You open an email, a post, or a school assignment and something sounds off. The grammar is correct, the sentences are well-arranged, and yet you suspect you're looking at AI-generated text. The hard part is saying why.
On October 1, 2026, an agency called Graphite published a measurement that helps answer that question. It compared thousands of articles from ten AI models (the programs behind chatbots like ChatGPT) with articles written by people, word by word, and listed what Claude Opus 5.5, Anthropic's newest model, repeats far more often than a person would.
This article shows what the study found, explains why these tics exist, and teaches a reading method that works without memorizing lists or relying on detectors.
Quick summary: the answer in 30 seconds. Opus 5.5 has almost stopped using em dashes and writes with words closer to those of a person, but still has 2,548 measured tics. The strongest is announcing that something matters: the phrase this matters appears 116 times more than in human text. Each model has its own tics, which change with each version. Detectors are not reliable. What works is looking for phrases that would fit any topic and the lack of concrete detail.
Contents
- What the study found in Opus 5.5
- The tics that give away AI-generated text
- Each model has its own accent
- Why AI writes this way
- Why AI detectors don't work
- How to recognize AI text without memorizing lists
- What this changes for those who write with AI
- Frequently asked questions
What the study found in Opus 5.5
The study found 2,548 writing tics in Opus 5.5, only 4% fewer than the 2,666 in Opus 5, the previous version. The model improved on famous tics, like the em dash, and ended up with vocabulary more similar to a person's. Even so, it continues using thousands of words and phrases in excess.
A tic, in this study, has an exact definition. The authors call a tell (the signal that gives it away, like a poker player's bluffing tic) any word, phrase, or sentence pattern that appears at least twice as often in AI text as in human text.
Think of a friend who says "like" in every sentence. Everyone uses the expression occasionally. What gives the friend away is the quantity. The same happens with AI: none of these words is forbidden or wrong, and the signal is in the frequency.
How the measurement was done
Graphite is a marketing agency with clients like Adobe and Webflow, and the research is signed by Gregory Druck, the company's AI director. The original study came out on September 16, 2026. The update with Opus 5.5 came two weeks later.
The human side of the comparison consists of 10,000 articles from websites, all published before November 30, 2022, the day ChatGPT was launched. The date ensures that no one used a chatbot to write them.
The AI side was assembled with extra care. Each human article was summarized, and each model received only the summary, with the task of writing a new article of the same length. This way the AI doesn't copy phrases from the original, and both texts deal with the same subject. In the update, the count included the 9,974 topics that had text from all ten models.
What truly improved
Three measures improved. The first is the em dash, that long dash that became synonymous with chatbot text. Opus 5 used 2.92 em dashes per 1,000 words. Opus 5.5 uses 0.015. That's a 99% drop, which in practice means one em dash every 67,000 words, roughly the length of a short novel.
The second is what the authors call affected prose: metaphor and embellishment instead of direct statement. On a scale from 0 to 100, given by another AI model that read the texts without knowing their origin, human text scored 6.65. Opus 5 scored 16.75 and Opus 5.5 dropped to 10.57, still 1.6 times the human score.
The third is the overall distance between the model's vocabulary and human vocabulary, which fell 19% compared to Opus 5. This matches what Anthropic promised in the Opus 5.5 announcement on September 22: a model "less prone to using jargon or idiosyncratic expressions," meaning its own language quirks.
In one sentence: Opus 5.5 writes more like people than Opus 5, but eliminated few tics and replaced many with others.
The tics that give away AI-generated text
Opus 5.5's tics group into four families: announcing that something matters, defining something by contrast with another, praising with vague adjectives, and stitching paragraphs together with transitional phrases. All were measured in English. The translations below help you recognize the pattern, and the proportions apply only to English. In the table, the underscore (_) marks the place for any word.
| Family | Measured example | Translation | How many times more than human |
|---|---|---|---|
| Announcing importance | this matters | "this matters" | 116 |
| Announcing importance | why _ matters | "why _ matters" | 92 |
| Ready-made contrast | is more than a _ it | "is more than a _, it" | 98 |
| Ready-made contrast | rather than simply | "rather than simply" | 32 |
| Transitional phrase | looking ahead | "looking ahead" | 40 |
| Transitional phrase | adds another layer | "adds another layer" | 27 |
| Praise adjective | dependable | "dependable" | 23 |
| Praise adjective | steady | "steady" | 11 |
Announcing that something matters
This is the model's strongest tic. Instead of showing the consequence of a fact, the text stops and declares that the fact is relevant. This matters appears 116 times more than in human articles, and the pattern why _ matters, with any word in the middle, 92 times more.
Compare two sentences about the same subject. The first: "The battery lasts two days. This matters." The second: "The battery lasts two days, so you can travel on the weekend without bringing the charger." The second doesn't need to announce anything, because the consequence is already written.
