The AI writing tells of 2026: what changed, in 7,554 rewrites and 1,177 drafts
We had 2026 models polish thousands of passages people wrote, and write over a thousand pieces from scratch, then counted what the models added. The famous tells are gone. Quieter ones replaced them.
"Delve" has largely gone out of style in 2026. If the AI is writing from scratch, the biggest red flag now is formatting. AI will add headers, bold certain parts and create bullet lists, even in something as simple as an email. The — character shows up six times as often as in human writing; people use nearly as many dashes, but tend to type them as "--" or " - ". If the AI is polishing your writing, it will add small words like "for instance", "genuinely" and "roughly", and even out the length of your sentences.
What did we measure?
We ran two studies. The first measures what AI does when it rewrites human text. The second measures what it does when it writes from scratch.
Study 1, rewriting. Our dataset consists of 7,554 pairs. Each pair contains a chunk of text from a blog written before ChatGPT launched (November 2022) and the same chunk of text rewritten by an LLM in 2026. The human-written text comes from four well-known blogs. In total, the human-written texts contain about 1.2 million words, and the AI-rewritten texts contain about 1.2 million words. We rewrote the human-written texts in four settings (a system prompt for helping polish a blog post, a system prompt for acting as an expert editor for a blog post, the system prompt "You are a helpful assistant", and no system prompt) using ten models: Claude Opus 5.5, OpenAI's GPT-5.5, GPT-6.1 and chat-latest, DeepSeek V4.1 Flash, GLM-5.3, gpt-oss-120b, Kimi K3, MiniMax M3 and Qwen 3.8 Max. For sentence rhythm, we measured a held-out set of 471 pairs.
Study 2, writing from scratch. First, we gathered 53,616 human-written documents written before November 2022. These documents had a combined length of 22.2 million words, and represent 20 different kinds of writing (emails, forum posts, LinkedIn posts, social posts, news, press releases, speeches, letters, short stories, reviews, blog posts, newsletters, how-tos, recipes, docs, product pages, marketing copy, event pages and academic abstracts, among others). Then for 692 of these documents (about 35 of each kind), we took notes on the facts discussed. Finally, we prompted four 2026 models (Claude Sonnet 5.5, GPT-5.4 mini, Gemini 3.8 Flash and GLM-5.3) to write a document of the same kind as the original, in the same tone and point of view, at the same length, containing the facts from our notes. This resulted in 1,177 drafts totaling about 500,000 words, compared to the 232,000 words of the original documents. We didn't say anything about formatting in the prompt. The drafts came out at a similar length to the originals (a median of 1.03 times).
In both studies, for each word, phrase and punctuation mark, we computed how often it occurs per 1,000 words in the human text and in the AI text.
Does AI add formatting when it writes from scratch?
Yes, and by far the biggest difference. Our AI-generated drafts used 10.1 bold phrases per 1,000 words and 4.2 markdown headers per 1,000 words. In other words, there was a bold phrase about every 100 words. They also used bullet points much more often: 8.9 bullet-point lines per 1,000 words compared with just 2.4 in the original, human-written documents. Last, there were 2.5 times as many colons in the AI-generated documents (9.6 per 1,000 words) compared with the originals (3.8). And remember, none of the prompts asked for any of this.
With the exception of fiction, this was true for all the kinds of writing we tried. For example, AI-generated email drafts used 5.8 bold phrases per 1,000 words. Docs, how-tos and event pages used even more (20.6, 19.1 and 22.1, respectively). But the AI-generated short stories had almost no bold at all. The AI-generated versions also used emoji much more, eight times as much in fact (0.92 per 1,000 words vs. 0.11).
| Formatting, per 1,000 words | Human originals | AI drafts |
|---|---|---|
| Bold phrases | see note | 10.08 |
| Markdown headers | see note | 4.21 |
| Bullet-point lines | 2.44 | 8.86 |
| Colons | 3.82 | 9.62 |
| Emoji | 0.11 | 0.92 |
What words do AI models use too much in 2026?
