By the Deslop desk · Published 12 Aug 2026 · Lexicon version 1.0 (dated because tells drift)
Researchers can measure the arrival of ChatGPT in scientific writing by tracking a single word. Usage of "delve" surged so hard after 2022 that a study of over 15 million PubMed abstracts concluded at least 13.5 percent of 2024 abstracts had been processed with a language model. One word gave the game away.
Instead of repeating someone else's list, we tested ours. We took a passage of AI-flavored marketing copy, the Gettysburg Address, and the opening of Darwin's Origin of Species. Then we counted lexicon hits per 1,000 words using the same engine that powers our free editor.
The gap is sixty to one. That gap is what "sounding like AI" actually is. It was never one forbidden word. It is density.
And notice Lincoln's score. Not zero. His single flagged word is "thus." Hold that thought, because it is the most important fact in this article.
delve leverage utilize foster garner harness navigate embark unlock empower elevate streamline optimize underscore
The fix is nearly always the plain verb. Examine. Use. Build. Earn. Start. Show. If the plain verb feels too small, the sentence lacks a fact, not a synonym.
tapestry landscape realm journey vibrant bustling seamless robust intricate multifaceted holistic comprehensive testament insights plethora myriad
These decorate instead of describe. Name the real thing. What is actually in the "landscape"? Say that instead.
furthermore moreover additionally consequently thus overall ultimately
Well-built sentences connect themselves. Try deleting the connector and reading the two sentences aloud. It almost always works.
crucial pivotal significantly notably arguably undoubtedly undeniably
When there is nothing concrete to say, machine prose compensates with emphasis. Every intensifier is a slot where a number or a fact should be.
Humans wrote with every word on this list for centuries before language models existed. Lincoln used "thus" in the most famous short speech in American history. So a word list can never prove authorship. It can only show density and habit.
This is not a small point. Students have received zeros from AI detectors on work they wrote themselves. In one 2023 case a student was cleared only after his family showed the same detector flagged the Gettysburg Address as machine-written. In January 2026 a New York court ordered a university to expunge a freshman's record, calling the AI finding "without valid basis and devoid of reason." Tools that treat tells as proof put real people in front of tribunals.
We built our editor on the opposite premise. It flags density and habit as writing problems you can fix, makes no authorship claim, and never uploads your text anywhere.
The tells move. A study of 4,820 undergraduate reports found markers like "delve" and "intricate" surged until 2024, then declined in 2025 as writers adapted. Any undated list of AI words you find is describing the past.
So this page is versioned like software. Lexicon v1.0, published 12 August 2026. When the data shifts, the version number changes and the changelog will say what moved. If you cite this list, cite the version.
Sounding like AI is a density problem, not a vocabulary crime. AI-style copy hit 222 tells per 1,000 words in our test. Lincoln hit 3.7 and still tripped a flag, which is why word lists prove nothing about authorship. Use the list to spot your habits, replace decoration with facts, and check your density in thirty seconds.
Run your text through Deslop, free and private →Sources: Kobak et al., Science Advances, excess-vocabulary analysis of 15M+ PubMed abstracts. ScienceDirect study of 4,820 undergraduate reports on marker decline in 2025. USA Today, April 2023, on the UC Davis case. Matter of Newby v. Adelphi Univ., 2026 NY Slip Op 26021. Our chart data: computed 12 Aug 2026 with Deslop lexicon v1.0; texts are public domain; arithmetic shown in the caption. Tool pages: Deslop editor · readability checker · word counter.