A new study from the marketing firm Graphite shows that despite efforts to make large language models (LLMs) sound more natural, frontier models continue to rely on specific linguistic "tells" that distinguish their prose from human writing.
AI writing "tells" persist in frontier models despite efforts to sound more human, study finds
The researchers identified 13,000 phrases that appear at least twice as frequently in AI-generated content compared to human writing. While well-known markers like the frequent use of em-dashes (—) are being successfully reduced—with Gemini 3.1 Pro almost entirely eliminating them—other unique patterns are emerging or persisting.
Graphite's analysis found that Anthropic's Claude models are becoming closer to human word distribution over time. Claude Opus 5.5, for example, uses em-dashes 99% less often than its predecessor, Opus 5. However, Opus 5.5 still exhibits specific habits, such as using the word "dependable" 23 times more often than human samples and frequently employing phrases like "this matters" to emphasize importance.
In contrast, OpenAI's Astra models appear to be moving further away from human-like word distribution. Astra's writing is characterized by "corrective framing"—defined as describing a topic as "not simply X" but "rather than relying on X"—which occurs more than 100 times more frequently in Astra's prose than in human writing.
Researchers suggest that as developers successfully eliminate the most obvious tells, new patterns emerge. Graphite's chief AI officer, Greg Druck, noted that because these are massive models with billions of parameters, it is difficult for labs to completely control all linguistic idiosyncrasies.
Sources
- Opus 5.5 loves to tell you ‘this matters’ (and other AI writing tells) （TechCrunch AI, 2026-10-01）
- Graphite