New Habits in AI Texts According to Graphite Test
Research conducted by Graphite revealed that advanced AI models are abandoning old language habits while developing their own unique new patterns.
A comprehensive study by marketing firm Graphite examined the unique word and phrase preferences frequently used in texts by AI models such as Claude Opus 5.5 and OpenAI's Astra, showing that old markers like the em-dash are disappearing, but new model-driven patterns persist stubbornly.
The Change in AI Texts
With the omnipresence of content generated by artificial intelligence, people have started looking for ways to detect them.
Although early markers like the em-dash have largely disappeared, researchers note that typical habits resorted to by AI models during the writing process still continue.
Comprehensive Methodology of the Research
To perform a scaled analysis, Graphite designated 10,000 articles published before the launch of ChatGPT as the human control group.
To minimize source bias, researchers applied a method where different AI models rewrote articles based on summaries.
Preferences of the Claude Opus 5.5 Model
According to Graphite's results, one of the most prominent features of the Claude Opus 5.5 model is words used much more frequently compared to human samples.
The model prefers to emphasize by using phrases like 'this matters' at much higher rates compared to human writings.
Structure of the OpenAI Astra Model
The Astra model developed by OpenAI displays a structure with distinctly different cues and expression patterns.
This model frequently prefers using phrases indicating that certain benefits can be provided to soften claims, alongside corrective framing patterns.
Change in Punctuation Marks
All pioneering laboratories have taken steps to make various adjustments against the tendency of models to overuse the em-dash sign.
As a result of examinations, significant decreases were observed in the usage rates of this punctuation mark by models such as Opus 5.5, Astra, and Gemini 3.1 Pro.
Control Limits of Laboratories
Despite the efforts of artificial intelligence developers to create human-like writing styles, the persistence of such markers continues to surprise.
Experts draw attention to the difficulties of completely controlling massive models possessing billions of parameters.