The words that are most often overused by AI are small, ordinary words, and they depend on the task. When rewriting, the most lopsided were "for instance" (used twice in the original 1.2 million words of human-written text, but 298 times in the AI rewrites), "genuinely" (12.5 times more common in the rewrites than in the originals), "roughly", "plenty of" and "essentially" (each about 7 times as common), and "truly", "honestly" and "simply" (each about 5 times as common).
When writing from scratch, the list changes. "Honestly" (8.3 times the human rate), "whether you're" (7.1), "roughly" (4.7), "rather than" (4.3), "genuinely" (4.2) and "here's" (3.9) lead it. "For instance" and "essentially" drop out: the drafts used them less than people did.
| Word or phrase | Rewriting: AI vs. human | From scratch: AI vs. human |
|---|---|---|
| for instance | ~150x (2 uses vs. 298) | 0.3x |
| genuinely / genuine | 12.5x | 4.2x |
| roughly | 7.7x | 4.7x |
| plenty of | 7.2x | 1.5x |
| essentially | 7.1x | 0.4x |
| truly | 5.4x | 1.5x |
| honestly | 5.3x | 8.3x |
| simply | 4.7x | 2.2x |
| rather than | 3.7x | 4.3x |
| here's | 3.9x | |
| whether you're | 7.1x | |
| actually | 2.0x | 1.7x |
| crucial | 3.3x | 0.7x |
| leverage | no difference | almost never used |
| delve | 0 uses in the rewrites | almost never used |
Most of these words are intensifiers or hedges. The models add them to sound careful and sincere, and the result sounds careful and sincere in exactly the same way every time. "Honestly", "genuinely", "roughly" and "rather than" show up in both studies, so they're the safest bets. For a longer list of words and phrases that AI tends to overuse, and suggestions for how to replace them, see words and phrases AI overuses.
Do AI models still use words like "delve"?
The answer to that is no. The word "delve" appears one time in the human-written passages in our rewriting test, and zero times in the AI rewrites. In the from-scratch drafts it is close to zero too. The words "tapestry" and "testament" show the same pattern, "leverage" barely appears, and "moreover", "furthermore" and "additionally" appear at the same rate as in human writing. If your list of tell-tale AI words starts with "delve", you need to update it, because it's still stuck in 2023. You can find the current list in signs of AI writing: the 2026 list.
Do AI models still use too many em dashes?
When writing from scratch, yes. The AI drafts used the em dash (—) 4.31 times per thousand words, compared to 0.72 for the original human writing. That's 6 times as often. The human figure is in line with the average of 0.73 for all 22 million words of human writing we have. There's also a big difference between models: for Claude Sonnet 5.5 it's 0.09, for GPT-5.4 mini it's 1.99, for Gemini 3.8 Flash it's 3.90, and for GLM-5.3 it's 11.13.
The caveat is that people use dashes almost as much as AI. They just use a different kind of dash. If we count all kinds of dashes (em dash —, en dash –, double hyphen -- and hyphen with spaces - ), then the average is 2.97 for the original human writing and 4.67 for the AI drafts. Most of the time people use double hyphens or hyphens with spaces, which AI almost never uses. For example, in 4,000 emails written by people, there are 3.8 dashes per thousand words, but only 0.06 of them are em dashes. In the emails AI drafted, there are 2.9 em dashes per thousand words. So the tell is the character, not the pause.
| Source | Em dashes (—) per 1,000 words |
|---|---|
| From scratch | |
| Claude Sonnet 5.5 | 0.09 |
| Human originals (20 kinds of writing) | 0.72 |
| GPT-5.4 mini | 1.99 |
| Gemini 3.8 Flash | 3.90 |
| GLM-5.3 | 11.13 |
| Rewriting | |
| Claude Opus 5.5 | 0.10 |
| Human originals (four blogs) | 2.56 |
| Qwen 3.8 Max | 3.66 |
| OpenAI (GPT-5.5, GPT-6.1, chat-latest) | 3.88 |
| DeepSeek V4.1 Flash | 6.02 |
| Kimi K3 | 6.84 |
| MiniMax M3 | 7.75 |
| gpt-oss-120b | 10.31 |
| GLM-5.3 | 10.75 |
The numbers for rewriting look smaller because the four bloggers we rewrote use a lot of em dashes: 2.56 per thousand words, which is 3.5 times the average across all kinds of writing. Even so, across all ten models, the AI rewrites use 4.14 em dashes per thousand words, which is 1.6 times the bloggers' rate. Claude Opus 5.5 uses very few (0.10), but GLM-5.3 and gpt-oss-120b both use over ten.
