20 New Signs AI Wrote It, Now That the Em Dash Is Gone
AI stopped using the em dash. Here's how to tell if ChatGPT, Claude, or Gemini wrote something now: the exact words and shapes each one leans on, with the numbers, what real people write that AI doesn't, and a sixty-second check for any text.
Save it, print it, read it offline. It's yours.
The em dash was the easy way to spot AI writing, and it stopped working in September. The newest ChatGPT, Claude, and Gemini dropped it almost completely. Keep hunting for it and you’ll flag real people and wave the AI through. Here are the 20 signs that work now, each with the number behind it, what real people write that AI can’t, and a sixty-second check for any text someone sends you.
An AI tell is a word, phrase, or sentence shape that an AI model uses far more often than people writing about the same thing. The em dash used to be the big one. Now the tells are things like "this matters," "not simply," "furthermore," no contractions, no exclamation points, and every sentence the same length.
Nothing to install. The signs below work on any text you can read, on your phone or your computer. Ctrl+F or Command+F is the only tool you need.
The companies haven’t said why, but the timing fits: the em dash became the one thing non-writers could spot, it got mocked everywhere, and the newest models shipped without it. Two separate measurements in late September 2026 caught the drop.
Graphite, a marketing research firm, had already compared 10,000 articles written before ChatGPT existed with 90,000 AI versions of the same topics. On October 1 they added Claude Opus 5.5, Anthropic’s newest model, which shipped in late September. Here is the em dash count per 1,000 words, from their data:
| Writer | Em dashes per 1,000 words |
|---|---|
| People (10,000 pre-ChatGPT articles) | 3.1 |
| Claude Opus 5 (the model before) | 2.9 |
| Claude Opus 5.5 (new) | 0.015 |
| ChatGPT-6 Astra | 0.4 |
| Gemini 3.1 Pro | 0.014 |
That’s a 99 percent drop between the two Claudes, and the new Claude now uses the em dash about 200 times less often than a person does. ChatGPT’s newest model is at about one eighth of the human rate. Gemini is near zero.
A second measurement, by Arena, looked at chat answers instead of articles and found the same thing from a different angle: the old Claude put 15.2 em dashes in every 1,000 words of its answers, the new one puts in 0.8. That’s the 95 percent number you’ll see quoted. Semicolons fell 73 percent in the same comparison, and sentences got 17 percent shorter.
The Economist ran its own test over the summer, 1.2 million words through ChatGPT, Claude, Gemini, and Grok, and reached the same conclusion before the new Claude even shipped: the em dash is no longer a reliable marker of AI writing, and ChatGPT already used fewer of them than their human writers did.
So the dash is gone. What didn’t go anywhere is everything else. Graphite counted 2,548 tells in the new Claude, only 4 percent fewer than in the old one. The tells didn’t disappear. They changed. Here are the twenty that matter.
The first six signs are things real people put in without thinking and AI leaves out. They show up in ChatGPT, Claude, and Gemini alike, so check for them first. Numbers are per 1,000 words from Graphite’s measurements.
1. No exclamation points. People use about 2.8 per 1,000 words. The three current models use between 0.02 and 0.04. That’s more than a hundred to one. One real exclamation point is one of the surest signs a person was there.
2. No parentheses. People drop a quick aside in parentheses 7.7 times per 1,000 words (like this). ChatGPT’s newest model does it 0.05 times. Claude, about once. Gemini, 0.04.
3. Almost no semicolons. This one is the opposite of what most people expect, because semicolons used to be an AI tell. People use 1.4 per 1,000 words; the new Claude uses 0.14, Gemini 0.3, ChatGPT 0.6. The AI got rid of its fancy punctuation. People never did.
4. Every sentence is the same length. Human sentences swing. In Graphite’s human articles the average sentence was 27 words and the longest in a piece ran past 70. AI sentences land between 16 and 19 words and the longest rarely passes 40. Read three sentences in a row. If they all feel the same size, that’s the shape.
5. No contractions. “It’s,” “don’t,” “we’ve,” “can’t,” “that’s.” People write them constantly. ChatGPT’s newest model writes “it’s” about 100 times less often than a person, and Gemini almost never writes it or “don’t” at all. The new Claude contracts more than it used to, so this sign is weakest on Claude and strongest on Gemini.
6. No year, no date, no place. People mention a year about six times per 1,000 words. The models mention one about once. People write “last March” and “in 2019” and “when we moved to Denver.” AI writes “in recent years.”
The next two signs are about how the text is built, not which words are in it. A September 2026 study caught AI blog posts from structure alone 97 percent of the time, and 96 percent even after the AI reworded its own work, so these two hold up when the words change.
7. It tells you what it’s about to say, says it, then says it again. That study, 2,250 company blog posts against 11,250 AI copies, found AI posts “state the thesis and announce the structure before the first section,” then close “with a stage that summarizes or restates the thesis.” Graphite measures the same thing: an AI article’s last paragraph mirrors its first about twice as often as a human one does.
