Signs of AI writing: the 2026 list

The main signs of AI writing in 2026 are bold, headers and bullets nobody asked for, overused words (genuinely, roughly, honestly, simply), sentences of about the same length, "rather than" corrections, hedging, generic examples, and tidy summaries and upbeat closers. A few together probably mean AI; one alone isn't a great signal. Delve is out! Em dashes are mixed: most models use more than people do, but Claude uses far fewer.

This page will update quarterly with new models. We use two studies of our own: thousands of human-written passages rewritten by 2026 AI models, and 1,177 drafts that 2026 models wrote from scratch, compared with human writing of the same kind. We also use the two best publicly available lists of AI tells: Wikipedia's Signs of AI writing page and Pangram's data on AI tells.

What changed in AI writing in 2026?

Three things have changed: the words that were common in 2023 are no longer that common, em dashes are no longer reliable, and new words and framings have cropped up.

Gone: Delve (and others)

Delve is gone. So are many other 2023-isms like tapestry, testament, moreover, etc. In our data, delve is close to zero for both human originals and AI rewrites. Wikipedia editors describe the same arc: delve was "famously overused by ChatGPT in 2023 and early 2024," became less frequent later in 2024, then dropped off sharply in 2025. Tapestry and testament are also near zero for both. Moreover, furthermore, and additionally are also rarely used by either humans or AI (0.01 per thousand words for humans, and 0.02 per thousand words for AI rewrites), and AI drafts used them at the same rate as people. If you're still scanning for these, you're looking for 2023's AI.

Em dashes are less reliable

AI writing (both rewrites and drafts) uses more em dashes. But the difference is much smaller for rewrites. The number of em dash characters per 1,000 words in our rewrite study: 4.14 for AI rewrites, 2.56 for the human originals (four bloggers who use many more em dashes than most people, which is 0.73 per 1,000 words in our corpus of 22 million words of all kinds of texts). In our draft study, the ratio was six times as much (4.31 for the AI drafts, 0.72 for the human originals). It also depends on the model: GLM-5.3 uses very many em dashes (10+ per 1,000 words in both studies) and Claude doesn't (Opus 5.5: 0.10 per 1,000 words for rewrites; Sonnet 5.5: 0.09 per 1,000 words for drafts). If a text doesn't have any em dashes, it may have been written by Claude.

Wikipedia editors have moved em dash overuse to their "historical indicators" section, dated 2022 to September 2026. They suggest that newer models are often set up to avoid em dashes entirely, which has a side effect: less varied punctuation, often leaning on commas and periods.

Depending on the model and task, you might find different results. Wikipedia cites a July 2026 Economist study that found Claude used more em dashes than professional writers. We looked at Claude Opus 5.5 polishing existing writing and Claude Sonnet 5.5 writing from scratch, and both used almost none. Both can be true, since results depend on the model, the task and the kind of writing. Just don't rely on this one.

New: genuinely, roughly, honestly, "here's" and "whether you're"

Some of the words that now distinguish AI from humans can be seen in the table below. Some show up in both studies; others only when AI rewrites, or only when it drafts.

New: "rather than" and "should not be confused with"

Wikipedia editors note a new 2026 tendency of AI that they call corrective framing. Since mid-2026, they say, chatbots often add disclaimers stating that something should be treated as one thing "rather than" another, noting that available information "does not establish" a particular conclusion, or stating that something "should not be confused" with another topic (a pattern Wikipedia rates as less reliable than the other two). Its editors write that these disclaimers were "extremely uncommon, verging on unheard-of" in AI output before mid-2026. In our data, we also find that "rather than" is much more common in AI text: 3.7 times the human rate when rewriting and 4.3 times when writing from scratch.

Which words are signs of AI writing?

When AI rewrites something, the most common words AI adds are intensifiers and softeners like genuinely, roughly, plenty of, essentially, truly, honestly, and simply: the words a model adds when it's told to make text sound friendlier. When AI writes from scratch, the most common words it adds are honestly, "whether you're", roughly, "rather than", genuinely, and "here's". Interestingly, "for instance" and "essentially" are no longer in the top words, because AI used them less than people did when writing from scratch.

