AI Writing Tells Shift from Punctuation to Phrases as Models Adapt
A new study by Graphite identifies approximately 13,000 words and phrases that distinguish AI-generated text from human writing. While older markers like em dashes have declined significantly in Anthropic's Claude Opus 5.5, new linguistic patterns have emerged. The phrase 'this matters' is now a primary indicator for Claude, while OpenAI's Astra model favors the phrase 'does not establish.' Although some models are becoming more human-like, others are developing new, distinct stylistic quirks.
Key points
- Graphite's analysis of 10,000 pre-ChatGPT texts and their AI rewrites found 13,000 distinct markers of AI authorship.
- Anthropic's Claude Opus 5.5 uses the phrase 'this matters' 116 times more frequently than human writers, replacing older tells like em dashes.
- OpenAI's Astra model frequently uses the phrase 'does not establish,' appearing 275 times more often in its output than in Claude's.
- Claude Opus 5.5 has reduced em dash usage by 99% compared to its predecessor, Opus 5, but has increased the use of hedges like 'perhaps' by 97%.
- OpenAI's Astra model is trending toward less human-like writing, while Claude is moving closer to human stylistic norms.
Background
This development follows a broader trend in 2026 where AI models are increasingly integrated into professional and creative workflows, as seen in Anthropic's recent scientific discoveries and the U.S. government's push for AI acceleration. Earlier concerns about AI detection focused on obvious punctuation errors, but the rapid evolution of models has necessitated more sophisticated linguistic analysis to distinguish machine from human output.
How outlets are covering it
Gizmodo and Mezha both highlight the shift from punctuation-based tells to phrase-based ones, noting that Claude Opus 5.5 is becoming more human-like while OpenAI's Astra is not. HackerNoon provides a more granular analysis, focusing on sentence length and vocabulary, noting that Claude Opus 5.5 uses 17% shorter sentences and 95% fewer em dashes than its predecessor. All sources agree that while some tells are disappearing, new ones are emerging, making detection a continuous challenge.
Why it matters
As AI-generated content becomes more prevalent, the ability to distinguish it from human writing is crucial for maintaining trust in information. The shift from obvious punctuation errors to subtle phrase-based tells means that AI detectors must evolve rapidly to keep up with the changing landscape of AI writing. This has implications for content moderation, academic integrity, and the overall reliability of online information.
What to watch
Researchers and AI developers will likely continue to refine their models to reduce these tells, while AI detection tools will need to adapt to the new linguistic patterns. The ongoing cat-and-mouse game between AI developers and detection tools will likely lead to more sophisticated methods for both generating and identifying AI-written content.
- ‘This Matters’: Researchers Identify Thousands of New Tells in AI Writing Gizmodo
- Opus 5.5 loves to tell you ‘this matters’ (and other AI writing tells) TechCrunch
- Claude Opus 5.5 uses em-dashes 99% less often and sounds more human — but still exhibits 2,548 AI writing tells VentureBeat
- Claude Opus 5.5 Writes With 17% Shorter Sentences: How It Reads More Human HackerNoon
- Researchers have identified new 'filler words' that can be used to identify text generated by AI Mezha
Want the full story? Read the original reporting
Read on Gizmodo