I once amused myself by coding a variant on "scramble the middle, keep first and last letter the same": instead of random scrambles, permute so that the resulting word is the most probable nonword according to a letter-trigram model of English.

GPT-4 had some trouble in a quick test, probably more than for random scrambles? But got most of it: https://chat.openai.com/share/51f1a94e-b35c-4dbc-945b-ef5983...

(It made one other mistake I didn't comment on there.)