Breaking01 / AI
AI-favoured words rise 29% on least-edited sites
A study linked vocabulary growth to how often answers were edited, but did not observe actual model use.

A paper's author reported on 13 Nov that words favoured by language models rose about 29% relative to matched controls on Stack Exchange sites where answers were edited least often.
The paper tracked 56 words that language models overuse across 170 sites on Stack Exchange, a network of question-and-answer communities. It analysed answers posted from 2019 to 2024.
According to the paper, use of those words rose about 14% above its pre-2023 trend. The largest increase among 1,000 lists of ordinary words matched for frequency was about 7%.
The paper reported no increase in the tracked vocabulary among the sites where answers were edited most often. The difference associated with editing rates also appeared among answers that were never edited.
Of 15 site characteristics measured beforehand, the paper identified the share of answers receiving edits as the strongest predictor of vocabulary growth. It defined that share as the proportion of answers whose body text was edited within 30 days of posting.
The paper reported that the relationship followed the share of answers edited by their own authors. Once that share was held constant, the share edited by other people added little explanatory value.
The study did not observe actual model use and could not distinguish lower usage from any effect of editing on the vocabulary measured.