AI
AI models still betray themselves through writing quirks, study finds
Researchers catalogued thousands of phrases and sentence patterns that expose AI authorship—and show each model has its own distinctive fingerprint.
TechCrunch — AI
As language models flood the internet with generated text, identifying machine-written prose has become a cat-and-mouse game. Early red flags—like overused em-dashes and the word "delve"—have largely disappeared as labs refined their systems. Yet new research from the marketing firm Graphite reveals that AI continues to leave behind thousands of linguistic traces that distinguish it from human writing.
The breadth of AI fingerprints#
Graphite's analysis uncovered a startling finding: roughly 13,000 distinct phrases appear at least twice as frequently in AI-generated content compared to human writing. Rather than relying on a single smoking gun, each frontier model exhibits its own constellation of quirks that collectively paint an unmistakable picture of machine authorship.
The research team established a rigorous methodology by assembling a baseline of 10,000 articles published before ChatGPT's launch. They then prompted various models to rewrite these pieces based on summaries, minimizing potential bias from source material. By comparing word choice and syntactic patterns across these matched pairs, they quantified how distinctly each system "speaks."
Each model's telltale habits#
Anthropic's Claude Opus 5.5 betrays itself through affection for the word "dependable"—appearing roughly 23 times more frequently than in comparable human samples. More conspicuously, the model peppers text with appeals to significance: "this matters" surfaces 116 times more often, while "why X matters" occurs 92 times above baseline.
OpenAI's Astra exhibits different markers. It gravitates toward phrases like "another dimension" and softens claims through constructions such as "may provide" or "can provide." Its most distinctive pattern is what researchers term "corrective framing," where statements take the form "not simply X" or "rather than relying on X"—these appear over 100 times more frequently in Astra's output.
The Great Em-Dash Purge#
Interestingly, all major labs appear to have heard the criticism about em-dash abuse and responded dramatically. Opus 5.5 has cut em-dash usage by 99 percent compared to its predecessor. Astra has reduced the mark by 88 percent below human baselines, while Gemini 3.1 Pro has nearly eradicated it entirely.
