The words AI overuses, measured
Plenty of pages list "words that sound like ChatGPT." Almost none of them tell you where the list came from, and most were themselves written by ChatGPT. This one is different in two ways: every multiplier below traces to a named study, and the trend column updates as the research does, because these lists rot fast. A word that screamed AI in 2024 may be gone from new model output in 2026 while still flooding the academic record.
The headline words
| Word | Measured excess | Direction in 2026 |
|---|---|---|
| delve / delves | 28x expected rate in 2024 abstracts (Kobak); 14x in indexed papers (Shapira) | Fading from new model output; rising in human speech |
| underscore(s) | ~14x in abstracts; used in ~30% of 2025 PMC papers vs ~3% in 2022 | Still climbing in print |
| meticulous | 34.7x in AI-era peer reviews (Liang) | Rising |
| showcasing | 10.7x (Kobak); 20x (GPTZero) | Stable |
| intricate | +611% occurrences per million, 2020 to 2024 (Juzek & Ward) | Plateauing in text, rising in speech |
| tapestry | On every editor blacklist | Tuned out of new models |
| pivotal, crucial | Among the largest common-word excesses in 2024 abstracts | Stable |
| boasts, garner | Post-2022 rise measured in podcasts and talks (Max Planck; FSU) | Rising in human speech |
| leverage, seamless, robust, transformative, holistic, myriad | Consensus blacklist across detectors and editor guides | Stable |
The phrases are worse than the words
Single words at 10x or 30x human rates are impressive. Whole phrases go further. By GPTZero's corpus of millions of documents, "plays a significant role in shaping" appears 182 times more often in AI text than human text, "notable works include" 120 times, and "today's fast-paced world" 107 times. Add the editor-guide classics: "stands as a testament to," "it's important to note," "in today's ever-evolving landscape," "let's dive in," and the summary twins "in conclusion" and "in summary."
The strange part: the words are becoming human
Here's the finding that complicates every list like this one. Researchers at the Max Planck Institute analyzed 737,083 hours of podcasts and YouTube talks and found people now say "delve," "boast," and "meticulous" significantly more than before ChatGPT launched. A Florida State study of 22 million words of unscripted conversation confirmed it: nearly three quarters of the tracked AI-favored words rose in human speech, while their synonyms didn't. The vocabulary is leaking out of the models and into us. Which is exactly why word lists alone can't identify AI text anymore, and why the structural signs matter more every month.
What to write instead
The boring answer is the correct one: the plain word. Use, not leverage. Careful, not meticulous. Show, not showcase. Key, not pivotal. Many, not myriad. New, not cutting-edge. If a sentence needs "underscores the importance of," the sentence usually needs deleting, not editing.
Check a draft against the live list. Paste your text and see every flagged word with a plain swap suggested. Runs in your browser; your text is never uploaded anywhere.
Check your draft freeSources: Kobak et al., Science Advances 2025 · Juzek & Ward, COLING 2025 · Liang et al., ICML 2024 · Shapira, "Delving into delve" · GPTZero AI vocabulary · Yakura et al., Max Planck · FSU speech study · Scientometrics 2026
Part of HumanSounding, the live index of AI writing tells. Figures quoted from their sources without adjustment; the trend data behind this site updates weekly.