736,000 computer-science and engineering abstracts on arXiv · January 2019 to September 2026
Delve is all but gone. Tech papers are on their third vocabulary.
2024This paper delves into the intricate design space, showcasing a pivotal trade-off.2025Our results underscore an underexplored failure mode, paving the way for safer systems.2026The gain persists in simulation but collapses sharply on real hardware, which reframes the benchmark.Invented sentences built from each year's marker words. No real paper is quoted.
After ChatGPT arrived, a set of words spread through tech research abstracts and then died out within two years. A second set replaced it, and a third is rising now. Each set lasted about a year at its height.
−96%fall in "delve", in any form, from 2024 to 2026. It is now at half its pre-ChatGPT rate
1 in 4carried a marker word at the peak in March 2025, up from 1 in 20 in 2022
1.0% → 11.8%share using the newest set of words, 2022 to August 2026
Share of tech abstracts containing at least one word from each setMonthly, arXiv papers whose main category is computer science or electrical engineering. Hover for values.
2024 words2025 words2026 words
The first set, with "delves", "intricate" and "showcasing", climbs from 2.3% of abstracts in 2022 to 15.4% in July 2024. By September 2026 it is at 1.5%, lower than before ChatGPT existed. Writers now avoid these words.
The second set, with "underscoring" and "underexplored", peaks at 15.3% in May 2025 and has fallen to 5.4%. The third set, with "collapses", "persists" and "reframes", stayed under 2% through 2024 and reached 11.8% in August 2026.
Taken together, an abstract with any marker word went from 5.0% in 2022 to 26.2% in March 2025, and stands at 17.0%.
The words take turns
Each bar is one word, with its forms combined, and shows how many times more common it is than in 2019 to 2022. Press play to watch the ranking change. "Delve" reaches about 21 times its old rate in January 2024 and drops below that old rate by 2026. In September 2026 the top word is "survives", at about 100 times its old rate.
2024 words2025 words2026 words× = multiple of the word's 2019–2022 rate, 3-month average
The newest words are plain, which makes them hard to catch
Both charts use the same scale. "Delve" was always an unusual word, so it was easy to spot and easy to ban. "Structurally" and "collapses" are ordinary technical English. "Structurally" has gone from 12 to 143 of every 10,000 abstracts, and "collapses" from 3 to 98.
Peaked in early 2024, then fellAbstracts per 10,000 containing the word, 3-month averageRising through 2026Abstracts per 10,000 containing the word, 3-month average
Tech leads every other field
arXiv also hosts physics, mathematics and other fields, which allows a like-for-like comparison. In 2025, 24.9% of computer-science abstracts carried a marker word. Physics was at 11.0% and mathematics at 3.4%.
Share of 2025 abstracts with any marker word, by fieldThe dark tick is the same field's share in 2022. Hover for paper counts.
20252022
Inside tech, the AI subfields are highest
Across 18 subfields the 2025 share runs from 5.0% in algorithms to 29.9% in artificial intelligence. Vision, language processing and information retrieval sit near the top. Theory subfields, where abstracts are short and mathematical, sit at the bottom.
Share of 2025 abstracts with any marker word, by subfieldSubfields with at least 1,500 papers in 2025, by main arXiv category. The dark tick is 2022.
How the words were chosen, and why the rise is real
The data picked the words. A word had to at least triple in tech abstracts and also at least triple in the other arXiv fields. The second condition removes subject terms such as "LLM" or "prompt", which rose in tech only because the research topic changed.
The same test finds almost nothing before ChatGPT. From 2019 to 2022 only two words pass it, "pandemic" and "GNNs", and both are subject terms. From 2022 to 2025, 85 words pass.
Each form of a word was tested on its own. The habit attaches to particular forms: "advancing" rose ten times, while "advanced" was already common and less than tripled. Only forms that passed are counted, 52 forms of 47 words.
The remaining subject terms were removed by hand, and each word was assigned to the year in which it peaked.
What a publisher or conference can take from this
A word list lasts about a year. A screen built on the 2024 words would flag 1.5% of today's abstracts. The current words flag 10.9%.
Preprints give early warning. The rise shows on arXiv by April 2023, months before it appears in journals, so the next vocabulary can be found here first.
Benchmark by subfield. A venue far above its subfield's rate has a concrete question to ask its reviewers.
What this does and does not show
It measures wording, and nothing about misconduct. A marker word suggests the text passed through an AI tool. It says nothing about whether the research is sound.
The link between the 2026 words and newer AI models is an inference. The timing and the spread across unrelated fields fit that reading, but this page cannot prove the cause.
It undercounts. Text edited to remove the tell-tale words is missed, and so are other forms that also rose, such as "delving".
It covers arXiv only. arXiv records no publisher or country, so those views are not available here.
Each paper is counted in its first-submission month with its latest abstract. A few older papers were revised later.