Claude models are moving closer to human word distribution over time while GPT models are drifting further away, according to a new study from marketing firm Graphite that analyzed the writing habits of frontier AI models.
The research examined a corpus of 10,000 articles published before AI-generated content became widespread, then compared those human writing patterns against output from leading language models. Graphite's chief AI officer Greg Druck said the findings reveal a clear divergence between model families.
"It turns out that Claude models are actually getting closer to the human word distribution over time," Druck said. "And for the GPT models, it's getting further away."
The study identified specific phrases that have become signatures of AI-generated text, including constructions like "it's not X, it's Y" and declarations that something "matters" — linguistic tics that human writers use far less frequently. These patterns give away AI authorship even when the content itself is accurate.
Graphite's methodology required what the firm described as careful study design to measure AI writing at scale. By establishing a baseline from pre-AI human articles, researchers could quantify exactly how much each model's vocabulary deviates from natural human expression.
The findings carry implications for anyone using AI to produce content that needs to pass as human-written — from marketing copy to news articles to academic work. As detection tools improve and audiences grow more attuned to AI tells, the gap between model families could influence which systems businesses choose for writing tasks.
Anthropic's Claude models have emphasized constitutional AI training methods that prioritize helpful, harmless, and honest outputs. OpenAI's GPT series, meanwhile, has focused on scaling and capability improvements. The Graphite data suggests those different philosophical approaches may produce measurably different writing styles.
The corpus of 10,000 pre-AI articles provided a control group representing authentic human prose before large language models influenced how people write. Comparing that baseline to current model outputs shows not just which words AI favors, but how those preferences have evolved across model generations.
Druck's assessment that Claude is converging toward human patterns while GPT diverges represents one of the first quantitative comparisons of writing style between major AI model families. The research adds to growing evidence that different training methodologies produce distinct linguistic fingerprints.
TechCrunch Disrupt, where AI writing tools are frequently showcased, will host its final demo day for exhibitors on Oct. 2, with ticket discounts ending Sept. 25 at 11:59 p.m. PT.
