Study identifies specific linguistic tells in AI-generated text from Claude and GPT models

AI Linguistic Patterns Identified

A new study by the marketing firm Graphite investigated the writing habits of frontier AI models. The research aimed to identify specific linguistic patterns, or "tells," that distinguish AI-generated prose from human writing. Graphite defined a tell as a phrase appearing at least twice as often in AI content compared to human content.

Diverging Trends in Model Development

The study found that Claude models are increasingly aligning with human word distribution over time. Conversely, GPT models are moving further away from human linguistic patterns. This indicates different developmental trajectories in how these AI systems generate text.

Methodology for Analysis

To conduct the study, Graphite used a control group of 10,000 articles published before ChatGPT's release. AI models then rewrote these articles from summaries to minimize source bias. This method allowed for a direct comparison of word and phrase frequency, as well as broader sentence construction patterns, between human and AI-generated text.

Specific Tells for Claude Opus 5.5

Graphite's findings indicate that Claude Opus 5.5 frequently uses the word "dependable," appearing 23 times more often than in human samples. Opus 5.5 also shows a strong preference for phrases like "this matters," which occurs 116 times more often, and "why X matters," used 92 times more often than in human writing.

Specific Tells for OpenAI's Astra

OpenAI's Astra model exhibits different linguistic tells. It often uses the phrase "another dimension" and tends to qualify claims with phrases such as "may provide" or "can provide." Astra's most prominent tell is the "corrective framing," where topics are defined as "not simply X" or presented as an alternative, "rather than relying on X," with these constructions appearing over 100 times more frequently than in human prose.