In March 2025, Dmitry Kobak and colleagues published a landmark study in Science Advances: "Delving into ChatGPT usage in academic writing through excess vocabulary." They analyzed millions of academic abstracts published before and after ChatGPT's November 2022 launch and identified specific words that spiked dramatically in frequency — words that were statistically normal before and suddenly overrepresented after.
These aren't guesses or heuristics. They're empirically measured vocabulary shifts. This is the strongest research basis we have for AI text detection.
Vocabulary Patterns
The research identified a set of words that spiked significantly in published writing after ChatGPT's launch. These are words that were already present in academic writing, but suddenly appeared at much higher rates — strong evidence that AI models trained on formal text default to these vocabulary choices.
The pattern is consistent: words that sound formal, abstract, and "polished" — the kind of vocabulary that makes writing sound authoritative — are dramatically overrepresented in AI output. Human writers use these words occasionally; AI uses them constantly.
Rather than listing specific words (which would help people game detection systems), our free detector automatically screens for these patterns. If your text scores high on our vocabulary signal, that's a strong indicator.
Structural Patterns Beyond Vocabulary
Vocabulary is the most measurable signal, but ChatGPT has other structural tells:
The "In today's..." opener
ChatGPT frequently opens essays and articles with "In today's rapidly evolving...", "In the modern world...", "In an era of unprecedented...". This grand contextualizing opener is a ChatGPT default. Human writers are more likely to start with a specific example or question.
The balanced both-sides structure
"While some argue X, others contend Y. Ultimately, the answer lies in understanding the nuanced relationship between..." — this framing appears in a huge percentage of ChatGPT responses to any opinion question. Human writers take sides.
The three-item list
ChatGPT defaults to presenting exactly three points, three reasons, or three examples. Always three. Human writers might give two, or five, or trail off after the first good one.
The transition overload
ChatGPT uses formal academic transition phrases at a significantly higher rate than human writers. These are the connectors of formal academic writing — ChatGPT learned them from training data and defaults to them even in casual contexts. Human writers stick to simpler connectors.