It's difficult to ascertain what justice is, but since the beginning of recorded history it's generally been assumed that if a decent chunk of the population are explained a law and don't think they're being screwed over then it's got an above average chance of being a good thing.

In other words, voting/rebelling is generally assumed to be within the rights of people in response to things they want to be different.

Under that assumption; One less-problematic means of assisting government through the use of LLMs/AI would be to simply feed legal texts and interpretations to the machine on one end and come up with a simplified more human readable version on the other end.

This would not determine what people do but how the existing laws for behavior would be abbreviated and more readable in order to ensure citizens can understand what they are to do and lawmakers can change rules in a way that is mutually transparent.