The Persistence of AI Linguistic Markers
As large language model-generated prose becomes ubiquitous, researchers are identifying specific habits that distinguish AI output from human writing. While early indicators like excessive em-dash usage have largely been addressed by model developers, new linguistic tells continue to emerge.
A study by the marketing firm Graphite analyzed the writing habits of frontier models, identifying 13,000 phrases that appear at least twice as frequently in AI-generated content compared to human-authored samples. The research suggests that while models are evolving, they maintain unique, model-specific linguistic quirks.
The study indicates divergent trends in model development. According to Graphite, Claude models are trending closer to human word distribution patterns over time, whereas GPT-series models appear to be moving further away.
Methodology and Model-Specific Tells
To isolate these patterns, researchers used a control group of 10,000 articles published prior to the release of ChatGPT. They tasked various AI models with rewriting these articles from summaries to minimize source bias, allowing for a direct comparison of sentence construction and vocabulary frequency.
Claude Opus 5.5 shows a distinct preference for the word “dependable,” which appears 23 times more frequently than in human samples. The model also frequently employs the construction “is more than an X, it’s a Y.”
Opus 5.5 also exhibits a tendency to emphasize importance, using the phrase “this matters” 116 times more often than human writers, and “why X matters” 92 times more often.
OpenAI’s Astra model displays different markers, such as referencing “another dimension” of a topic and hedging claims with phrases like “may provide” or “can provide.” It also frequently uses “corrective framing,” such as defining a topic as “not simply X” or “rather than relying on X,” constructions found to be over 100 times more common in Astra’s output than in human writing.
The Evolution of AI Writing Styles
Frontier labs have successfully reduced the use of em-dashes across their models. Opus 5.5, Astra, and Gemini 3.1 Pro have all significantly curtailed their use of this punctuation mark compared to earlier versions or human benchmarks.
Despite these adjustments, the overall volume of linguistic tells remains stable. As developers eliminate known markers, new ones emerge, suggesting that each model version retains its own set of identifiable habits.
The persistence of these patterns is notable given the focus labs place on human-like communication. Anthropic, for instance, marketed Opus 5.5 as having a more natural, clearer writing style than its predecessors.
OpenAI has made similar claims regarding the GPT-6 versions, Sol and Luna, promising increased clarity and reduced jargon.
Challenges in Model Control
Experts suggest there are inherent limitations to how effectively labs can curate the output of large-scale models. The complexity of these systems, which contain billions of parameters, makes it difficult to fully eliminate specific linguistic tendencies.
The sheer scale of these models means that despite rigorous testing, certain patterns and quirks inevitably persist in the final output.
