A recent study by marketing firm Graphite analyzed the writing patterns of cutting-edge AI language models, focusing especially on Anthropic’s Opus 5.5. Researchers compared AI-generated text to human writing by having models rewrite thousands of pre-existing articles from summaries, reducing source bias. Their findings reveal that despite efforts to mimic human prose and weed out obvious giveaways like overused em-dashes, AI still exhibits distinct linguistic habits, with over 13,000 phrases appearing at least twice as frequently in AI text compared to human samples.

Opus 5.5’s most telling habits include an unusually frequent use of the word “dependable,” appearing 23 times more often than in human writing, as well as phrases emphasizing importance such as “this matters” and “why X matters,” which occur dramatically more in its output. Other models show different traits; for example, OpenAI’s Astra model commonly uses “corrective framing,” stating topics as “not simply X,” and frequently hedges claims with phrases like “may provide.” Interestingly, most leading models have greatly reduced em-dash usage, responding to criticism that earlier AI relied too heavily on this punctuation.

Despite continuous improvements and claims from companies like Anthropic and OpenAI about making AI writing more natural and clearer, these linguistic “tells” persist. Greg Druck, Graphite’s chief AI officer, expressed skepticism about whether labs can ever fully eliminate such patterns given the complexity and scale of large language models. Although models can remove familiar markers, new unique tells often emerge with each iteration, reflecting the inherent challenges in perfecting human-like text generation.

This research highlights that even state-of-the-art AI still leaves detectable fingerprints in its writing style, which has implications for detecting AI-generated content and for ongoing development in the field. As AI writing becomes more widespread, understanding these nuances could help both developers and users discern AI contributions from human text, amidst evolving efforts to create more seamless, natural language models.