Which AI sounds least like AI? The best writers are the heaviest em-dash users

Readers say a text “sounds like AI” long before they can say why. When pressed, they name the same things: the em dash in every paragraph, “delve” and “tapestry”, the sentence that says what something is not before saying what it is, the list that always has three items. None of these is proof of anything. All of them are countable. So we counted them, in every English text that entered the ranking, and turned the count into one number per model: marks per 1,000 words. This article explains what that number shows. The figures and ranks in the text date from 20 September 2026; the charts are recomputed at each update and may differ.

The two best writers are also the two heaviest em-dash users

How much the 15 best writers sound like AI

Marks of machine prose per 1,000 words in the ranked texts: em dashes, stock phrases, “not X but Y”, lists of three. Lower is better; #n is the rank on the writing leaderboard.

View as a data table

	Marks per 1,000 words	Rank	Texts
Claude Opus 5.5 💡 adaptive high	2.1	2	33
Claude Fable 5.1 💡 adaptive high	2.4	6	33
Claude Sonnet 5.5 💡 adaptive high	2.4	7	34
GPT-6 Sol 💡 high	2.8	7	32
GPT-6 Luna 💡 high	7.4	19	36
Grok 4.7 💡 high	7.4	12	36
Muse Spark 1.3 💡 high	8.3	19	41
MiniMax M3 💡 enabled	8.4	18	56
GPT-5.6 Sol 💡 high	8.7	16	135
GLM 5.2 💡 enabled	8.8	23	57
Claude Opus 5 💡 adaptive high	9.1	10	44
MiMo V2.6 Pro fp8 💡 high	11.6	17	41
DeepSeek V4.1 Flash 💡 high	12.0	21	31
Kimi K3 💡 enabled	14.4	3	46
Claude Fable 5 💡 adaptive high	14.6	5	199

Among the fifteen best-ranked models on 20 September, the two at the top had the most marks. Kimi K3, then first on the leaderboard, scores 14.3 marks per 1,000 words; Claude Fable 5, then third, scores 14.6. Almost all of it is the em dash: 11.8 and 12.2 per 1,000 words, one every 80 words. Claude Opus 5, then second, is noticeably more restrained at 9.1, with a dash every 125 words. GPT-5.6 Sol, then ninth, had the lowest index of the top ten at 8.7, with a dash every 230 words.

The models with the fewest marks are not the best writers. Qwen3.7 Plus had the lowest index of the whole ranking, 4.5, and ranked 42nd; Qwen3.7 Max, at 5.6, ranked 31st. The median model had 8.9 marks per 1,000 words.

The em dash is the whole story; stock phrases barely register

What the marks are made of, for the ten best writers

The same index split by family, per 1,000 words.

• Em dashes
• Stock phrases
• “Not X, but Y”
• Lists of three

View as a data table

	Em dashes	Stock phrases	“Not X, but Y”	Lists of three	Marks per 1,000 words
Claude Opus 5.5 💡 adaptive high	0.2	0.0	0.3	1.6	2.1
Kimi K3 💡 enabled	11.9	0.0	0.5	2.0	14.4
Claude Fable 5 💡 adaptive high	12.2	0.1	0.6	1.7	14.6
Claude Fable 5.1 💡 adaptive high	0.6	0.1	0.6	1.1	2.4
Claude Sonnet 5.5 💡 adaptive high	0.4	0.0	0.4	1.6	2.4
GPT-6 Sol 💡 high	1.5	0.1	0.1	1.1	2.8
Claude Opus 5 💡 adaptive high	7.9	0.0	0.2	0.9	9.1
Grok 4.7 💡 high	4.5	0.0	0.8	2.1	7.4
GPT-5.6 Sol 💡 high	4.3	0.1	1.0	3.3	8.7
MiMo V2.6 Pro fp8 💡 high	9.1	0.1	0.5	1.9	11.6

Split the index into its parts and one part dominates. In the whole English corpus, 1.2 million words, em dashes account for 6.7 marks per 1,000 words, the “not X but Y” construction for 1.5, lists of three for 1.8, and stock phrases for 0.5. The word that readers cite most, “delve”, appears so rarely that it does not make the top ten; “vibrant” leads with 104 occurrences, then “additionally,” (99), “furthermore,” (95) and “crucial” (61). One stock phrase every 2,000 words is not what the memes suggest.