Different models have different tells: for example, Claude Opus 5.5 uses "plenty of" a lot (0.68 times per 1,000 words), as well as "isn't," "roughly" and "genuinely." GLM-5.3 and MiniMax M3 use "genuinely" the most (1.25 and 1.09, respectively). MiniMax M3 uses "actually" more than any other model (2.43).

Wikipedia has a list of AI-associated words for GPT-6-era models (mid-2026 onwards): challenge, clearer, dependable, echoed, foster, leverage, matters, multifaceted, practical, prioritize, quietly, steady and universally. Note that leverage made Wikipedia's list but barely appeared in either of our studies. Lists like these vary by model and task, so these are weak tells individually. For what to write instead, see our guide to words and phrases that AI tends to use too much.

Is punctuation still a sign of AI writing?

Partly. The em dash character is still more common in AI drafts, but rates vary a lot by model, and people type dashes in other ways. When AI writes from scratch, colons are a clearer signal than they are in rewrites.

Em dashes when writing from scratch

Em dashes when rewriting

People use nearly as many dashes as AI, but they use different types of dashes. If we count all kinds of dashes (em dash character, en dash character used as a dash, double hyphens "--", hyphen surrounded by spaces " - "), the difference between people and AI is smaller: 4.67 for AI drafts, 2.97 for the human originals (1.6 times as much). People often use "--" (0.49 per 1,000 words), " - " (0.62) and en dashes (1.13), and AI doesn't (0.07, 0.01, 0.28). Emails show it best: 4,000 human emails have 3.8 dashes per 1,000 words (all kinds of dashes), but only 0.06 of them are em dash characters, while AI-drafted emails have 2.9 em dash characters.

For text that the AI writes from scratch, Pangram found that AI used em dashes about ten times as much as humans (though Gemini 3 Pro used them 0.7x as much as humans). We found about six times, so we agree on the direction. Also, in November 2025 OpenAI said that ChatGPT would follow custom instructions not to use em dashes, so a lot of ChatGPT-generated text will not use em dashes at all. Wikipedia lists a common style of AI-generated text as being the use of em dashes with a space on each side, which most style guides don't recommend.

In our data, for text where the AI is rewriting another text, the AI uses colons slightly more often (4.55 vs. 3.67 per thousand words, a ratio of 1.2x). They don't use semicolons (0.85 vs. 0.79) or exclamation marks (0.65 vs. 0.72) at different rates. Another feature that AI changes is colons, which grew 2.5x (9.6 vs. 3.8) when AI wrote from scratch. AI also tends to use fewer exclamation marks (1.77 vs. 3.01): AI drafts toned them down in LinkedIn posts (8.3 to 3.4) and newsletters (10.0 to 5.4). Finally, AI tends to use more emoji than people do (0.92 vs. 0.11). Wikipedia also lists the use of curly quote marks as being part of AI style. For more on the use of the dash, see why ChatGPT uses so many em dashes.

What does AI sentence rhythm look like?

AI-written text has more consistent sentence lengths. People write a long sentence, then a short one, then something in between. Models settle into a steady middle.

In our rewriting study, we calculated the coefficient of variation in sentence length over a held-out set of 471 passages. The coefficient of variation of the bloggers' passages was 0.65, compared to 0.59 for OpenAI rewrites and 0.56 for Claude rewrites. The bloggers also started sentences with And, But or So 12.0% of the time, compared to 9.7% for OpenAI and only 5.0% for Claude.

The difference between AI and people is smaller for AI drafts written from scratch: 0.50 for the human originals and 0.47 for the AI drafts. GPT-5.4 mini is the most uniform (0.43), then Gemini 3.8 Flash (0.46), Claude Sonnet 5.5 (0.49), GLM-5.3 (0.51). And/But/So turned out to be partly a blogger habit: it's used almost as often by people as by AI (3.8% vs 3.4% for the human originals and AI drafts, but note that the texts are different for these two studies, so don't compare numbers across the studies).

