AI Texts Use 13 000 Expressions at Least 2 Times as Often as Human Writing

A new study by the marketing company Graphite compared the writing styles of modern AI models. It identified 13 000 expressions used at least 2 times as often in AI-generated texts as in human writing. Signs that once gave AI away, such as the em dash, have become less common, but other recurring patterns have taken their place. Each model version uses its own distinctive vocabulary and sentence structures.

The researchers selected 10 000 articles published before ChatGPT was made available to the public as samples of human writing. Various AI models were then given brief summaries and asked to rewrite those articles. This reduced the influence of the original text’s style on the results. The researchers then compared which words, expressions, and sentence structures were used more frequently in each group.

The word that most often gave away the Claude Opus 5.5 model was “dependable.” This word, meaning “reliable,” appeared 23 times as often in the model’s texts as in human writing. The model has largely moved away from the pattern of contrasting ideas as “not X, but Y,” but it continues to favor similar structures. The expression “This matters” appeared 116 times as often as in human texts, while the “why X matters” structure appeared 92 times as often. According to the study, the vocabulary choices of Claude models are becoming closer to human writing over time.

OpenAI’s Astra model frequently adds “another dimension,” meaning “another aspect,” to a topic. When discussing benefits, it chooses cautious expressions such as “may provide” or “can provide.” The model generally frames an idea through corrective comparisons such as “not simply X” and “rather than relying on X.” These expressions were used more than 100 times as often as in human writing. The overall vocabulary choices of GPT models, however, have moved further away from human language over time.

AI labs have responded to criticism about the excessive use of em dashes. Claude Opus 5.5 used this punctuation mark 99% less often than the previous Claude Opus 5 model. In Astra, em dashes appear 88% less often than in human texts. Gemini 3.1 Pro has almost completely eliminated the punctuation mark. However, as familiar signs disappear, new vocabulary and sentence patterns emerge. As a result, the overall number of indicators that give AI away is not declining significantly.

When Anthropic introduced the Opus 5.5 model, it said the model communicated more naturally than previous versions. According to the company, the model’s writing is clearer and easier to read. OpenAI also promised that the GPT-6 Sol and GPT-6 Luna models would produce clearer texts with less jargon. The study’s authors believe it is difficult to identify and eliminate every language habit in models consisting of billions of parameters. Therefore, when one pattern is removed, another may take its place.