8. “Rather than” everywhere. This two-word phrase is the single most widespread tell in the data. ChatGPT’s newest model writes it 22 times more often than people and puts it in 84 percent of everything it writes. The new Claude writes it 8 times more often, in half of its articles. People write “instead of,” or they write the one thing they mean.
Claude’s tells are quiet. They sound like a thoughtful consultant. These are from Graphite’s October 1 measurement of Claude Opus 5.5, with how many times more often Claude writes each one than a person does.
| What Claude writes | Times more than a person | Where it shows up |
|---|---|---|
| “this matters” | 116x | the new number one |
| “why [this] matters” | 92x | in 756 of 9,974 articles |
| “it’s more than a [tool], it’s a” | 98x | the shape, with any nouns |
| “article explains” / “explains what” | 90x / 35x | it narrates itself |
| “together these” / “taken together” | 56x / 42x | the wrap-up line |
| “the result is a” / “a real difference” | 38x / 45x | the payoff line |
| “often called” | 58x | before a definition |
| “what comes next” / “looking ahead” | 24x / 40x | the closing turn |
9. “This matters.” Claude’s biggest single tell. The word “matters” alone appears in 45 percent of everything the new Claude writes, 13 times more often than in human writing. If a paragraph pauses to tell you that something matters, or why it matters, instead of showing you, that’s Claude.
10. “It’s more than a tool, it’s a.” The old “it’s not X, it’s Y” didn’t die, it changed clothes. Claude’s version is “more than a,” 98 times the human rate. People say the one thing. They don’t build a little staircase to it.
11. “Together these” and “the result is a.” Claude’s way of tying a bow on a list. A person who just gave you three points stops. A Claude paragraph ends with a line like “Together, these three changes make a real difference,” and that’s two tells in one sentence.
12. The calm adjectives. Dependable (23x), steady (11x), practical (11x, in a third of its articles), thoughtful (9x), genuine (9x), quietly (11x), meaningful (8x). None of these is a bot word on its own. All of them together is a Claude paragraph.
ChatGPT’s newest model, ChatGPT-6 Astra, writes in contrasts. It defines things by what they are not, and it almost never contracts. These are from Graphite’s September measurement.
| What ChatGPT writes | Times more than a person | Where it shows up |
|---|---|---|
| “is not simply” | 270x | its single biggest phrase |
| “not automatically” | 233x | in 1,614 articles |
| “rather than relying” | 187x | the contrast move |
| “not simply” | 157x | in 3,082 of 9,974 articles |
| “extends beyond” | 146x | reaching for scope |
| “is straightforward” | 102x | reassurance |
| “together these” | 95x | the same bow Claude ties |
| “not merely” | 77x | the formal cousin |
| “adds another” | 75x | “adds another layer” |
13. “Not simply” and “is not simply.” ChatGPT’s signature. “Good onboarding is not simply a checklist.” A person would write “Good onboarding is more than a checklist,” or more likely skip the sentence and tell you what good onboarding is.
14. “Rather than relying” and “not automatically.” Both are ways of saying “instead of,” and both are a hundred-plus times more common in ChatGPT than in people. When a sentence tells you what not to rely on before it tells you what to do, that’s the pattern.
15. The clarity words. Distinction (46x), clearer (43x), manageable (34x), broader (31x), practical (26x, in more than half its articles), useful (17x), dependable (59x, even more than Claude). ChatGPT loves to tell you a thing is clear, practical, and useful. People say the thing.
16. “Does not,” never “doesn’t.” ChatGPT writes “does not” almost nine times more often than people and puts it in more than half of what it writes. Paired with almost no “it’s,” “can’t,” or “we’ve,” you get prose that’s grammatically perfect and slightly stiff. That stiffness is the tell.
Gemini is the loud one. Where Claude is calm and ChatGPT is precise, Gemini is impressed by everything. These are Gemini 3.1 Pro’s numbers from Graphite’s September measurement.
| What Gemini writes | Times more than a person | Where it shows up |
|---|---|---|
| “ultimately this” | 78x | the closing turn |
| “this comprehensive” / “comprehensive guide” | 61x / 47x | it names its own format |
| “furthermore the” | 43x | the stacked transition |
| “we will explore” | 47x | announcing the structure |
| “is a massive” / “as a powerful” | 60x / 56x | the hype |
| “whether you are” | 35x | in 2,197 articles |
| “a testament to the” | 27x | the cliché |
| “is not just a [thing], it is” | 153x | its “not just X, it’s Y” |
| “just a [thing], it” | 46x | the same shape, shorter |
17. The stacked transitions. Furthermore (26x, in almost a third of its articles), ultimately (7.7x), “furthermore the,” “ultimately this.” People rarely start a sentence with “furthermore.” Gemini can’t stop.