Wikipedia claims that GPT-6-era models tend to write shorter sentences, often between 8 and 20 words.

Try reading three sentences out loud. If they have the same rhythm, that can be a sign of AI.

Which structure and formatting habits are AI tells?

If the text was generated from scratch, formatting is the biggest clue in our data: headings, bold and bullet lists where a person would just write sentences.

We didn't ask AI to use any formatting, but it does: 10.1 phrases in bold per 1,000 words, 4.2 markdown headers, 8.9 lines in bullet lists (the human originals have 2.4 bullet lines per 1,000 words). Bold text is also used in emails (5.8 phrases in bold per 1,000 words) and is especially common for event pages (22.1), docs (20.6) and how-tos (19.1). We don't show the ratio between people and AI here because most of our human documents were stored as plain text (~1% of documents have bold text). But Pangram compared formatted texts and found much more formatting in AI-generated texts than in human texts (per 10k words): all kinds of formatting (12 times as much), bold text (43x), headers (23x), bullet lists (9x), groups of three like "fast, simple and secure" (4x).

Groups of three. According to Wikipedia, AI text often includes groups of three as a way to give a deeper air to a shallow analysis. Our AI drafts had 1.4 times as many as the human originals (5.60 vs. 4.00 per 1,000 words), a smaller gap than Pangram's 4x. One group of three is normal. Three groups of three in a paragraph is a clue.

"It's not X, it's Y" and "not just X, but Y". These are the most well-known clues. Wikipedia refers to this as negative parallelism. Pangram found that AI text written from scratch has about 3 times more "not just X but Y"s. This signal didn't show up overall in our two studies: AI rewrites have slightly less (0.12 vs 0.17 per 1,000 words) and AI drafts have the same as the human originals (0.12). But it is more common for LinkedIn posts (0.51 for AI drafts, 0 for the human originals) and academic abstracts (0.32), and it is much more common for some models (2.4 times as much for GPT-5.4 mini). More on why AI loves "it's not X, it's Y".

What tone and content habits give AI writing away?

What AI says is often a better clue than any specific word: it hedges, summarizes, reaches for generic examples and closes on a cheerful line.

• Using hedging language: Phrases like "X may be," "X can help," "based on available information" or "does not by itself establish." Wikipedia lists these as a pattern from mid-2026 on.
• Summarizing key points at the end: A final paragraph that repeats everything it just said. Wikipedia also flags "outline-like conclusions about challenges and future prospects."
• Using generic examples: Phrases like "For instance, a small business might use this to save time," without citing specific people or events, suggest AI-generated content.
• Referencing vague sources: Phrases like "experts say" or "studies show" without naming specific experts or studies can indicate AI. Wikipedia notes that AI also exaggerates how many sources back a claim.
• Using dramatic language: Elevating something to "a pivotal moment" or "a lasting legacy" when it isn't particularly noteworthy is a common AI trait.
• Concluding with overly positive statements: Phrases like "I hope this helps," "Let me know if you have any questions" or "Exciting times ahead." Wikipedia lists "I hope this helps" as a phrase to watch.

All of these behaviors can be traced back to one core principle: if the AI doesn't have specific information, it won't provide specific details. A person writing about their own work mentions the client's name, the date it broke, the number that surprised them.

What isn't a sign of AI writing?

Correct grammar, a formal style, bland writing and transition words aren't signs of AI on their own: they're very common in human writing. These are all listed as "ineffective indicators" in Wikipedia's Signs of AI writing guide.

AI doesn't use contractions (like "don't") much more than people do (15.0 vs. 12.7 per 1,000 words), so they aren't a tell. The same can be said for em dashes: we find 0.73 em dashes per 1,000 words in 22 million words of pre-ChatGPT writing, and the bloggers in our rewriting study used 2.56 per 1,000 words. This might be considered normal for some types of writing but not others. For example, a single em dash is unusual for someone's email writing (0.06 em dashes per 1,000 words) but normal for someone writing a news story (1.62). And people are worse at this than they think: according to Wikipedia, a 2025 study found people were no better than random chance at detecting AI writing, and even AI detectors can have "non-trivial error rates." If you want to tell if someone else's writing was written by AI, check out our guide for how to tell if something was written by AI.