The ready-made contrast
In April 2026, TechCrunch showed how the construction "it's not just X, it's Y" had invaded company press releases. Opus 5.5 abandoned this form, which became too well known. In its place came a cousin: "is more than an X, it's a Y," 98 times more common than in human text.
The trick is the same. The sentence invents a weak idea just to knock it down and seem profound. "A cellphone is more than a phone, it's a window to the world" doesn't inform anything the reader didn't already know.
Praise adjectives and transitional phrases
The third family consists of adjectives that approve without describing. Dependable appears 23 times more, steady 11 times, thoughtful 9 times. These are words that fit in any review of any product.
The fourth are bridges between paragraphs: what comes next (24 times), looking ahead (40 times), adds another layer (27 times). A human writer usually moves from one idea to another without announcing the transition.
Each model has its own accent
Models err each in their own way. In the original study, Graphite counted 12,877 different tics across nine models, and 65% of them appear in a single model line (only in Claudes, only in GPTs, or only in Geminis). Opus 5.5 wasn't yet in that count. Someone who learns one model's list won't recognize another's.
| Model | Typical tic | Em dash | Trend |
|---|---|---|---|
| Claude Opus 5.5 (Anthropic) | announces that something matters; superlatives like "the most popular" | 99% less than Opus 5 | vocabulary increasingly closer to human |
| GPT-6 Astra (OpenAI) | hedges: may provide, not simply | 88% below human level | vocabulary increasingly farther from human |
| Gemini 3.1 Pro (Google) | intensifiers like "absolutely" and "incredibly," 13 times more than human | almost zero | affected prose at about 2.5 times the human score |
The direct comparison between Opus 5.5 and GPT-6 Astra helps hear the difference. Astra hedges: it uses may provide 18 times more than Opus 5.5. Opus 5.5 exaggerates: it writes "the most popular" 45 times more than Astra. One sounds like a cautious lawyer, the other like an enthusiastic salesperson.
The two are also moving in opposite directions. The study measures the distance between each model's vocabulary and human vocabulary, on a scale where zero would be text identical to people's. In Claude it has been falling, from 0.064 in Opus 5 to 0.052 in Opus 5.5. In OpenAI it rose, from 0.101 in GPT-5.6 Sol, the previous version, to 0.109 in GPT-6 Astra.
Why AI writes this way
AI has tics because it writes by choosing, at each step, a likely continuation for the sentence. What appears often in the texts it read and pleases those who evaluate responses before launch becomes habit. The labs, as the companies that create the models (Anthropic, OpenAI, and Google) are called, cut the tics that become famous, but don't control all of them.
A language model is a program that learned to predict the next word by reading an enormous amount of text. It works like a phone keyboard that suggests the next word, just on a much larger scale. Ask for a sentence about a product and "dependable" is a safe bet, because it fits almost anything.
After training, companies adjust the model to sound helpful and clear. This is where excesses are born. A text that announces "this matters" seems organized to someone evaluating one response at a time. The problem only appears when someone reads a thousand texts and sees the same phrase in all of them.
The em dash shows what happens next. The signal became a joke on the internet, and usage changed in all three labs. Opus 5.5 and Gemini 3.1 Pro almost eliminated it, and Astra started using fewer em dashes than people use. The correction overshot: today, the total absence of the dash in a long text is what deviates from the human pattern.
In an interview with TechCrunch, Druck said the tics are not decreasing: the best-known ones leave and others emerge, and each version has its own. His hypothesis is that labs control this less than imagined, because models have billions of parameters (think of billions of adjustment knobs, regulated on their own during training) and the number of tests is finite.
The numbers support the hypothesis. The 11 most famous tics, like the word delve and paragraphs that always end with a punch line, lost 53% of their strength from Opus 4 to Opus 5.5. In the same interval, the total count of tics fell from 3,746 to 2,548, a 32% drop. The famous ones shrink much faster than the set.
Why AI detectors don't work
AI detectors don't work because they err in both directions: they let lightly edited machine text pass and accuse human text. Since tics change with each version, the tool is always trained for the previous model. A detector's result serves as a clue and never as proof.
A detector is another AI program, trained to say whether a text seems machine-made. It gives a probability, the same way a weather forecast gives chance of rain. A warning of "80% AI" means the tool estimates an 80% chance the text is machine-made, not that 80% of it was generated.
In 2023, a group led by researcher Debora Weber-Wulff tested 14 of these tools, including Turnitin, an academic assignment checking program. The conclusion of the published paper is direct: they are neither accurate nor reliable, tend to classify AI text as human, and any simple disguise worsens performance.
The inverse error is more serious. A study by Weixin Liang and colleagues showed that detectors classify as AI texts by people who don't have English as their native language. The simpler and more predictable writing of someone who learned the language later resembles, to the tool, that of a chatbot. For a student, this can become an accusation of fraud.
