. Only multi-model can catch an error the first model is blind to.
Does Gemini Use Em Dashes? Fewer, but More AI Words
We gave Gemini 3.7 Flash and Gemini 3.8 Flash the same 35 writing jobs as Claude and GPT, three times each. Then we counted every em dash and 22 other AI writing tells. In October we added Gemini 4 Argon, and in November three more Gemini 4 models.
Best AI Model for Writing: 11 Models Tested
We counted 23 AI writing tells in 1,155 drafts. Two pairs split, and no model won overall.
Long Context vs RAG in 2026
Under about 100 pages, put your documents in a long context window. From 100 to 500 pages, Karpathy's wiki method works, and past 500, graph RAG wins. Better retrieval still does not make the answer true.
What Is Context Rot?
An agent gets less reliable as notes, history and tools pile up. Here are four ways it fails.
Check the claims in your GPT draft
A clean GPT draft can still cite a study that does not exist. Paste yours into TrueStandard. Four models from different labs check each claim and flag what none of them can back up, in about 60 seconds.
TrueStandard is for writers who publish AI-assisted drafts. Paid plans check every claim and citation across four AI models before you hit publish.
Free: one check a day, no signup. Pro is $20 a month for 250 drafts.
Check Your Draft →