What does removing these tells look like?

Here's a short product update written by an AI model, with most of the 2026 tells in it. After this, we'll show a rewrite from Zavus.

The first version uses words and phrases like "genuinely," "crucial," "navigate," "rather than," "essentially," "for instance," "plenty of," "roughly," and "simply." It also uses the "triad" structure, an em dash, a summary line, and an upbeat closer.

A list of AI tells to paste into your agent's instructions

To fix this in your own writing, just paste the below into your agent's instructions, either as a CLAUDE.md or AGENTS.md file, in custom instructions in ChatGPT, a Claude project, Cursor rules, etc. It should work in any agent you're using.

AVOID THESE AI WRITING TELLS (zavus.ai/guides/signs-of-ai-writing, Nov 2026)

Words
- Cut intensifiers: genuinely, genuine, truly, honestly, essentially, simply, actually.
- "roughly": give the number, or use "about" once.
- "plenty of": say how many.
- "for instance": just give the example.
- No "Here's..." setups and no "whether you're X or Y".
- Avoid: crucial, navigate, delve, tapestry, testament, landscape, leverage,
  foster, multifaceted, underscore, showcase, pivotal, seamless, robust.

Punctuation
- No em dashes. Use a comma, a period, a colon or parentheses.
- Straight quotes, not curly quotes.
- Don't put a colon in every paragraph.

Rhythm
- Vary sentence length. Put short sentences next to long ones.
- Starting a sentence with And, But or So is fine.

Structure
- No "It's not X, it's Y" and no "not just X, but Y".
- Don't group things in threes by habit.
- No headings, bold text or bullet lists in emails and short posts.
- Headings in sentence case, not Title Case.

Tone and content
- Don't hedge. State the claim, or say plainly what you don't know.
- No "rather than" corrections of claims nobody made.
- No "should not be confused with" unless readers really would confuse them.
- Use specific examples: names, numbers, dates. No generic "a small business might..."
- Attribute claims to a named source, or drop them. No "experts say".
- Don't inflate significance ("pivotal moment", "lasting legacy").
- Don't summarize at the end. No "In short", "Overall" or "Ultimately".
- No upbeat closers: "I hope this helps", "Let me know if you have any
  questions", "Exciting times ahead", "Can't wait to hear what you think".

Important note: This isn't a total cure, but for short pieces of text (like emails), it might be enough. For longer pieces of text, the model will drift back to its default style, particularly for harder-to-pin-down tells like rhythm and structure. And if you have the model check its own writing against the list, it will usually switch to different tells. See why your AI can't fix its own writing for more.

How often is this list updated?

We will revisit this page every quarter to rerun our dataset on the latest models, and to also incorporate updates to Wikipedia's or Pangram's lists. Our next planned update for this page is February 2027.

• November 2026 (update): Added our second study, 1,177 drafts written from scratch by Claude Sonnet 5.5, GPT-5.4 mini, Gemini 3.8 Flash and GLM-5.3, compared with 692 human originals and a 22.2-million-word human baseline across 20 kinds of writing. The formatting section now leads with our own drafting numbers. We corrected the human em dash baseline (the 2.56 was the bloggers; across all kinds of writing it's 0.73) and the claim that "not X but Y" is a from-scratch tell.
• November 2026: First version of this page; based on 7,554 passages of human writing that were rewritten by models from 2026, including models from Anthropic, OpenAI, DeepSeek, Zhipu (GLM), Moonshot (Kimi), MiniMax, and Alibaba (Qwen).

If you don't want to spend time manually removing AI tells from your writing, consider using Zavus. Zavus is a separate model that rewrites text to remove AI tells. You first generate a draft with your favorite model, and then your agent sends that draft to Zavus to be rewritten. This way, all of your agent's writing (emails, posts, announcements, etc.) can sound human. See our guide to making AI-generated writing sound human, or if you work in marketing, our page for marketing teams.