As
The research, which analyzed over 55,000 sentences and 1.2 million words, involved testing various chatbots—including ChatGPT, Claude, Gemini, and Grok—by tasking them with rewriting existing articles. The resulting AI-generated content was then compared against human-authored pieces from major news outlets and a collection of novels published over the last several decades to identify distinct patterns in structure, punctuation, and word choice.
Evolving Linguistic Patterns in AI Writing
The study found that while early iterations of chatbots were often associated with specific buzzwords like "delve" or "tapestry," newer models have shifted their vocabulary. Current systems are more likely to employ terms such as "significant," "increasingly," and "consequences," alongside more technical or scientific terminology like "methodology" and "parameter." Furthermore, AI models often exhibit a preference for nominalization, frequently converting simple verbs into longer noun forms, such as using "expansion" instead of "expand."
Structural habits also provide clues, though these are becoming less reliable as technology advances. While it was once common to identify AI through the excessive use of em-dashes, the study noted that recent updates have made this trait inconsistent across different platforms. For instance, Claude was found to use em-dashes more frequently than human writers, whereas ChatGPT used them less. Generally, AI-generated text tends to feature longer sentences, a higher frequency of the word "and," and a noticeable reduction in the use of direct quotes from experts compared to traditional journalism.
The Limitations of Detection Tools
Beyond
The researchers concluded that as AI models undergo frequent updates, the gap between machine and human prose continues to narrow. Because chatbots are becoming more sophisticated at mimicking human writing styles, the report suggests that there is no single "smoking gun" for detection. Instead, identifying AI-generated content requires looking for a combination of stylistic markers rather than relying on any individual word or punctuation habit, a task that is likely to become more difficult as the technology matures.
