AI writing tells are the habits that give machine-written prose away, and Claude Opus 5.5 has a clear favourite: it keeps telling readers that things matter. A study published on 1 October 2026 by Graphite, a growth marketing agency, found that Anthropic’s flagship model uses the phrase “this matters” 116 times as often as human writers, and the pattern “why _ matters” 92 times as often.
The research is an update to Graphite’s September study of AI writing tells across nine models. It adds Opus 5.5, released on 22 September, and finds a mixed picture. The model’s word choice is closer to human writing than its predecessor’s, and it has almost stopped using em dashes. Yet Graphite still counts 2,548 words, phrases and patterns that Opus 5.5 overuses, only 4% fewer than Opus 5.
This article explains how Graphite measured the AI writing tells, which ones belong to Opus 5.5, how they compare with OpenAI’s GPT-6 Astra, why the tells keep returning, where the study is weak, and how to edit them out of a draft. For the wider question of spotting machine text, see our guides on how to detect AI writing and on how reliable AI detectors are.
Table of contents
- How Graphite Measured AI Writing Tells
- Opus 5.5’s Biggest AI Writing Tells
- How Opus 5.5 Differs From Opus 5
- Opus 5.5 Against GPT-6 Astra
- Why AI Writing Tells Keep Coming Back
- The Limits of the Study
- How to Edit AI Writing Tells Out of a Draft
- What AI Writing Tells Mean for Businesses
- AI Writing Tells FAQ
- References and Further Reading
How Graphite Measured AI Writing Tells
The study’s design is the reason its numbers carry some weight. Graphite compared each model with human writers on the same topics, so a word only counts as a tell if the model uses it far more than people writing about the same thing.
A human baseline from before ChatGPT
Graphite gathered 10,000 articles from Common Crawl, the public web archive, all published before ChatGPT launched on 30 November 2022. That cut-off makes the human corpus very unlikely to contain machine text.
Each model rewrote the same topics
GPT-4.1 summarised each human article. Every model under test then wrote a fresh article from that summary, at roughly the length of the original. Graphite calls this a “bottleneck” method: the AI article covers the same ground without copying the human one. The first study used nine models and produced 90,000 AI articles. The Opus 5.5 update compares 9,974 topics where every model and the human article are present.
What counts as a tell
A word, two- or three-word phrase, or “frame” counts as one of the AI writing tells if it appears at least twice as often in a model’s articles as in the human ones, after adjusting for length. Frames are patterns with gaps, written with an underscore, such as “why _ matters”. Each candidate also has to appear in hundreds of articles, so one-off quirks drop out.
| Study detail | Graphite’s approach |
|---|---|
| Human corpus | 10,000 web articles published before 30 November 2022 |
| AI corpus | One article per topic from each model, written from a GPT-4.1 summary |
| Topics compared in the update | 9,974 |
| Tell threshold | At least twice the human rate, after length normalisation |
| Minimum spread | 500 articles for words, 250 for two-word phrases, 100 for three-word phrases, 120 for frames |
| Prompt | One fixed instruction: an “expert content writer” in a “professional yet friendly tone” |
Opus 5.5's Biggest AI Writing Tells
Graphite groups the Opus 5.5 tells into recurring themes. Four stand out.
Flagging importance
This is the headline finding among Opus 5.5’s AI writing tells. The model tells the reader that something is important instead of letting the point carry itself. “This matters” appears at 116 times the human rate, “why _ matters” at 92 times, “just as important” at 13 times and “matters most” at 12 times.
Rating words
Graphite calls these evaluative adjectives: words that rate something without describing it. Opus 5.5’s biggest single-word tell is “dependable”, at 23 times the human rate. “Steady” appears at 11 times. Gizmodo’s write-up gives the “dependable” figure as 26 times, but Graphite’s own table, and TechCrunch, say 23.
Transitions
Stock bridges between ideas make up another group of AI writing tells. “Looking ahead the” runs at 40 times the human rate, “adds another layer” at 27 times and “what comes next” at 24 times.
Contrast frames
Earlier Claude models were known for the “it’s not X, it’s Y” construction. TechCrunch notes that Opus 5.5 now avoids it, but it still says something “is more than an X, it’s a Y”. Graphite measures that frame at 98 times the human rate. “Rather than simply” sits at 32 times and “not only about” at 13 times.
Claude Opus 5.5’s AI writing tells, times the human rate (bar length relative to 116x)
All figures come from Graphite’s Opus 5.5 update. Bars are scaled to the largest value, so 98 ÷ 116 = 84.5%, 92 ÷ 116 = 79.3%, 40 ÷ 116 = 34.5%, 32 ÷ 116 = 27.6%, 27 ÷ 116 = 23.3%, 24 ÷ 116 = 20.7% and 23 ÷ 116 = 19.8%.
How Opus 5.5 Differs From Opus 5
Comparing two versions of the same model shows that AI writing tells are not fixed traits. They move with each release.
More talk of helping
Opus 5.5 uses more language about making things easier for the reader, a set of AI writing tells that Opus 5 showed far less often. “Can help you” appears eight times as often as in Opus 5, and “is especially helpful” 12 times as often.
