"Foster" was another word that increased 10-fold (10.84x, to be exact) in prevalence when comparing pre- and post-ChatGPT data.
We also noted massive jumps in mean answer length (both tokens and sentences), Flesch Reading Ease and the mean Type-Token Ratio starting 3mo post-ChatGPT.
Kobak et al. (2025, arxiv.org/pdf/2406.07016) analyzed pubmed abstract from 2010-2024, identifying a bunch of words that are at least highly suspicious.
In upcoming work for AIED'25, we looked at student MOOC essays from 2020-2024 and identified very similar trends, e.g. "delve" increasing 10-fold.
/3 But here’s the great thing, the vendor explained: you can tell the AI not to include hallucinations in its work product.
I beg your pardon, I said to the paralegal relating this to me.
You can set hallucinations to “no,” he explained.
