Graphite counted 12,877 words, phrases and recurring word-pattern frames that appeared at least twice as often in AI-generated articles as in its pre-ChatGPT human comparison set. In an update focused on Claude Opus 5.5, the phrase “this matters” appeared 116 times more often than in that comparison. Those are corpus-level frequency results—not proof that any particular sentence or writer used AI.
What Graphite means by an AI writing tell
A “tell” in Graphite’s report is a feature that appeared in generated articles at least twice as often as in the human comparison, after the researchers normalized counts for text length and applied minimum occurrence thresholds. Features included individual words, two- and three-word phrases, and “frames”: word patterns with a gap of up to three less-common words.
The label describes a statistical difference between two collections of text. It is not a verdict about authorship, and Graphite says it chose an interpretable method for identifying patterns rather than building the most accurate AI detector. Graphite’s original report
How the study compared human and AI articles
For its September 2026 report, Graphite used 10,000 articles collected from Common Crawl and published before ChatGPT launched on November 30, 2022. It summarized each article with GPT-4.1, then asked nine models to produce an article from that summary. The common comparison covered 9,984 matched topics, with a human article and AI-generated article for each topic.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
The original model set was GPT-4.1, GPT-5, GPT-5.6 Sol, GPT-6 Astra, Claude Opus 4, Claude Opus 4.6, Claude Opus 5, Gemini 2.5 Pro and Gemini 3.1 Pro. The setup makes the texts comparable by topic, but it also means the findings reflect a particular prompt, summarization step, source collection and article genre—not all writing by people or all output from those models.
What the reported counts show
Nearly 13,000 features across nine models
Graphite reported 12,877 unique tells across its nine-model study. Counts for individual models ranged from 2,355 to 3,746. Its combined total for GPT-6 Astra, Claude Opus 5 and Gemini 3.1 Pro included 7,043 tells. These figures count features that met the report’s frequency rules; they are not a checklist of 12,877 universally reliable AI clues.
Most tells were not universal
Graphite found that 65% of the tells were unique to one model family. A phrase associated with one family therefore should not automatically be treated as a signature of every chatbot. The report also found that tell profiles changed across model versions. Graphite’s original report
“This matters” in the Opus 5.5 update
In its October 2026 update, Graphite said Claude Opus 5.5 used “this matters” at 116 times the rate in the human comparison; the frame “why _ matters” appeared at 92 times that rate. Graphite counted 2,548 tells for Opus 5.5, four percent fewer than for Opus 5, using its original method. Graphite’s Opus 5.5 update
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallWhy fewer tells do not necessarily mean more human-like writing
A tell count and an overall comparison of word distributions measure different things. Graphite’s original report said the combined frequency of nine well-known features fell between its earliest and latest tested models by 41% to 86%, depending on model family. Across tested versions, it counted 29% fewer total tells for Claude and 32% fewer for Gemini, while GPT’s total rose 48%. Yet Graphite reported that Claude’s overall word distribution became more similar to the human comparison, while GPT’s and Gemini’s diverged in the versions tested. Graphite’s original report
The Opus 5.5 update illustrates the distinction: its total tell count was four percent lower than Opus 5’s, while its overall word distribution was closer to the human comparison. Separately, Graphite’s measure of well-known tells was six percent lower for Opus 5.5 than Opus 5 and 53% lower than Opus 4. These measures should not be collapsed into one claim that a model is simply “more human.” Graphite’s Opus 5.5 update
What these findings can—and cannot—say about a passage
A corpus-level rate does not establish who wrote an individual sentence. “This matters” is ordinary language; encountering it in a draft is not evidence that a person used Claude, another model or AI at all. The phrase’s reported ratio compares how often it occurred across Graphite’s collected texts, not the probability that any one text containing it was generated by AI.
Other studies also caution against treating linguistic signals as proof. A 2026 study of 512,970 psychology abstracts across 975 journals estimated that at least 16% of 2025 abstracts showed linguistic traces consistent with LLM editing. Its authors describe those markers as indirect rather than definitive evidence of AI use; this is a separate study, not a Graphite result. Botes et al., “Finding the Fingerprints of Generative Artificial Intelligence in Psychology Publications”
Authorship attribution is also distinct from identifying which generator produced text. Penn State’s account of attribution research discusses those as different tasks: deciding whether writing is machine- or human-produced is not the same as identifying a particular model. Penn State’s overview of AI-content detection research
Rank #4
Why the baseline and prompt matter
- Different publication eras: the human articles predate ChatGPT, while the generated articles came from newer models. Changes in web-writing conventions over time can affect the comparison.
- Different text-generation process: researchers first summarized each human article with GPT-4.1, then used a fixed general-writing prompt to produce AI articles. Remaining boilerplate or effects of the prompt may contribute to apparent differences.
- Limited genre: the study concerns web articles. Its counts should not be assumed to describe emails, fiction, school assignments, technical documentation or every other kind of writing.
- Changing systems: new model versions and different prompts can change the patterns a corpus study would find. A tell list is therefore time- and setup-dependent.
How to use the findings responsibly
For readers, editors and educators, the useful conclusion is that generated writing can have recurring statistical habits, but no single phrase is a dependable authorship test. Treat a pattern as a reason to look for more context—not as a basis for accusing a writer. Stronger judgments require evidence beyond a handful of word choices, and tools that flag patterns should not be presented as proving who wrote a text.
Graphite chief AI officer Greg Druck told TechCrunch that Claude models were “getting closer to the human word distribution over time,” while GPT models were moving further away in the versions studied. He also said models can remove familiar tells while new ones emerge, with each version developing its own patterns. These observations refer to Graphite’s comparisons, not a general rule for every model or every writing task. TechCrunch’s report on the Opus 5.5 findings
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