The one model that bucks the pattern is GPT-5.6 Sol: few dashes, but the highest rate of lists of three among the top models, 3.3 per 1,000 words, and a taste for the “not just X” construction. Its tells are grammatical rather than typographic, which makes them harder to edit out.

Fewer tells does not mean better writing

Writing score against machine-prose marks

One dot per model with at least 30 ranked texts; local models in the second colour. Correlation between the two: -0.42.

View as a data table

	Marks per 1,000 words	Writing score
Claude Opus 5.5 💡 adaptive high	2.1	94.8
Kimi K3 💡 enabled	14.4	92.2
Claude Fable 5 💡 adaptive high	14.6	89.7
Claude Fable 5.1 💡 adaptive high	2.4	89.4
Claude Sonnet 5.5 💡 adaptive high	2.4	89.1
GPT-6 Sol 💡 high	2.8	89.1
Claude Opus 5 💡 adaptive high	9.1	88.1
Grok 4.7 💡 high	7.4	87.6
GPT-5.6 Sol 💡 high	8.7	80.5
MiMo V2.6 Pro fp8 💡 high	11.6	79.7
MiniMax M3 💡 enabled	8.4	79.5
GPT-6 Luna 💡 high	7.4	79.4
Muse Spark 1.3 💡 high	8.3	79.4
DeepSeek V4.1 Flash 💡 high	12.0	78.0
GLM 5.2 💡 enabled	8.8	77.6
Hy3 💡 high	7.9	77.3
Gemini 3.8 Flash 💡 high	7.7	76.9
Muse Glimmer 30B K-Quant Dynamic 💡 high 🏠	7.3	72.4
GPT-5.6 Terra 💡 high	8.9	65.9
Muse Glimmer 30B K-Quant 17GB 💡 high 🏠	9.6	61.4
GPT-5.6 Luna 💡 high	8.2	60.0
DeepSeek V4 Pro 💡 enabled	11.8	59.3
Qwen3.8 27B Q4_K_M 💡 xhigh** 🏠	9.0	58.7
Qwen3.6 27B Q8_0 💡 on* 🏠	7.5	58.6
Qwen3.7 Max 💡 enabled	5.6	57.7
Qwen3.8 27B Q8_0 💡 xhigh** 🏠	10.8	57.5
Qwen3.8 27B Q6_K 💡 xhigh** 🏠	10.1	56.5
DeepSeek V4 Flash 💡 enabled	13.5	55.2
Gemma 4 31B Q6_K 💡 on* 🏠	7.6	53.9
Gemma 4 31B Q8_0 💡 on* 🏠	8.2	53.5
Inkling 💡 high	17.1	52.2
Qwen3.6 27B Q4_K_M 💡 on* 🏠	6.9	51.6
Gemma 4 31B QAT-Q4_0 💡 on* 🏠	6.0	51.5
Gemma 4 26B A4B Q6_K 💡 on* 🏠	6.7	50.9
Qwen3.7 Plus 💡 enabled	4.4	50.2
Qwen3.6 35B A3B Q8_0 💡 on* 🏠	7.2	47.5
Gemma 4 31B QAT-Q4_0 ↯ 🏠	8.5	47.0
MiMo V2.5 Pro 💡 enabled	11.2	45.3
Gemma 4 31B Q8_0 ↯ 🏠	7.7	43.3
Gemma 4 31B Q6_K ↯ 🏠	7.3	42.4
Gemma 4 12B Q6_K 💡 on* 🏠	7.7	42.2
MiMo V2.5 💡 enabled	9.4	42.1
Qwen3.6 35B A3B Q4_K_M 💡 on* 🏠	6.7	41.8
Gemma 4 12B QAT-Q4_0 💡 on* 🏠	7.6	41.7
Gemma 4 12B Q8_0 💡 on* 🏠	9.9	41.3
Gemma 4 26B A4B Q8_0 💡 on* 🏠	9.1	41.3
Gemma 4 26B A4B Q6_K ↯ 🏠	7.2	41.1
Laguna S 2.1 UD Q4_K_M 💡 on* 🏠	15.9	40.1
Gemma 4 26B A4B QAT-Q4_0 ↯ 🏠	6.8	38.2
Gemma 4 26B A4B QAT-Q4_0 💡 on* 🏠	8.7	38.1
Ling 3.0 Flash Q4_K_M ↯ 🏠	11.0	36.2
Gemma 4 12B Q8_0 ↯ 🏠	9.0	34.9
Seed 2.1 Turbo ↯	8.8	34.2
Gemma 4 26B A4B Q8_0 ↯ 🏠	7.6	32.0
Agnes 3.0 Flash 💡 high	8.7	29.9
Gemma 4 12B QAT-Q4_0 ↯ 🏠	7.1	29.5
Gemma 4 12B Q6_K ↯ 🏠	8.7	29.0
Nemotron 3 Super 120B A12B Q4_K_M 💡 full 🏠	10.3	28.3
Qwen3.6 27B Q8_0 ↯ 🏠	8.7	26.5
Qwen3.8 27B Q6_K ↯ 🏠	7.8	26.4
Ling 3.0 Flash Q4_K_M 💡 high 🏠	11.5	25.8
Granite 4.2 30B Q8_0 💡 full 🏠	14.2	25.7
Mistral Medium 3.5 128B Q4_K_M ↯ 🏠	17.2	24.4
Mistral Medium 3.5 💡 high	10.1	22.4
Qwen3.6 35B A3B Q8_0 ↯ 🏠	8.0	21.2
Qwen3.8 27B Q4_K_M ↯ 🏠	7.7	20.6
Qwen3.6 35B A3B Q4_K_M ↯ 🏠	7.2	19.8
Granite 4.2 30B Q6_K 💡 full 🏠	14.2	19.5
Qwen3.8 27B Q8_0 ↯ 🏠	7.0	19.3
Granite 4.2 8B Q8_0 💡 full 🏠	17.6	18.5
Nemotron 3 Nano Omni 30B A3B Q8_0 💡 on* 🏠	11.6	18.4
Granite 4.2 8B Q4_K_M 💡 full 🏠	17.5	17.9
Gemma 4 E2B Q8_0 💡 on* 🏠	12.6	16.6
Nemotron 3 Nano Omni 30B A3B Q4_K_M 💡 on* 🏠	14.2	14.7
Gemma 4 E4B Q6_K 💡 on* 🏠	15.7	14.6
Gemma 4 E4B Q8_0 💡 on* 🏠	14.7	13.4
Gemma 4 E4B Q8_0 ↯ 🏠	12.2	11.8
Granite 4.2 30B Q4_K_M 💡 full 🏠	14.5	11.8
Gemma 4 E2B Q6_K 💡 on* 🏠	12.1	10.9
Gemma 4 E4B Q4_K_M 💡 on* 🏠	14.6	10.7
Gemma 4 E2B Q4_K_M 💡 on* 🏠	10.4	10.4
Gemma 4 E2B Q8_0 ↯ 🏠	9.5	10.4
Gemma 4 E2B Q4_K_M ↯ 🏠	9.9	7.8

