@artemis IIRC there was a huge thing on African Twitter about this: a lot of the LLM training/tagging was done by Nigerian workers (underpaid ghost labour) and Nigerian English is gorgeously full of expressive and specific terms which introduces a bias in LLMs to use those words like delve, devoid, etc, and AI checkers were updated in turn to look for those as markers and people automatically assuming stuff was written by LLMs.