What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
In Graphite’s October 1, 2026 analysis, Claude Opus 5.5 used the phrase “this matters” at 116 times the rate found in its matched human corpus. The frame “why something matters” appeared at 92 times the human rate. Those are corpus-level frequency ratios—not proof that a passage containing either phrase was generated by AI.
What the “116 times” figure actually measures
Graphite defines a “tell” as a word, phrase or grammatical frame that appears at least twice as often in model output as in the human comparison, after length normalization, frequency filters and feature-specific minimum-occurrence requirements. A frame is a pattern that can include gaps of up to three words.
Therefore, “this matters” appearing 116 times the human rate means its normalized frequency in the Opus 5.5 sample was 116 times the normalized frequency in Graphite’s human sample. It does not mean 116% of Opus writing contains the phrase, nor that one use identifies an AI author.
Other Opus 5.5 patterns Graphite found
| Word, phrase or frame | Rate versus human corpus | What it signals stylistically |
|---|---|---|
| “this matters” | 116 times | Explicit importance-signalling |
| “why _ matters” | 92 times | Explanations organized around significance |
| “rather than simply” | 32 times | Contrastive, corrective framing |
| “what comes next” | 24 times | Forward-looking transitions |
| “dependable” | 23 times | Evaluative reassurance |
These are recurring tendencies in a large sample. They can help an editor notice formulaic prose, but none is a standalone authorship test.
Recommended Free Tools
#1 Best Overall
How Graphite built the comparison
The Opus 5.5 update reused Graphite’s earlier approach across 9,974 aligned topics. For each topic, the dataset paired one human article with an article generated by each model. The human articles predate ChatGPT, so the baseline is matched by subject matter but is not a representative sample of writing published today.
Graphite’s original study covered 10,000 human articles and 90,000 AI-generated articles from nine models. It identified nearly 13,000 qualifying tells. In that study, 65% of tells were unique to one model family, a result that argues against a universal checklist.
Opus 5.5 versus Opus 5: three different ways to compare them
Graphite reports several statistics that answer different questions. They should not be collapsed into one score for “human-likeness.”
| Measure | Opus 5 | Opus 5.5 | How to read it |
|---|---|---|---|
| Qualifying words, phrases and frames | 2,666 | 2,548 | Count of features meeting Graphite’s tell criteria |
| Word-distribution divergence | Baseline | 19% lower than Opus 5 | Overall difference in word-use distribution from the human corpus |
| Em dashes per 1,000 words | 2.92 | 0.015 | Selected punctuation tendency; Opus 5.5 was 99% lower |
| Mannered-prose score | 16.75 | 10.57 | Separate 0–100 rating; evaluated by Opus 5 on a 1,000-topic subsample |
A lower divergence or mannered-prose score does not mean every sentence is more natural, and a lower tell count does not mean the remaining tells are unimportant. Opus 5.5 can move closer to the human word distribution while still overusing particular constructions.
Rank #3
Why familiar AI clues keep changing
Graphite Chief AI Officer Gregory Druck told TechCrunch that Claude models were “actually getting closer to the human word distribution over time.” He also said, “It’s not like the tells are decreasing. They are managing to remove the most well-known tells, but other ones pop up. And every model version has its own.” Those statements describe Graphite’s interpretation of its measurements, not an independently replicated general law.
The em-dash result illustrates the problem. Opus 5 used 2.92 em dashes per 1,000 words in Graphite’s sample; Opus 5.5 used 0.015. A rule such as “many em dashes means AI” would therefore age badly, even before accounting for human writers who favor that punctuation.
Rank #4
What these patterns can—and cannot—tell an editor
Useful signals
- Repeated importance language can reveal vague claims that announce significance instead of demonstrating it.
- Predictable contrast and transition frames can make paragraphs sound templated when they appear in nearly every section.
- Comparing several passages from the same author can expose abrupt shifts in diction, rhythm or structure for human review.
Unsafe conclusions
- One phrase is not evidence that a passage came from Opus 5.5 or any other model.
- High frequency in Graphite’s corpus does not establish that all AI systems use the same phrase.
- A detector score or stylistic impression cannot substitute for provenance, drafts, source notes or a conversation with the writer.
A practical review method for suspected AI prose
- Read for substance first. Check whether claims are specific, sourced and responsive before looking for stylistic clues.
- Mark repeated frames. Highlight recurring constructions such as importance announcements, “rather than simply” contrasts or identical transition patterns.
- Check the whole sample. A single sentence has little evidentiary value; repeated patterns across a substantial body of writing are more informative as an editing signal.
- Compare with the author’s verified work. Look for changes in vocabulary, sentence length, punctuation and level of specificity, while allowing for legitimate assignment differences.
- Ask for process evidence when stakes are high. Draft history, notes, citations and revision records are stronger evidence of authorship than a phrase checklist.
- Edit the prose directly. Replace abstract importance claims with concrete evidence, remove redundant transitions and vary sentence openings where the writing genuinely sounds mechanical.
What remains uncertain
Graphite is both the publisher of this analysis and, as TechCrunch describes it, a growth-marketing agency. The Opus 5.5 update says the original study’s limitations apply, but it does not reproduce all of those details. The results have not been independently established here, and the pre-ChatGPT human corpus limits how confidently they can be generalized to present-day human writing.
The safest interpretation is narrow: Graphite found model- and version-specific overuse patterns in a controlled corpus. Those patterns are useful prompts for careful editing, not forensic proof of who wrote an individual passage.
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.