Across the 36 models with at least 30 ranked texts, the correlation between the index and the writing score is almost zero. The scatter shows why: the best writers are spread from 8.7 to 14.6, and the models with the most marks include both the leaders and some of the weakest files, such as Mistral Medium 3.5 128B at 17.2 marks per 1,000 words and rank 71 at the time. The judges, who never see the model names, reward the prose and ignore the punctuation, or at least do not punish it.

What the index can and cannot tell you

The index counts surface marks, nothing more. It is not an AI detector, and a human text full of em dashes would score just as high. It also depends on the prompt: a request for a list gets a list. That is why markdown structure and emoji are reported separately in the data and not added to the index.

What it does tell you is practical: which model’s first draft needs the least work before it stops reading as machine prose. For English, that is not the best writer. It is either a mid-table model with a low index, or the best writer followed by a find-and-replace on the em dash, which costs a minute and removes most of the signal. The charts above give both numbers for each model, so you can weigh them for the kind of writing you do.

The same count in French, German and Spanish

The four editions do not sound like AI in the same way. German texts carry the most marks, 11.7 per 1,000 words, again mostly dashes; English 10.5; French 5.8; Spanish 4.9, where the em dash is rare and the tells are “además,” and lists of three. Each edition has its own version of this article with its own lists of phrases, because a French reader flinches at “véritable” and a German reader at “zudem”, not at “delve”.

Conclusion

Sounding like AI and writing well are two different things. The em dash is the one mark that separates models, and the best writers use it most: Kimi K3 and Claude Fable 5 put one every 80 words, Claude Opus 5 one every 125, GPT-5.6 Sol one every 230. Stock phrases are rarer than the internet claims, one every 2,000 words. If a draft must not read as machine prose, pick a model with a low index and a high score, or pick the best writer and cut the dashes yourself: it takes a minute.