Comment by boothby

This is why misspellings and homophones are tells of human righting. LLMs strongly prefer word-level tokens, and word substitutions follow semantic similarity and not the more human auditory similarity.