Less emphasis
Opus 5 was the more emphatic writer. It used “is genuinely” 26 times as often as Opus 5.5, and “matters enormously” 15 times as often. In the original study, Graphite had found that “is genuinely” appeared 1,021 times in Opus 5 articles and four times in the human ones.
Less mannered prose
Anthropic uses the term “mannered prose” for writing that swaps metaphor and flourish for a direct statement, such as “a dial worth turning” where “a parameter worth varying” would do. Graphite had Opus 5 score each article from 0 to 100. Opus 5.5 scored 10.57 against 16.75 for Opus 5, a drop of 37%. Human articles averaged 6.65, so Opus 5.5 is still about 1.6 times as mannered as people, against 1.2 times for GPT-6 Astra.
The em dash nearly disappears
The em dash was the best-known of all AI writing tells. Opus 5.5 uses 0.015 em dashes per 1,000 words, against 2.92 for Opus 5, a fall of 99%. Graphite says it has “nearly stopped using them”, and now uses far fewer than human writers did before ChatGPT.
| Measure | Opus 5 | Opus 5.5 | Change |
|---|---|---|---|
| Tells found | 2,666 | 2,548 | -4% |
| Em dashes per 1,000 words | 2.92 | 0.015 | -99% |
| Mannered-prose score (human 6.65) | 16.75 | 10.57 | -37% |
| Word-distribution divergence from humans | 0.064 | 0.052 | -19% |
| Strength of 11 well-known tells | Baseline | 6% lower | -6% |
Opus 5.5 Against GPT-6 Astra
Each lab’s models have their own AI writing tells. In the original study, 65% of the tells Graphite found were unique to one model family.
Astra hedges
Astra’s AI writing tells lean towards caution. OpenAI’s model qualifies its claims far more often than Opus 5.5. Compared directly, Astra uses “may provide” 18 times as often, “not necessarily” 17 times, “not simply” 10 times and “does not establish” 275 times as often as Opus 5.5.
Opus reaches for superlatives
Opus 5.5 goes the other way. It uses “the most popular” 45 times as often as Astra, “perhaps the most” 29 times, “the most powerful” 24 times and the frame “one of the best _ about” 79 times as often.
Astra’s corrective framing
Against human writing, Astra’s signature habit is defining something by what it is not. TechCrunch reports that “not simply X” and “rather than relying on X” were more than 100 times as common in Astra’s prose as in human writing. Graphite’s September study put “rather than relying” at 187 times the human rate and the frame “the _ is not simply” at 576 times.
| Habit | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|
| Typical move | Flags importance, uses superlatives | Hedges, corrects, qualifies |
| Signature phrase | “this matters” | “does not establish” |
| Contrast style | “is more than an X, it’s a Y” | “not simply X”, “rather than relying on X” |
| Mannered prose vs humans | About 1.6x | About 1.2x |
| Word choice over time | Moving closer to human writing | Moving further away |
Why AI Writing Tells Keep Coming Back
The most useful finding in the study is about direction. Labs are removing the famous tells, but the total barely shrinks.
The famous ones fade
Graphite tracks 11 of the best-known AI writing tells, including “delve”, “not X, it’s Y” antithesis, boilerplate such as “it is important to note”, formulaic closes and sycophantic phrases. Their average strength is 6% lower in Opus 5.5 than in Opus 5, and 53% lower than in Opus 4. For OpenAI’s models it is 46% lower in Astra than in GPT-4.1, with little change after GPT-5.
The count barely moves
Claude’s total number of AI writing tells has fallen from 3,746 in Opus 4 to 3,059 in Opus 4.6, 2,666 in Opus 5 and 2,548 in Opus 5.5. That is a 32% fall over four versions, but only 4% in the latest step.
Number of AI writing tells found in each Claude Opus version (bar length relative to 3,746)
Counts are from Graphite’s Opus 5.5 update. Bars are scaled to Opus 4, so 3,059 ÷ 3,746 = 81.7%, 2,666 ÷ 3,746 = 71.2% and 2,548 ÷ 3,746 = 68.0%. The fall from Opus 4 to Opus 5.5 is (3,746 − 2,548) ÷ 3,746 = 32.0%, and from Opus 5 to Opus 5.5 it is (2,666 − 2,548) ÷ 2,666 = 4.4%.
New tells replace old ones
Tells churn between versions. In the September study, 55% to 72% of tells were unique to one of the two latest versions in each model family, and only 45% of the tells found in GPT-5.6 Sol and Astra were shared. “It’s not like the tells are decreasing,” Graphite’s chief AI officer Greg Druck told TechCrunch. “They are managing to remove the most well-known tells, but other ones pop up. And every model version has its own.”
Claude drifts closer, GPT drifts away
Graphite also compares overall word use with a statistic called Jensen-Shannon divergence, where lower means closer to human writing. Opus 5.5 fell from 0.064 to 0.052, a 19% improvement. GPT-6 Astra rose from 0.101 for GPT-5.6 Sol to 0.109, 8% further from human writing.