There is a technical solution, watermarking, similar to that on a banknote: invisible in common use and clear to those who know where to look. The idea, described in 2023, is for the model itself to give slight preference to a secret group of words while writing. The reader doesn't notice, and a program that knows the group can measure. It only works if the company that owns the model activates the feature, and nothing forces someone who wants to deceive to use a watermarked model.
How to recognize AI text without memorizing lists
The safest path is to look for what the lists have in common: phrases that would fit any topic, importance announced without consequence, praise without description, and absence of detail that only someone who lived the situation would know. These signals depend on the idea of the sentence, and therefore tend to work in other languages and versions, something a word list doesn't do.
The best public reference was born on Wikipedia. The volunteer editors, who receive chatbot-generated contributions every day, maintain a guide called "Signs of AI writing." TechCrunch called it the best available guide on the subject.
Next time you suspect a text, ask these five questions:
- Would the sentence work in another text? "An important milestone for the sector" fits any news story. Human sentences are usually tied to the subject.
- Does the text say something is important or show the effect? Look for the concrete consequence. If it doesn't come, it's empty emphasis.
- Who said so? "Experts point out" and "studies indicate," without name or link, are vague attributions that Wikipedia's guide lists as signals.
- Are there names, dates, places, numbers? Text by people who reported has verifiable detail. Generic text talks about "researchers" and "companies."
- Does the rhythm vary? Paragraphs of the same size, all closing with a punch line, indicate formula.
See the five questions in action in an example we assembled with the study's tics: "The bank's new app is more than a tool, it's a dependable partner. This matters. Looking ahead, experts point out that the innovation adds another layer to the customer experience."
None of these sentences says what the app does. Now the version from someone who used it: "The bank's new app shows the balance without asking for a password and does transfers in two taps. It crashed twice in the first week." Less elegant, and with facts you can check.
One signal helps in the opposite direction. In Graphite's study, the most human markers were those of informality: first person, loose expressions, and exclamation points. The exclamation point alone appeared in human articles more than 100 times more than in AI ones. Back in 2022, Daphne Ippolito, a Google researcher, told MIT Technology Review that typos are a good indication of human authorship.
Now the brake. Wikipedia's guide itself warns that these patterns are possible signals of a problem, and that many people write this way without any AI. It also asks not to rely only on detectors and cites a 2025 study in which people trying to distinguish AI text were right as often as chance. An isolated signal proves nothing. Several together justify a question.
What this changes for those who write with AI
It changes revision. If you use AI to write, the real risk is delivering generic text, with the model's quirks in place of your voice. Being "caught" by a detector is the least of the problems. And if you read or evaluate others' texts, the lesson is caution before accusing.
| If you are | What to do |
|---|---|
| Student or professional who writes with AI | Reread cutting every announcement of importance and every adjective that doesn't describe. Add the detail only you know. |
| Teacher or manager who evaluates texts | Don't use a detector as proof. Talk about the content: whoever wrote it knows how to explain what they wrote. |
| Regular reader | Use the five questions to measure text quality. Empty text is bad, wherever it comes from. |
| Content creator for websites or social media | Give the model writing rules and your own examples. Anthropic says Opus 5.5 follows the writing rules it receives. |
Switching models doesn't solve it, because each has its own accent. If you want to understand what Anthropic's model delivers beyond text, it's worth reading our analysis of Claude Opus 5.5. To follow other releases, the artificial intelligence page gathers everything we've published on the subject.
Frequently asked questions
Does the em dash still give away AI text?
Less and less. Opus 5.5 uses 99% fewer em dashes than Opus 5, Gemini 3.1 Pro hardly uses them, and GPT-6 Astra uses fewer than a person. A text full of em dashes today may come from an old model or from a human who likes the punctuation mark.
Can you trust an AI detector?
As proof, no. The test with 14 tools concluded they are neither accurate nor reliable, and another study showed they unfairly accuse those who write in English as a second language. The result serves, at most, as reason to look at the text more carefully.
How to tell if a text is AI in other languages?
Graphite's study only provides examples in English, so there is no measured list for other languages. What you can take are the patterns: importance announced without consequence, ready-made contrast, vague praise, and lack of concrete detail. They depend on the idea of the sentence, and the idea survives translation.
Will AI writing tics disappear?
By the data, they change more than they vanish. From Opus 5 to Opus 5.5 the count fell only 4%. And in the original study, comparing two consecutive versions from the same manufacturer, more than half the tics appeared in only one of them. It's more useful to learn the pattern than the word.
Does the study have limitations?
Yes, and the authors list several. The texts are website articles, and the tics may be different in a conversation with the chatbot. All models received the same instruction, and different instructions would give different results. The affected prose score was given by another model, Claude Opus 5, which may distort the score.
Next time a text sounds off, you already know where to look: less at the em dash, more at the sentence that promises importance and delivers nothing. And if the text is yours, the revision is the same.
Originally published on kalkazu.org.
Top comments (0)