AI writing tells: Opus 5.5 says 'this matters' 116x too often
Graphite's AI writing tells study finds Claude Opus 5.5 dropped the em-dash but overuses 'this matters' 116x. Every model has new tells. Edit for patterns, not words.
Marketing firm Graphite updated its AI writing tells study this week. The headline: Claude Opus 5.5 nearly stopped using the em-dash, then picked up new habits. Per TechCrunch, Opus 5.5 writes "this matters" 116 times more often than humans. Full disclosure: our post template has a heading called "Why it matters for your business." We read the study with some interest.
What actually happened
Graphite's original study, published September 16, compared 10,000 human articles from before ChatGPT's launch with 90,000 AI rewrites of the same topics across nine models. It found nearly 13,000 words and phrases that appear at least twice as often in AI text. Each model had between 2,355 and 3,746 of them. 65% were unique to one model family.
The update, reported by VentureBeat on September 30, adds Opus 5.5:
- 2,548 tells for Opus 5.5
- Em-dash: 0.015 per 1,000 words, down from 2.92 in Opus 5, a 99% drop
- "This matters": 116x the human rate. "Why X matters": 92x. "Dependable": 23x (per TechCrunch)
- OpenAI's GPT-6 Astra: corrective framing like "not simply X" at more than 100x, and em-dashes 88% below the human rate
The pattern across versions: well-known tells fell 41% to 86% from the earliest to the latest model in each family. New ones took their place. GPT-5.6 Sol and GPT-6 Astra share only 45% of their tells. "Every model version has its own," Graphite's chief AI officer Greg Druck told TechCrunch.
Why AI writing tells matter for your business
A banned-word list goes stale every model release. Teams that grep drafts for "delve" and em-dashes are fighting the last war. The vendor already fixed those. When you switch models, the tells switch with you.
Readers notice patterns, not words. "Why X matters," "not just X, it's Y," three-item lists in every paragraph. Edit for structure. If every section opens by telling the reader something is important, cut the claim and show the number.
Your own voice is the fix. The tells are what a model writes when it has nothing specific to say. Feed it your prices, your job counts, your customer's actual words. Specific facts crowd out filler.
Key takeaways
- Graphite found nearly 13,000 AI writing tells across nine models; 65% are unique to one family
- Claude Opus 5.5 cut em-dashes 99% but has 2,548 tells, led by "this matters" at 116x
- Old tells drop 41–86% per model family, then new ones appear
- Static banned-word lists go stale with every model update
- Edit for repeated structures and replace filler with your own specifics
Your content pipeline sounds like everyone else's? We build AI drafting workflows that pull from your real data and run a per-model style check before anything ships. See what we build or show us a draft.
Sources: Graphite, VentureBeat, TechCrunch.
- #ai-writing
- #content-marketing
- #claude
- #seo
- #editing
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
OpenAI and America's SBDC: free AI training for small business
OpenAI and America's SBDC will train SBDC advisors and run AI workshops for 1,000+ small businesses, with free ChatGPT Plus. Bring one real workflow.
Read itBain: AI needs $6T a year by 2031. Guess who pays
Bain's Global Technology Report says AI must earn $6 trillion annually by 2031, with $4.2 trillion of it uncreated. Price your vendor contracts for that gap.
Read it