The labs’ own claims
Both companies say their latest models write better. TechCrunch notes that Anthropic said Opus 5.5 “communicates more naturally than prior models”, while OpenAI promised “more clarity, less jargon, [and] fewer odd turns of phrase” from GPT-6 Sol and Luna. Druck is doubtful the labs can finish the job: “These are giant models with billions of parameters. They have some finite number of tests they can run, and things slip through.”
The Limits of the Study
The findings are interesting, and they come with caveats that Graphite itself spells out.
Who ran it
Graphite, which ran the AI writing tells research, is a growth and search marketing agency rather than a university lab. Gizmodo’s coverage told readers to “take it with a grain of salt” for that reason. The researchers include a former Yahoo Research scientist and a long-serving professor, and Graphite has published the tells data and raw articles for anyone to check.
One prompt, one genre
Every model wrote web articles from the same fixed prompt. Graphite says a different prompt “would likely change the writing and the tells we identify”, and that tells may differ in other settings, such as chat. Because the human articles predate 2022, some changes in human style since then may be mistaken for AI writing tells.
A model graded the style
The mannered-prose scores were assigned by Claude Opus 5, which was not told who wrote each article. Graphite accepts that “some evaluator-specific bias is possible.”
Not a detector
Graphite is clear on this point: “we do not recommend classifying authorship based on individual tells.” A person can write “this matters” too. Accusing someone of using AI because of a phrase is unfair and unreliable, a problem we covered in our report on AI detectors.
Watch the numbers in the coverage
Two figures have drifted in retelling. The “13,000 phrases” quoted by TechCrunch and Gizmodo comes from September’s nine-model study, which found 12,877 tells across all models. Opus 5.5 alone has 2,548. And as noted above, “dependable” is 23 times the human rate, not 26.
How to Edit AI Writing Tells Out of a Draft
Most businesses now use AI somewhere in their writing. Graphite’s list of AI writing tells gives editors a useful checklist, as long as it is used to improve drafts and not to police people.
Cut the importance flags
When a draft says “this matters”, delete the line and check whether the paragraph still makes its case. If it does not, add the consequence: the cost, the deadline or the risk. Readers trust a fact more than a label.
Swap ratings for evidence
“Dependable”, “powerful” and “the most popular” assert a quality without proving it. Replace them with the figure behind the claim, such as an uptime percentage, a customer count or a test result.
Question every contrast
Contrast frames are among the easiest AI writing tells to spot. Phrases such as “is more than an X, it’s a Y” and “not simply X” often set up a claim nobody made. Say the positive point directly.
Put the person back in
Graphite found that human writing uses the first person, direct address, casual qualifiers, small anecdotes and exclamation marks far more often than AI. Human writers use exclamation marks more than 100 times as often as each current model. Sentence lengths also vary much more in human writing. A draft that sounds like someone was actually there reads as human because, after editing, it is.
Do not chase a blocklist
Because AI writing tells change with every model version, a banned-words list goes stale fast. Edit for clarity and evidence, and most tells disappear as a side effect.
| Tell | Example | Edit |
|---|---|---|
| Importance flag | “This matters because…” | State the consequence |
| Rating word | “A dependable platform” | Give the uptime or track record |
| Contrast frame | “It’s more than a tool, it’s a partner” | Say what it does |
| Stock transition | “Looking ahead, the…” | Name the date or the next step |
| Hedge stack | “May provide potential benefits” | Say what it provides, or cut it |
| Superlative | “Perhaps the most powerful” | Compare with a named alternative |
What AI Writing Tells Mean for Businesses
For companies that publish, sell or advise through text, the study has three consequences.
Readers notice
Graphite says separate research of its own shows that half of the articles published online are now AI-generated. As readers learn the patterns, unedited AI copy risks reading as generic. Brands that publish it at scale may find their content blends into everyone else’s.
Search and answer engines read it too
Content that repeats stock phrases and makes no specific claims gives search engines and AI answer engines little reason to cite it. Facts, first-hand detail and named sources are what get quoted. Our SEO services and AEO services teams help businesses publish content that earns citations.
Set a standard, not a ban
The workable policy is to use AI for drafts, require a human edit for facts and voice, and disclose where rules demand it. Our report on AI content label rules covers the EU’s transparency duties, and our AI strategy team can help write a policy that fits your business.
AI Writing Tells FAQ
What are AI writing tells?
AI writing tells are words, phrases and patterns that AI models use far more often than human writers. Graphite counts a tell when a model uses it at least twice as often as people writing on the same topics.
What is Claude Opus 5.5’s biggest tell?
The phrase “this matters”, at 116 times the human rate. Its biggest single-word tell is “dependable”, at 23 times.
Does Opus 5.5 still use em dashes?
Barely. It uses 0.015 per 1,000 words, 99% fewer than Opus 5.
Is Claude getting harder to spot?
By one measure, yes: its overall word choice is moving closer to human writing. But it still has 2,548 AI writing tells, only 4% fewer than Opus 5.
Can I use these tells to prove someone used AI?
No. Graphite itself advises against judging authorship from individual AI writing tells. People use these phrases too.
Who is Graphite?
A growth and search marketing agency that publishes research under the name Five Percent. Its data is available to download.
References and Further Reading
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
