Graphite Study Finds AI Writing Tells Persist in Frontier Models

A comprehensive study by the marketing analytics firm Graphite has identified persistent linguistic patterns that reliably distinguish artificial intelligence-generated text from human writing. Despite ongoing efforts by major AI developers to produce more natural prose, frontier language models continue to exhibit model-specific phrasing habits, revealing that current alignment techniques successfully eliminate well-known stylistic markers while simultaneously generating new ones. The research methodology involved analyzing a corpus of 10,000 human-published articles predating the release of ChatGPT as a baseline. Researchers instructed various AI models to rewrite these articles from summaries to minimize source bias. By comparing the rewritten outputs against the original human text, the team identified 13,000 specific words and phrases that appeared at least twice as frequently in AI-generated content. According to Graphite Chief AI Officer Greg Druck, the scope of detectable patterns is remarkably broad and highly adaptive. Model-specific analysis revealed distinct linguistic fingerprints. Anthropic’s Claude Opus 5.5 heavily favors the word dependable, appearing 23 times more frequently than in human samples, and frequently employs the construction more than an X it is a Y. Most notably, the model consistently emphasizes significance, utilizing the phrase this matters 116 times more often than human writers. OpenAI’s Astra model relies on hedge phrases like may provide or can provide, frequently introduces another dimension of a topic, and heavily utilizes what Graphite terms corrective framing, defining subjects as not simply X or rather than relying on X. These constructions appear over 100 times more often than in human text. Meanwhile, Google’s Gemini 3.1 Pro has largely abandoned the em-dash, a punctuation mark once ubiquitous in early AI prose. While individual models successfully suppress outdated stylistic quirks, the overall volume of AI-specific tells remains stable. Druck notes that developers are excelling at removing widely recognized markers, but these gaps are consistently filled by novel phrasing patterns unique to each model iteration. This persistence challenges industry marketing claims that newer versions communicate more naturally with greater clarity and fewer artificial turns of phrase. The discrepancy likely stems from the fundamental architecture of modern language models. With parameters measured in the billions, labs cannot comprehensively test every linguistic combination during development. Consequently, subtle statistical biases and phrase preferences inevitably surface in generation, operating beyond the scope of current editorial controls. The findings underscore a shifting landscape in AI content detection. As developers refine alignment techniques to produce smoother, more human-like prose, AI systems continue to generate identifiable linguistic signatures. Rather than signaling a failure in model training, the emergence of new tells highlights the complex interplay between large-scale probabilistic generation and stylistic refinement. For content professionals, the research provides a continuously updated framework for identifying AI authorship, suggesting that detection will rely on evolving pattern recognition rather than static keyword filtering.