Graphite researchers have found that leading AI models still produce thousands of recurring words, phrases, and sentence patterns at substantially higher rates than human writers, even as familiar signals such as em-dashes disappear. Its latest AI Tells study found 2,548 such patterns in Claude Opus 5.5 alone, while the broader research identified nearly 13,000 across multiple models.
Graphite defines a “tell” as a word, phrase, or structural pattern that appears at least twice as frequently in AI-generated writing as in comparable human writing. The researchers analyzed 9,974 aligned topics in the latest update, using human articles published before ChatGPT as a control and asking AI models to produce articles covering the same topics.
Different Models Develop Different Writing Habits
Claude Opus 5.5 used “dependable” 23 times more often than human writers, according to Graphite. Phrases signaling importance were even more distinctive, with “this matters” appearing 116 times as often and “why X matters” appearing 92 times as often.
The model also frequently relied on contrastive constructions such as describing something as “more than” another thing. However, its overall word distribution moved closer to human writing than its predecessor, with Graphite measuring a 19% reduction in divergence compared with Opus 5.
OpenAI’s GPT-6 Astra showed a different pattern. Graphite found it was more likely to use corrective framing, including constructions built around “rather than” and “not simply,” while its overall word distribution moved 8% further from human writing than GPT-5.6 Sol.
Em-Dashes Fade While Other Tells Remain
One widely recognized AI-writing signal has nearly disappeared from newer models. Opus 5.5 used em-dashes 99% less often than Opus 5, while Astra also used them much less frequently than human writers.
Graphite found that removing prominent habits does not eliminate model-specific patterns. Opus 5.5 still had only 4% fewer tells than Opus 5, suggesting that recognizable tendencies change between versions rather than disappearing entirely.
That finding comes as AI developers continue promoting improvements in writing quality. Anthropic says Opus 5.5 communicates more naturally than earlier versions and that testers found its writing clearer and easier to follow.
OpenAI has similarly updated its guidance for GPT-6 models to discourage stock expressions, unnecessary contrastive framing, jargon, and other recurring writing habits. Its current developer documentation specifically advises models to avoid several phrases commonly associated with AI-generated prose.
Graphite chief AI officer Greg Druck told TechCrunch that model developers appear able to reduce well-known tells, but new ones continue appearing. He said the scale and complexity of large models can make every writing tendency difficult to control.
Featured image credits: Linkedin
For more stories like it, click the +Follow button at the top of this page to follow us.