Before November 2022, "delve" lived a quiet life. British academics used it occasionally. Fantasy novelists sent characters delving into dungeons. Corpus data from the preceding decade shows a flat, modest frequency: the profile of a word nobody thought about.
Then the curve bent upward, hard.
Studies of biomedical abstracts found the word appearing at many times its historical rate within eighteen months. Editors began spotting it everywhere: consultancy reports, student essays, LinkedIn posts. A verb that had been mildly bookish became the single most recognisable fingerprint of machine-assisted prose.
Why this word, of all words? The training and feedback process behind large language models appears to have concentrated certain lexical preferences, and reinforcement from human raters amplified them. One plausible contributing factor, reported by several investigations into the data-labelling industry, involves regional English: the word enjoys higher everyday currency in Nigerian English, and Nigerian annotators did a substantial share of feedback work. Whatever the exact mechanism, the outcome is measurable. The word now carries information about how a text was produced.
Tellmark counts it under entry L-01, weighted at the maximum. Not because the verb is bad (it is a perfectly good verb) but because its post-2023 frequency is so far above any organic baseline that readers have learned the association. A hiring manager who meets it in a cover letter no longer thinks "bookish". They think "template".
The remedy
The remedy costs nothing. Look at, examine, dig into, cover, get into: English offers half a dozen plainer alternatives, and every one reads as more direct. If the sentence genuinely needs the mildly archaic flavour, keep it. Once, deliberately, where the register earns it.
Editors should treat the word as a prompt rather than a conviction. One instance in a long report from a writer with a bookish streak means little. Three instances in a five-hundred-word cover letter, sitting alongside ornamental metaphors and stock scene-setting, tells a different story, and the surrounding evidence is what makes it credible. This is why the catalogue insists on counting rather than concluding: a single lexical fingerprint, however famous, never justifies an accusation. What it justifies is a closer read.
There is also a practical warning for anyone tempted to fix the symptom alone. Search-and-replace scrubs this word in seconds, and detection-evasion tooling already does exactly that. Rhythm, hedging saturation and template constructions survive the synonym pass, which is why the structural half of the catalogue exists. A text scrubbed of its most famous fingerprint but untouched at the sentence level still reads as machine-flavoured, and increasingly, readers can tell without any tool at all.
A note on our own score.
This article discusses the specimen, so it contains the specimen twice, and Tellmark marks both instances plus every other pattern the running prose commits. Determinism means no exemptions, including for us. The number in the footer below is real, and during drafting the engine caught this very page using an ornamental metaphor as a casual example, which we then had to cut. The catalogue disciplines its own authors first.