What's normal also depends on the type of writing. Across our baseline of 22 million words, people use em dashes very rarely in emails (0.06), forum posts (0.11) and short stories (0.28), but they're common in blog posts (1.27) and more common still in news stories and speeches (1.62). It's unusual for a person to use an em dash in an email, but it's very normal in a news story. For more, see why ChatGPT uses so many em dashes.
Does AI writing have a different rhythm?
Yes, and it's one thing models can't fix well. It's easy to change vocabulary, but harder to match how varied your sentence lengths are. In our rewriting study's held-out set, the coefficient of variation (standard deviation divided by the mean) of sentence length is 0.65 in the human writing, but only 0.59 in the rewrites by OpenAI's models, and 0.56 in the rewrites by Claude. Also, where the bloggers start about 12% of sentences with "And", "But" or "So", OpenAI's rewrites do it 9.7% of the time and Claude's only 5%.
There's not a very big difference when the models write from scratch: we get 0.50 for the original human-written texts, and 0.47 for the AI drafts. Among the models, GPT-5.4 mini is the most uniform (0.43), followed by Gemini 3.8 Flash (0.46). Claude Sonnet 5.5 (0.49) and GLM-5.3 (0.51) are close to the human figure. Starting sentences with "And", "But" or "So" turned out to be partly a blogger habit: across the 20 kinds of writing, people did it 3.8% of the time and the drafts 3.4%. This study is based on different texts than the rewriting study, so only compare numbers within the same study.
This is probably a big part of why AI writing has such a distinctive sameness, one that readers notice without being able to name it: sentences of similar length, joined in similar ways, with the rough edges smoothed off. To see where that comes from, read why your AI can't fix its own writing.
Do AI models use "it's not X, it's Y"?
Not that much, overall. When they rewrite an existing text, the models use this pattern slightly less than the human authors (0.12 vs. 0.17 per 1,000 words). When they write from scratch, the number is the same as in the original texts (0.12).
It varies from one kind of writing to another and from one model to another. The human LinkedIn posts in our sample never contain these phrases, while the AI versions do (0.51 per 1,000 words). Other kinds of writing that see an increase are academic abstracts (from 0 to 0.32), letters (from 0 to 0.18) and forum posts (from 0 to 0.13). Among the models, GPT-5.4 mini stands out with 0.29 per 1,000 words, about two and a half times the human rate. On the other end, we have Claude Sonnet 5.5 (0.04), which almost never uses these phrases. We only have a limited number of texts of each kind (about 35 originals), so take this with a grain of salt. See why AI loves "it's not X, it's Y".
What else is different in AI drafts?
Writing in threes. Groups of three such as "fast, simple and secure" were 1.4 times as common in the drafts (5.60 vs. 4.00 per 1,000 words).
Fewer exclamation marks. AI writing comes across as calmer, with an average of 1.77 exclamation marks per 1,000 words vs. 3.01 for people. In some kinds of writing the drop is bigger: LinkedIn posts went from 8.3 to 3.4, and newsletters from 10.0 to 5.4.
Contractions are not a giveaway. AI uses slightly more contractions than people (15.0 vs. 12.7 per 1,000 words). How much people use them depends on the kind of writing, from 4.4 in academic abstracts to 21 in forum posts.
What are the limits of this data?
For the rewriting study, the human passages were taken from four bloggers whose writing style may not reflect everyone's (for one, they use far more em dashes than most people). For the from-scratch study, we fed notes to the AI models, which may bias the results towards bullet points. We used four models, and for each kind of writing we had about 35 originals. The kinds of writing are our own categories. For both studies, the counts come from simple pattern matching that doesn't consider the context of the piece of writing. So the ratios here should be taken as directions, not as exact constants. This page will be updated with new models as they become available.
For the full picture of why AI text sounds like AI and what to do about it, start with how to make AI writing sound human.