There's no way anyone knows where the limit is of a feature set that is valid for AI enhancement. Substitutions and changes in amounts across a dataset that is not fully pre-taxonomized seems like a pretty good use of AI to be honest.

We've used it for modifications to enjoyable effect a few years ago https://news.ycombinator.com/item?id=46131280

So it's a matter of execution etc. to see which features stick.