18. The intensity words. Incredibly (18x, in 37 percent of its articles), massive (19x), profound (26x), immense (22x), highly (16x, in more than half), deeply (18x), heavily (18x), crucial (9x), fundamentally (29x). One of these is a word. Four in a paragraph is Gemini.
19. “Whether you are a…” and “this comprehensive guide.” Gemini writes for an imaginary reader it names at the top (“whether you are a small business owner or a freelancer”) and describes the thing you’re reading (“this comprehensive guide will explore”). People write to you, and they don’t describe the article inside the article.
20. “A testament to,” “navigating,” “the world of.” The old buzzword set is mostly gone from Claude and ChatGPT. It’s alive in Gemini: “a testament to” (27x), navigating (26x), “the [modern] world of” (10x), “the power of” (10x). If a paragraph reads like a 2024 LinkedIn post, check Gemini first.
This is the other half of the data, and it’s the half that helps you write. Graphite’s human articles are full of things the models almost never produce. Each one below is how many times more often a person writes it than the model named.
| What people write | Times more than AI | Against which model |
|---|---|---|
| “it’s” | 108x / 1,942x | ChatGPT / Gemini |
| “don’t” | 803x | Gemini |
| “we’ve” | 30x / 256x | Claude / ChatGPT |
| “didn’t,” “can’t,” “won’t,” “that’s” | 97x to 223x | ChatGPT |
| “I think” / “I like” / “I would” | 32x / 38x / 19x | Claude |
| “pretty” | 19x / 41x | Claude / Gemini |
| “a bit” / “a lot of” | 182x / 171x | ChatGPT |
| “definitely” | 80x / 116x | Claude / ChatGPT |
| “etc.” | 100x | Claude |
| “going to” / “going to be” | 19x / 74x | Claude |
| “check out” / “be sure to” | 19x to 141x | Claude, ChatGPT |
| “amazing” | 41x | Claude |
Graphite’s own list of the most human words adds “my favorite,” “mom,” “last year,” “pretty,” and “sort of,” the small personal words a person drops in without thinking. Add the punctuation: the exclamation points and parentheses from Part 2, question marks (people ask about three per 1,000 words, the models about one), and real lists. Graphite’s human articles had about 15 bulleted lines per 1,000 words. The models, writing the same topics, had almost none.
Read that table as a recipe. A person writes “it’s,” says “I think,” calls something pretty good, mentions last year, throws in an aside (like this one), and ends a sentence with an exclamation point when they mean it. None of that is sloppy. It’s what being there sounds like.
Someone sends you a paragraph and you want to know. Do this, in order, and stop when you have your answer.
- Search for “rather than.” Press Ctrl+F or Command+F and type it. Two or more in a short piece is a strong sign on ChatGPT and Claude.
- Search for “matters.” One “this matters” or “why this matters” is Claude.
- Search for “not simply.” That’s ChatGPT.
- Search for “furthermore” and “incredibly.” That’s Gemini.
- Count the contractions. A long piece with no “it’s” or “don’t” anywhere is AI, most likely ChatGPT or Gemini.
- Look for one exclamation point, one set of parentheses, one year. If none of the three is there, lean AI. If all three are, lean human.
- Read the first and last sentences together. If the last one restates the first one, that’s the shape.
Three or more hits and you can be fairly sure. Zero hits and a few of the human habits from Part 7, and a person wrote it, em dashes or not.
These are odds, not proof. "Practical" is a normal word, and a person can write "this matters" on a good day. Every multiplier in this guide means "far more often than people writing about the same thing," not "only AI says this." Use the signs to decide where to look closer, and never to accuse someone from a single phrase.
- Graphite, AI Tells: Opus 5.5 Update (October 1, 2026), for the Claude Opus 5.5 em dash count, the 2,548 tells, and the “this matters,” “why _ matters,” and “is more than a _, it” multipliers.
- Graphite, AI Tells (September 16, 2026), for the method (10,000 human articles, 90,000 AI articles, 9,974 topics), the ChatGPT-6 Astra and Gemini 3.1 Pro tells, and the human tells. Their measurements are published under CC BY 4.0 and the per-model tables in Parts 2 through 6 are drawn from their downloadable data.
- BleepingComputer: Claude Opus 5.5 uses 95 percent fewer em dashes, for Arena’s chat-answer measurement (15.2 to 0.8 em dashes per 1,000 words, semicolons, sentence length).
- SlopShape: Identifying AI-Generated Commercial Web Content (September 14, 2026), for the structure findings and the 97.0 and 96.1 detection figures.
- The Economist’s analysis, as reported August 4, 2026, for the 1.2-million-word comparison and the punctuation, nominalization, and “it’s not X, it’s Y” findings.
Want more like this?
One simple AI win in your inbox every Sunday. No jargon, no overwhelm, unsubscribe whenever you like.
Want more like this?
Free AI guides added regularly. No jargon, no overwhelm.
kellystrattonai.com