Delve is gone. Its replacements are everywhere.

The line sits between 1.1% and 1.4% for four years. It starts to climb in mid-2023, roughly six months after ChatGPT's release, which matches the time a paper takes to go from draft to publication. It reaches 12.2% in November 2025, then drifts down to 8.1% by September 2026.

The drop at the end should be read with care. The 20 words were chosen from 2024 and 2025 text, and the next sections show that the vocabulary keeps changing. A fixed word list loses its grip over time, so a falling line can mean the list is ageing while the practice continues.

The words take turns

Each bar shows how many times more common a word is than it was before 2023. Press play to watch the ranking change. "Delves" peaks at about 80 times its old rate in mid-2024, then sinks as writers and newer AI models learn to avoid it. By 2026 the top of the table belongs to words few people have been warned about.

The replacements are far bigger than "delve" ever was

Both charts use the same scale. At its height "delves" appeared in about 45 of every 10,000 abstracts. "Integrates" now appears in 208, and "synthesizes" has gone from 4 to 101. These words are harder to police because each one is ordinary on its own.

Ordinary academic vocabulary did not move. "However" appeared in 23% of abstracts in 2019 and 25% in 2026.

Every country moved. Some moved twice as far.

The share in 2025 runs from 7.5% for papers with an author in Japan to 16.8% for Saudi Arabia. India is at 15.2%. In 2022 every country sat between 0.5% and 2.3%. Much of this is likely researchers who write in a second language using AI to polish their English, which most publishers permit when it is disclosed.

For publishers, the gap is between journals

Across the eight largest publishers in this set, the 2025 share sits in a narrow band from 10.5% to 14.5%. None stands apart. Between individual journals the range is 2.7% to 21.5%, so the useful unit for an integrity team is the journal and its field.

Field explains a lot of this. Chemistry and taxonomy abstracts are terse and formulaic, so they score low. Public health, psychology and medical computing abstracts are discursive, so they score high.

What a publisher can take from this

• Word lists expire. A screen built on "delve" in 2024 would miss most of 2026. Detection has to be re-derived from current text on a schedule.
• Benchmark by journal. A journal at 20% when its field peers are at 10% is a concrete question for its editor.
• Pair it with disclosure. Most policies allow AI for language editing if declared. Comparing these rates with declared use shows how large the undeclared share is.

What this does and does not show

• It measures wording, and nothing about misconduct. A marker word shows the text likely passed through an AI tool. It says nothing about whether the science is sound.
• It undercounts. Text edited to remove the tell-tale words, or written by a newer model, is missed.
• It covers biomedical and life-science research only, the scope of PubMed. Other fields may differ.
• The 20 words were picked by data. About 43,000 abstracts from 2021–22 were compared with about 43,000 from 2024–25, and the words with the largest jump were kept. Subject terms that rose for their own reasons, such as new disease names, were removed by hand.
• Country is counted from any author's affiliation, so an international paper counts for each country on it.
• The most recent months are still being indexed and may shift.