A limitation is that you have to add a ton of details to your definitions of characters to get around the fact that AI doesn't actually have the ability to intuit what a person would or would not believe, it only knows about the associations in its dataset.

As an example, look here: the neo-Nazi character expectedly admires Japan for "ethnic purity" but in real life he wouldn't go off about how bad the atrocities they committed in the war were.

it also has some unnatural statements that are the result of the dataset so ideally someday in the future there will be neutral datasets that have "...and that's awful/good" statements removed from them.