Anthropic's Claude models are producing writing that increasingly resembles human text, while OpenAI's GPT models are drifting further from natural human word patterns, according to a new study from the marketing firm Graphite.
The research examined the writing habits of frontier AI models to identify each system's most frequently used words and phrases. Graphite's chief AI officer Greg Druck summarized the divergent trajectories: "It turns out that Claude models are actually getting closer to the human word distribution over time. And for the GPT models, it's getting further away."
To establish a human baseline, Graphite assembled a corpus of 10,000 articles published before 2022, predating the widespread release of ChatGPT and other generative AI tools. Comparing AI outputs against this pre-AI writing sample allowed researchers to measure how far each model's vocabulary and sentence patterns deviate from authentic human prose.
The study's methodology matters because it isolates genuine stylistic fingerprints rather than superficial word frequency. By using only pre-2022 text, researchers avoided contaminating their human baseline with AI-generated content that has flooded the web since late 2022.
Graphite's findings point to a recognizable pattern in AI writing: the "it's not X, it's Y" construction, which the study identified as a common tell across frontier models. This parallel structure appears frequently in AI-generated marketing copy, essays, and social media posts, often deployed to create an illusion of insight without adding substantive information.
The practical consequence for readers and editors is that AI writing detection remains possible through vocabulary analysis, but the reliability of such methods depends heavily on which model produced the text. Claude's convergence toward human patterns suggests its outputs will become harder to flag through word-distribution analysis alone.
Druck's assessment implies that model developers are making different choices about training data and fine-tuning objectives. Anthropic's apparent emphasis on natural language patterns contrasts with OpenAI's approach, which may prioritize other capabilities at the expense of stylistic authenticity.
Graphite's research adds to a growing body of work attempting to characterize AI-generated text at scale. The firm's decision to publish its methodology and baseline corpus selection criteria provides other researchers a foundation for replicating or challenging the results.
The study did not specify which versions of Claude or GPT were tested, nor did it disclose the full list of telltale words and phrases identified for each model. Those details would help writers and editors recognize model-specific patterns in submitted work.
